snapshot of https://github.com/seaweedfs/seaweedfs.git (master, 2026-09-16; blobs > 131072B omitted)

seaweed-worker/.gitignore

@@ -0,0 +1 @@
+target/

seaweed-worker/Cargo.toml

@@ -0,0 +1,38 @@
+# Rust plugin workers for SeaweedFS.
+#
+# `core` is the plugin.proto contract and nothing else; a worker crate beside it
+# supplies job handlers and a binary. Adding a worker means adding a member here,
+# not touching the protocol. `sort` is neither: it is the sort specification the
+# sorting jobs share, so that two of them cannot drift on what an order means.
+[workspace]
+resolver = "2"
+members = ["crates/core", "crates/lance", "crates/sort"]
+
+[workspace.package]
+version = "0.1.0"
+edition = "2024"
+# The edition needs 1.85; the dependency tree needs more. Verified with
+# `cargo +1.94.1 check --all-targets` (1.94.0 fails on the AWS SDK that
+# lance's `aws` feature pulls in).
+rust-version = "1.94.1"
+
+[workspace.lints.clippy]
+# Protobuf message literals keep `..Default::default()` on purpose: it is
+# what lets a proto gain a field without touching every constructor.
+needless_update = "allow"
+
+[workspace.dependencies]
+anyhow = "1"
+async-trait = "0.1"
+prost = "0.14"
+prost-types = "0.14"
+tokio = { version = "1", features = ["full"] }
+tokio-stream = "0.1"
+tonic = { version = "0.14", features = ["tls-aws-lc"] }
+tonic-prost = "0.14"
+# Already in the tree via tonic; named here so the metrics server can use them.
+axum = "0.8"
+prometheus = { version = "0.13", default-features = false }
+tonic-prost-build = "0.14"
+tracing = "0.1"
+tracing-subscriber = { version = "0.3", features = ["env-filter"] }

seaweed-worker/README.md

@@ -0,0 +1,110 @@
+# SeaweedFS Rust workers
+
+`weed/pb/plugin.proto` is a language-agnostic contract: a maintenance worker
+connects out to admin, announces the job types it can detect and execute, and
+answers requests on that one stream. `weed worker -admin=host:23646` is the Go
+implementation of it from outside the admin process. This workspace is the Rust
+one.
+
+    crates/core     the contract: stream, handshake, heartbeat, registry, config forms
+    crates/lance    maintenance jobs for Lance tables, and a binary
+
+`core` knows nothing about any job. A second worker is a new crate beside
+`lance` that depends on it, not a fork of the protocol.
+
+## Building
+
+Requires Rust 1.94.1+ (2024 edition), matching `rust-version` in `Cargo.toml`.
+The patch release matters: 1.94.0 does not build. The edition itself only needs
+1.85; the higher floor comes from the dependency tree — lance's `aws` feature
+pulls in the AWS SDK — so it moves with those crates. CI builds on the latest
+stable.
+
+`core` compiles `plugin.proto` with the protoc that protoc-bin-vendored ships,
+the way seaweed-volume does, so it needs no system install.
+
+The lance crates compile protos of their own, in their own build-script
+processes, which nothing our build script sets can reach. They need a protoc of
+their own: either one on PATH — `brew install protobuf`, `apt install
+protobuf-compiler` — or `PROTOC` naming one. CI points it at the vendored
+binary for the runner's platform, resolved from the version in `Cargo.lock`.
+
+## Running
+
+    cargo run -p weed-lance-worker -- --admin 127.0.0.1:23646
+
+The admin's *HTTP* address is what an operator has; the gRPC port is derived
+from it the way the Go side does. Dialling the HTTP port fails as "frame with
+invalid size", which reads like a protocol bug rather than a wrong port.
+
+The binary is `weed-worker`, not `weed-lance-worker`: it is the Rust side of
+`weed worker`, and lance is the first family of jobs it carries rather than the
+only one it ever will.
+
+Released builds do not need a toolchain. The worker ships inside the SeaweedFS
+image, beside the Rust volume server, under the verb that mirrors
+`volume-rust`:
+
+    docker run chrislusf/seaweedfs worker-rust --admin admin:23646
+
+and as `weed-worker_linux_{amd64,arm64}.tar.gz` on each GitHub release. Both are
+linux amd64/arm64 only — lance, arrow and datafusion make every extra target an
+expensive build, and the worker runs beside the cluster it maintains. On an
+architecture without a build the image carries an empty placeholder and the
+entrypoint says so rather than failing as "not found".
+
+## Metrics
+
+    cargo run -p weed-lance-worker -- --admin 127.0.0.1:23646 --metrics-port 9328
+
+Serves `/health`, `/ready` and `/metrics` on that port, the same three the Go
+worker serves under `weed worker -metricsPort`, so one scrape config covers
+workers in either language. Off by default, and bound to loopback unless
+`--metrics-ip` says otherwise, because the endpoint is unauthenticated. 9328
+continues the series the other components use (master 9324, volume 9325, filer
+9326, s3 9327); an IPv6 address works with or without brackets.
+
+Grafana: the "Plugin Workers" row of `other/metrics/grafana_seaweedfs.json`
+graphs these. Its panels filter on `$cluster`, which comes from the scrape job's
+labels, so scrape the worker the way the rest of the cluster is scraped or the
+row stays empty.
+
+Names are `SeaweedFS_worker_*`, matching the Go side's convention. The pair
+worth alerting on is `objects_seen_total` and `objects_skipped_total`: a sweep
+that proposes nothing and a sweep that could read nothing look identical from
+`proposals_total` alone.
+
+    SeaweedFS_worker_connected 1
+    SeaweedFS_worker_objects_seen_total{job_type="lance_compact"} 7
+    SeaweedFS_worker_proposals_total{job_type="lance_compact"} 2
+    SeaweedFS_worker_jobs_total{job_type="lance_compact",result="ok"} 2
+    SeaweedFS_worker_lance_fragments_removed_total 25
+
+`/ready` follows the control stream: a worker whose admin has gone away is
+running but is not going to do anything.
+
+## Credentials
+
+The worker holds none. It asks the namespace to describe a table with
+`vend_credentials` and hands the `storage_options` that come back to lance. A
+gateway without STS configured vends no credentials at all, so `--access-key`
+and `--secret-key` supply a fallback; anything the namespace does vend wins over
+them.
+
+## State
+
+All three jobs are implemented and tested end to end against a live gateway:
+
+    compaction result: 12 fragments became 1
+    reindex result:    512 uncovered rows became 0
+    cleanup result:    removed 14 versions and 24272 bytes
+
+`cargo test -p weed-lance-worker` runs them when `WEED_LANCE_NAMESPACE` names a
+live namespace and skips otherwise, the way the Go integration tests skip
+without Docker. Each test seeds the table it needs, including building a vector
+index and then appending rows outside it, so a run does not depend on what the
+previous one left behind — the first version of these did, and quietly stopped
+testing anything once it had done its job.
+
+The handshake, descriptor exchange and heartbeat work against a live admin,
+which logs the worker connecting and prefetches all three descriptors.

seaweed-worker/crates/core/Cargo.toml

@@ -0,0 +1,32 @@
+[package]
+name = "seaweed-worker-core"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+description = "SeaweedFS plugin.proto worker contract"
+
+[lib]
+name = "seaweed_worker_core"
+
+[dependencies]
+anyhow.workspace = true
+async-trait.workspace = true
+prost.workspace = true
+prost-types.workspace = true
+tokio.workspace = true
+tokio-stream.workspace = true
+tonic.workspace = true
+tonic-prost.workspace = true
+axum.workspace = true
+prometheus.workspace = true
+tracing.workspace = true
+
+[build-dependencies]
+tonic-prost-build.workspace = true
+# Ships protoc with the build so neither CI nor a developer needs a system
+# install, and so the version is pinned rather than whatever the platform's
+# package manager happens to carry. The same crate seaweed-volume uses.
+protoc-bin-vendored = "3"
+
+[lints]
+workspace = true

seaweed-worker/crates/core/build.rs

@@ -0,0 +1,25 @@
+fn main() -> Result<(), Box<dyn std::error::Error>> {
+    // Use the protoc that ships with protoc-bin-vendored rather than a system
+    // one, so the build needs no package manager and always sees the same
+    // version. An explicit PROTOC still wins, for packagers supplying their own
+    // and for the lance crates, whose own build scripts read the same variable.
+    if std::env::var_os("PROTOC").is_none() {
+        // SAFETY: a build script's main runs single-threaded before anything
+        // else in this process, so no other thread can be reading the
+        // environment concurrently.
+        unsafe {
+            std::env::set_var("PROTOC", protoc_bin_vendored::protoc_bin_path()?);
+        }
+    }
+
+    // Compiled straight out of the Go tree, the way seaweed-volume already reads
+    // filer.proto, so the contract cannot drift from a vendored copy.
+    tonic_prost_build::configure()
+        // The server half is only for tests, which stand up a fake admin.
+        .build_server(true)
+        .build_client(true)
+        .protoc_arg("--experimental_allow_proto3_optional")
+        .compile_protos(&["../../../weed/pb/plugin.proto"], &["../../../weed/pb/"])?;
+    println!("cargo:rerun-if-changed=../../../weed/pb/plugin.proto");
+    Ok(())
+}

seaweed-worker/crates/core/src/address.rs

@@ -0,0 +1,73 @@
+//! SeaweedFS addresses the way the Go tree does.
+//!
+//! An operator gives a worker the admin's HTTP address, and the gRPC port is
+//! derived from it rather than asked for separately. Dialling the HTTP port
+//! instead fails as "frame with invalid size", which reads like a protocol bug
+//! rather than a wrong port, so getting this right is worth its own module.
+//! Mirrors pb.ServerToGrpcAddress in weed/pb/grpc_client_server.go.
+
+const GRPC_PORT_OFFSET: u16 = 10000;
+
+/// Converts `host:port` to the gRPC address, and accepts the explicit
+/// `host:port.grpcPort` form the Go side also understands.
+pub fn server_to_grpc_address(server: &str) -> Option<String> {
+    let (host, port_part) = server.rsplit_once(':')?;
+
+    // "port.grpcPort" states the gRPC port outright.
+    if let Some((_, grpc_port)) = port_part.split_once('.')
+        && let Ok(port) = grpc_port.parse::<u16>()
+    {
+        return Some(join_host_port(host, port));
+    }
+
+    let port: u16 = port_part.parse().ok()?;
+    Some(join_host_port(host, port.checked_add(GRPC_PORT_OFFSET)?))
+}
+
+fn join_host_port(host: &str, port: u16) -> String {
+    // An IPv6 literal has to keep its brackets or the port reads as part of it.
+    if host.contains(':') && !host.starts_with('[') {
+        format!("[{host}]:{port}")
+    } else {
+        format!("{host}:{port}")
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::server_to_grpc_address;
+
+    #[test]
+    fn derives_the_grpc_port() {
+        assert_eq!(
+            server_to_grpc_address("localhost:23646").as_deref(),
+            Some("localhost:33646")
+        );
+        assert_eq!(
+            server_to_grpc_address("127.0.0.1:9333").as_deref(),
+            Some("127.0.0.1:19333")
+        );
+    }
+
+    #[test]
+    fn honours_an_explicit_grpc_port() {
+        assert_eq!(
+            server_to_grpc_address("localhost:23646.33999").as_deref(),
+            Some("localhost:33999")
+        );
+    }
+
+    #[test]
+    fn brackets_ipv6_literals() {
+        assert_eq!(
+            server_to_grpc_address("::1:23646").as_deref(),
+            Some("[::1]:33646")
+        );
+    }
+
+    #[test]
+    fn rejects_what_it_cannot_parse() {
+        assert!(server_to_grpc_address("localhost").is_none());
+        assert!(server_to_grpc_address("localhost:notaport").is_none());
+    }
+}

seaweed-worker/crates/core/src/config.rs

@@ -0,0 +1,49 @@
+use std::time::Duration;
+
+/// How one worker process connects and how much work it will take on.
+#[derive(Clone, Debug)]
+pub struct WorkerOptions {
+    /// Admin gRPC address, e.g. "localhost:23646".
+    pub admin_address: String,
+    pub worker_id: String,
+    pub worker_version: String,
+    /// Advertised address; empty when the worker takes no inbound connections.
+    pub worker_address: String,
+    pub heartbeat_interval: Duration,
+    pub reconnect_delay: Duration,
+    pub max_detection_concurrency: i32,
+    pub max_execution_concurrency: i32,
+    /// mTLS for the control stream, mirroring the Go worker's `[grpc.worker]`
+    /// section of security.toml. None means plaintext, which is the default the
+    /// Go worker also takes when no certificates are configured.
+    pub tls: Option<TlsOptions>,
+}
+
+/// Certificates for the control stream. All three are required together: the
+/// cluster's gRPC TLS is mutual, so a CA without a client identity gets refused
+/// by admin rather than falling back to one-way TLS.
+#[derive(Clone, Debug)]
+pub struct TlsOptions {
+    pub ca_path: String,
+    pub client_cert_path: String,
+    pub client_key_path: String,
+    /// Name to verify the server certificate against, when the address a worker
+    /// dials is not the name the certificate carries.
+    pub server_name: Option<String>,
+}
+
+impl Default for WorkerOptions {
+    fn default() -> Self {
+        Self {
+            admin_address: "localhost:23646".to_string(),
+            worker_id: String::new(),
+            worker_version: env!("CARGO_PKG_VERSION").to_string(),
+            worker_address: String::new(),
+            heartbeat_interval: Duration::from_secs(10),
+            reconnect_delay: Duration::from_secs(5),
+            max_detection_concurrency: 1,
+            max_execution_concurrency: 1,
+            tls: None,
+        }
+    }
+}

seaweed-worker/crates/core/src/config_form.rs

@@ -0,0 +1,106 @@
+//! Builders for the config forms a worker returns in its JobTypeDescriptor.
+//!
+//! Admin renders these into the job's settings page, so a worker written in any
+//! language gets a UI without touching Go or templ. That only holds if the
+//! field types and defaults are right, which is why they are built here rather
+//! than spelled out at each call site.
+
+use std::collections::HashMap;
+
+use crate::pb::{
+    ConfigField, ConfigFieldType, ConfigForm, ConfigSection, ConfigValue, config_value::Kind,
+};
+
+pub fn int_value(value: i64) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::Int64Value(value)),
+    }
+}
+
+pub fn bool_value(value: bool) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::BoolValue(value)),
+    }
+}
+
+pub fn string_value(value: impl Into<String>) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::StringValue(value.into())),
+    }
+}
+
+/// Reads an integer a request carried, falling back when admin sent nothing.
+pub fn int_or(values: &HashMap<String, ConfigValue>, name: &str, fallback: i64) -> i64 {
+    match values.get(name).and_then(|v| v.kind.as_ref()) {
+        Some(Kind::Int64Value(value)) => *value,
+        Some(Kind::DoubleValue(value)) => *value as i64,
+        _ => fallback,
+    }
+}
+
+pub fn bool_or(values: &HashMap<String, ConfigValue>, name: &str, fallback: bool) -> bool {
+    match values.get(name).and_then(|v| v.kind.as_ref()) {
+        Some(Kind::BoolValue(value)) => *value,
+        _ => fallback,
+    }
+}
+
+pub fn string_or(values: &HashMap<String, ConfigValue>, name: &str, fallback: &str) -> String {
+    match values.get(name).and_then(|v| v.kind.as_ref()) {
+        Some(Kind::StringValue(value)) if !value.is_empty() => value.clone(),
+        _ => fallback.to_string(),
+    }
+}
+
+pub fn number_field(name: &str, label: &str, description: &str, min: i64, max: i64) -> ConfigField {
+    ConfigField {
+        name: name.to_string(),
+        label: label.to_string(),
+        description: description.to_string(),
+        field_type: ConfigFieldType::Int64 as i32,
+        min_value: Some(int_value(min)),
+        max_value: Some(int_value(max)),
+        ..Default::default()
+    }
+}
+
+pub fn text_field(name: &str, label: &str, description: &str, placeholder: &str) -> ConfigField {
+    ConfigField {
+        name: name.to_string(),
+        label: label.to_string(),
+        description: description.to_string(),
+        field_type: ConfigFieldType::String as i32,
+        placeholder: placeholder.to_string(),
+        ..Default::default()
+    }
+}
+
+pub fn bool_field(name: &str, label: &str, description: &str) -> ConfigField {
+    ConfigField {
+        name: name.to_string(),
+        label: label.to_string(),
+        description: description.to_string(),
+        field_type: ConfigFieldType::Bool as i32,
+        ..Default::default()
+    }
+}
+
+pub fn form(
+    form_id: &str,
+    title: &str,
+    fields: Vec<ConfigField>,
+    defaults: HashMap<String, ConfigValue>,
+) -> ConfigForm {
+    ConfigForm {
+        form_id: form_id.to_string(),
+        title: title.to_string(),
+        description: String::new(),
+        sections: vec![ConfigSection {
+            section_id: format!("{form_id}-main"),
+            title: title.to_string(),
+            description: String::new(),
+            fields,
+        }],
+        default_values: defaults,
+    }
+}

seaweed-worker/crates/core/src/lib.rs

@@ -0,0 +1,29 @@
+//! The SeaweedFS plugin worker contract, in Rust.
+//!
+//! `weed/pb/plugin.proto` is a language-agnostic gRPC stream: a worker connects
+//! out to admin, announces the job types it can detect and execute, and then
+//! answers requests on that one stream. `weed worker -admin=host:23646` is the
+//! Go implementation of the same contract from outside the admin process; this
+//! is the Rust one, and it knows nothing about any particular job.
+
+pub mod address;
+pub mod config;
+pub mod config_form;
+pub mod metrics;
+pub mod registry;
+pub mod senders;
+pub mod stream;
+
+/// Generated plugin.proto types.
+pub mod pb {
+    // prost gives every oneof its own enum; the variant sizes are the
+    // messages' own, not a choice made here.
+    #![allow(clippy::large_enum_variant)]
+    tonic::include_proto!("plugin");
+}
+
+pub use config::{TlsOptions, WorkerOptions};
+pub use metrics::Metrics;
+pub use registry::{JobHandler, Preview, PreviewProvider, Registry};
+pub use senders::{DetectionSender, ExecutionSender};
+pub use stream::run;

seaweed-worker/crates/core/src/metrics.rs

@@ -0,0 +1,404 @@
+//! What a worker will say about itself when nobody is watching the logs.
+//!
+//! The Go plugin worker serves `/health`, `/ready` and `/metrics` on an optional
+//! port (`weed worker -metricsPort`); this is the same contract, so one scrape
+//! config covers workers in either language. Names follow the Go side's
+//! convention, `SeaweedFS_<subsystem>_<name>`, under a `worker` subsystem that
+//! nothing else uses.
+//!
+//! Nothing here is Lance-specific. A worker for another format registers its own
+//! collectors on the same registry and gets the same endpoint.
+
+use std::net::SocketAddr;
+use std::sync::Arc;
+
+use anyhow::{Context, Result};
+use prometheus::{
+    Encoder, HistogramOpts, HistogramVec, IntCounter, IntCounterVec, IntGauge, IntGaugeVec, Opts,
+    Registry, TextEncoder,
+};
+use std::sync::atomic::{AtomicBool, Ordering};
+use tracing::{info, warn};
+
+const NAMESPACE: &str = "SeaweedFS";
+const SUBSYSTEM: &str = "worker";
+
+/// The metrics one worker process publishes, and the registry they live on.
+///
+/// Cloneable because the stream loop, the job handlers and the HTTP server all
+/// hold it; everything inside is already shared.
+#[derive(Clone)]
+pub struct Metrics {
+    registry: Registry,
+    connected: IntGauge,
+    connects: IntCounterVec,
+    slots_used: IntGaugeVec,
+    slots_total: IntGaugeVec,
+    detections: IntCounterVec,
+    detection_seconds: HistogramVec,
+    proposals: IntCounterVec,
+    objects_seen: IntCounterVec,
+    objects_skipped: IntCounterVec,
+    jobs: IntCounterVec,
+    job_seconds: HistogramVec,
+    previews: IntCounterVec,
+    ready: Arc<AtomicBool>,
+}
+
+impl Metrics {
+    pub fn new(worker_id: &str, worker_version: &str) -> Result<Self> {
+        let registry = Registry::new();
+
+        // Build info as a constant 1, the way the Go side does it, so a scrape
+        // can tell which worker and which build answered.
+        let build = IntGaugeVec::new(
+            Opts::new("build_info", "Worker build information.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+            &["worker_id", "version"],
+        )?;
+        build.with_label_values(&[worker_id, worker_version]).set(1);
+        registry.register(Box::new(build))?;
+
+        let connected = IntGauge::with_opts(
+            Opts::new("connected", "1 while the admin control stream is up.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+        )?;
+        let connects = IntCounterVec::new(
+            Opts::new(
+                "stream_events_total",
+                "Control stream lifecycle events by outcome (connected, closed, failed, shutdown).",
+            )
+            .namespace(NAMESPACE)
+            .subsystem(SUBSYSTEM),
+            &["event"],
+        )?;
+        let slots_used = IntGaugeVec::new(
+            Opts::new("slots_used", "Slots currently held, by lane.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+            &["lane"],
+        )?;
+        let slots_total = IntGaugeVec::new(
+            Opts::new("slots_total", "Slots this worker advertises, by lane.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+            &["lane"],
+        )?;
+        let detections = IntCounterVec::new(
+            Opts::new(
+                "detections_total",
+                "Detection sweeps by job type and outcome.",
+            )
+            .namespace(NAMESPACE)
+            .subsystem(SUBSYSTEM),
+            &["job_type", "result"],
+        )?;
+        let detection_seconds = HistogramVec::new(
+            HistogramOpts::new("detection_seconds", "How long a detection sweep took.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM)
+                .buckets(prometheus::exponential_buckets(0.01, 2.0, 14)?),
+            &["job_type"],
+        )?;
+        let proposals = IntCounterVec::new(
+            Opts::new("proposals_total", "Jobs proposed to admin, by job type.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+            &["job_type"],
+        )?;
+        let objects_seen = IntCounterVec::new(
+            Opts::new(
+                "objects_seen_total",
+                "Objects a detection sweep read, by job type.",
+            )
+            .namespace(NAMESPACE)
+            .subsystem(SUBSYSTEM),
+            &["job_type"],
+        )?;
+        let objects_skipped = IntCounterVec::new(
+            Opts::new(
+                "objects_skipped_total",
+                "Objects a sweep could not read, by job type and reason. A sweep that \
+                 proposes nothing looks the same as one that could read nothing; this is \
+                 what tells them apart.",
+            )
+            .namespace(NAMESPACE)
+            .subsystem(SUBSYSTEM),
+            &["job_type", "reason"],
+        )?;
+        let jobs = IntCounterVec::new(
+            Opts::new("jobs_total", "Jobs executed by job type and outcome.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+            &["job_type", "result"],
+        )?;
+        let job_seconds = HistogramVec::new(
+            HistogramOpts::new("job_seconds", "How long a job took to run.")
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM)
+                .buckets(prometheus::exponential_buckets(0.05, 2.0, 16)?),
+            &["job_type"],
+        )?;
+        let previews = IntCounterVec::new(
+            Opts::new(
+                "previews_total",
+                "Object previews served to admin, by outcome.",
+            )
+            .namespace(NAMESPACE)
+            .subsystem(SUBSYSTEM),
+            &["result"],
+        )?;
+
+        for collector in [
+            Box::new(connected.clone()) as Box<dyn prometheus::core::Collector>,
+            Box::new(connects.clone()),
+            Box::new(slots_used.clone()),
+            Box::new(slots_total.clone()),
+            Box::new(detections.clone()),
+            Box::new(detection_seconds.clone()),
+            Box::new(proposals.clone()),
+            Box::new(objects_seen.clone()),
+            Box::new(objects_skipped.clone()),
+            Box::new(jobs.clone()),
+            Box::new(job_seconds.clone()),
+            Box::new(previews.clone()),
+        ] {
+            registry.register(collector)?;
+        }
+
+        Ok(Self {
+            registry,
+            connected,
+            connects,
+            slots_used,
+            slots_total,
+            detections,
+            detection_seconds,
+            proposals,
+            objects_seen,
+            objects_skipped,
+            jobs,
+            job_seconds,
+            previews,
+            ready: Arc::new(AtomicBool::new(false)),
+        })
+    }
+
+    /// The registry, so a worker can add collectors of its own.
+    pub fn registry(&self) -> &Registry {
+        &self.registry
+    }
+
+    pub fn stream_connected(&self) {
+        self.connected.set(1);
+        self.ready.store(true, Ordering::Relaxed);
+        self.connects.with_label_values(&["connected"]).inc();
+    }
+
+    /// `event` is why the stream ended: closed, failed, or shutdown.
+    pub fn stream_ended(&self, event: &str) {
+        self.connected.set(0);
+        self.ready.store(false, Ordering::Relaxed);
+        self.connects.with_label_values(&[event]).inc();
+    }
+
+    pub fn set_slots(&self, lane: &str, used: i64, total: i64) {
+        self.slots_used.with_label_values(&[lane]).set(used);
+        self.slots_total.with_label_values(&[lane]).set(total);
+    }
+
+    pub fn detection_finished(&self, job_type: &str, result: &str, seconds: f64, proposals: usize) {
+        self.detections
+            .with_label_values(&[job_type, result])
+            .inc_by(1);
+        self.detection_seconds
+            .with_label_values(&[job_type])
+            .observe(seconds);
+        self.proposals
+            .with_label_values(&[job_type])
+            .inc_by(proposals as u64);
+    }
+
+    pub fn object_seen(&self, job_type: &str) {
+        self.objects_seen.with_label_values(&[job_type]).inc();
+    }
+
+    pub fn object_skipped(&self, job_type: &str, reason: &str) {
+        self.objects_skipped
+            .with_label_values(&[job_type, reason])
+            .inc();
+    }
+
+    pub fn job_finished(&self, job_type: &str, result: &str, seconds: f64) {
+        self.jobs.with_label_values(&[job_type, result]).inc();
+        self.job_seconds
+            .with_label_values(&[job_type])
+            .observe(seconds);
+    }
+
+    pub fn preview_finished(&self, result: &str) {
+        self.previews.with_label_values(&[result]).inc();
+    }
+
+    /// A counter this worker's own jobs can raise, e.g. fragments removed.
+    /// Registered lazily so a format's numbers live beside the generic ones
+    /// without core having to know what they are.
+    pub fn counter(&self, name: &str, help: &str) -> Result<IntCounter> {
+        let counter = IntCounter::with_opts(
+            Opts::new(name, help)
+                .namespace(NAMESPACE)
+                .subsystem(SUBSYSTEM),
+        )?;
+        self.registry.register(Box::new(counter.clone()))?;
+        Ok(counter)
+    }
+
+    fn gather(&self) -> Result<String> {
+        let mut buffer = Vec::new();
+        TextEncoder::new().encode(&self.registry.gather(), &mut buffer)?;
+        Ok(String::from_utf8(buffer)?)
+    }
+
+    fn is_ready(&self) -> bool {
+        self.ready.load(Ordering::Relaxed)
+    }
+}
+
+/// Serves the metrics endpoints until the process stops. Failing to bind is
+/// logged rather than fatal: a worker that cannot publish metrics should still
+/// do its work.
+pub async fn serve(metrics: Metrics, addr: SocketAddr) -> Result<()> {
+    use axum::Router;
+    use axum::extract::State;
+    use axum::http::StatusCode;
+    use axum::routing::get;
+
+    let app = Router::new()
+        .route("/health", get(|| async { StatusCode::OK }))
+        .route(
+            "/ready",
+            get(|State(metrics): State<Metrics>| async move {
+                if metrics.is_ready() {
+                    StatusCode::OK
+                } else {
+                    StatusCode::SERVICE_UNAVAILABLE
+                }
+            }),
+        )
+        .route(
+            "/metrics",
+            get(|State(metrics): State<Metrics>| async move {
+                match metrics.gather() {
+                    Ok(body) => (StatusCode::OK, body),
+                    Err(err) => (StatusCode::INTERNAL_SERVER_ERROR, format!("{err:#}")),
+                }
+            }),
+        )
+        .with_state(metrics);
+
+    let listener = tokio::net::TcpListener::bind(addr)
+        .await
+        .with_context(|| format!("bind the metrics server to {addr}"))?;
+    info!("serving worker metrics on http://{addr}/metrics");
+    axum::serve(listener, app)
+        .await
+        .context("metrics server stopped")
+}
+
+/// Starts the metrics server in the background, warning rather than failing when
+/// it cannot start.
+pub fn spawn(metrics: Metrics, addr: SocketAddr) {
+    tokio::spawn(async move {
+        if let Err(err) = serve(metrics, addr).await {
+            warn!("worker metrics server: {err:#}");
+        }
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn metrics() -> Metrics {
+        Metrics::new("worker-1", "0.1.0").expect("build metrics")
+    }
+
+    // A scrape has to be able to tell these apart: a worker that swept and found
+    // nothing to do, and one that could not read anything it swept.
+    #[test]
+    fn a_sweep_reports_what_it_read_and_what_it_could_not() {
+        let metrics = metrics();
+        metrics.object_seen("lance_compact");
+        metrics.object_seen("lance_compact");
+        metrics.object_skipped("lance_compact", "open");
+        metrics.detection_finished("lance_compact", "ok", 0.5, 0);
+
+        let text = metrics.gather().expect("gather");
+        assert!(text.contains("SeaweedFS_worker_objects_seen_total{job_type=\"lance_compact\"} 2"));
+        assert!(text.contains(
+            "SeaweedFS_worker_objects_skipped_total{job_type=\"lance_compact\",reason=\"open\"} 1"
+        ));
+        assert!(text.contains(
+            "SeaweedFS_worker_detections_total{job_type=\"lance_compact\",result=\"ok\"} 1"
+        ));
+        assert!(text.contains("SeaweedFS_worker_proposals_total{job_type=\"lance_compact\"} 0"));
+    }
+
+    // /ready follows the stream, because a worker with no admin behind it is
+    // running but not doing anything.
+    #[test]
+    fn readiness_follows_the_stream() {
+        let metrics = metrics();
+        assert!(!metrics.is_ready(), "not ready before the stream is up");
+
+        metrics.stream_connected();
+        assert!(metrics.is_ready());
+        assert!(
+            metrics
+                .gather()
+                .unwrap()
+                .contains("SeaweedFS_worker_connected 1")
+        );
+
+        metrics.stream_ended("closed");
+        assert!(!metrics.is_ready());
+        assert!(
+            metrics
+                .gather()
+                .unwrap()
+                .contains("SeaweedFS_worker_connected 0")
+        );
+        assert!(
+            metrics
+                .gather()
+                .unwrap()
+                .contains("SeaweedFS_worker_stream_events_total{event=\"closed\"} 1")
+        );
+    }
+
+    #[test]
+    fn build_info_names_the_worker() {
+        let text = metrics().gather().expect("gather");
+        assert!(
+            text.contains(
+                "SeaweedFS_worker_build_info{version=\"0.1.0\",worker_id=\"worker-1\"} 1"
+            )
+        );
+    }
+
+    // A format's own numbers land on the same registry, so one endpoint serves
+    // both and a second worker implementation needs no new plumbing.
+    #[test]
+    fn a_worker_can_add_counters_of_its_own() {
+        let metrics = metrics();
+        let counter = metrics
+            .counter("lance_fragments_removed_total", "Fragments merged away.")
+            .expect("register");
+        counter.inc_by(16);
+
+        let text = metrics.gather().expect("gather");
+        assert!(text.contains("SeaweedFS_worker_lance_fragments_removed_total 16"));
+    }
+}

seaweed-worker/crates/core/src/registry.rs

@@ -0,0 +1,89 @@
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use anyhow::Result;
+use async_trait::async_trait;
+
+use crate::pb::{ExecuteJobRequest, JobTypeCapability, JobTypeDescriptor, RunDetectionRequest};
+use crate::senders::{DetectionSender, ExecutionSender};
+
+/// One job type, worker side. Mirrors the Go JobHandler interface in
+/// weed/plugin/worker/worker.go so the two stay readable against each other.
+#[async_trait]
+pub trait JobHandler: Send + Sync {
+    fn capability(&self) -> JobTypeCapability;
+    /// The descriptor admin renders as this job's settings page.
+    fn descriptor(&self) -> JobTypeDescriptor;
+    async fn detect(
+        &self,
+        request: &RunDetectionRequest,
+        sender: &dyn DetectionSender,
+    ) -> Result<()>;
+    async fn execute(
+        &self,
+        request: &ExecuteJobRequest,
+        sender: &dyn ExecutionSender,
+    ) -> Result<()>;
+}
+
+/// Sample rows of one object, already rendered as text. The worker formats
+/// them because it is the only side that knows the object's types.
+pub struct Preview {
+    pub columns: Vec<String>,
+    pub rows: Vec<Vec<String>>,
+    /// Rows in the object, which is not the number sampled.
+    pub total_rows: i64,
+}
+
+/// Reads sample rows of a format admin cannot parse itself.
+///
+/// This is deliberately not a JobHandler: a preview is answered while someone
+/// waits on a page, so it neither schedules nor reports progress.
+#[async_trait]
+pub trait PreviewProvider: Send + Sync {
+    /// The format this provider reads, matched case-insensitively against what
+    /// the catalog recorded.
+    fn format(&self) -> &str;
+    async fn preview(&self, object_id: &[String], row_limit: usize) -> Result<Preview>;
+}
+
+/// The handlers one worker process serves. A process may serve several job
+/// types, which is why WorkerHello carries a list of capabilities.
+#[derive(Default, Clone)]
+pub struct Registry {
+    handlers: HashMap<String, Arc<dyn JobHandler>>,
+    previews: HashMap<String, Arc<dyn PreviewProvider>>,
+}
+
+impl Registry {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn register(mut self, handler: Arc<dyn JobHandler>) -> Self {
+        self.handlers.insert(handler.capability().job_type, handler);
+        self
+    }
+
+    pub fn with_preview(mut self, provider: Arc<dyn PreviewProvider>) -> Self {
+        self.previews
+            .insert(provider.format().to_ascii_uppercase(), provider);
+        self
+    }
+
+    pub fn get(&self, job_type: &str) -> Option<Arc<dyn JobHandler>> {
+        self.handlers.get(job_type).cloned()
+    }
+
+    pub fn preview_provider(&self, format: &str) -> Option<Arc<dyn PreviewProvider>> {
+        self.previews.get(&format.to_ascii_uppercase()).cloned()
+    }
+
+    pub fn capabilities(&self) -> Vec<JobTypeCapability> {
+        self.handlers.values().map(|h| h.capability()).collect()
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.handlers.is_empty()
+    }
+}

seaweed-worker/crates/core/src/senders.rs

@@ -0,0 +1,203 @@
+use anyhow::Result;
+use tokio::sync::mpsc;
+
+use crate::pb::{
+    ActivityEvent, DetectionComplete, DetectionProposals, JobCompleted, JobProgressUpdate,
+    WorkerObservations, WorkerToAdminMessage, worker_to_admin_message::Body,
+};
+
+/// Replies to one detection request.
+pub trait DetectionSender: Send + Sync {
+    fn send_proposals(&self, proposals: DetectionProposals) -> Result<()>;
+    fn send_complete(&self, complete: DetectionComplete) -> Result<()>;
+    fn send_activity(&self, activity: ActivityEvent) -> Result<()>;
+    /// Reports what the worker saw while deciding. Admin caches the last one
+    /// per object and serves it back for display; nothing is scheduled from it.
+    fn send_observations(&self, observations: WorkerObservations) -> Result<()>;
+}
+
+/// Replies to one execution request.
+pub trait ExecutionSender: Send + Sync {
+    fn send_progress(&self, progress: JobProgressUpdate) -> Result<()>;
+    fn send_completed(&self, completed: JobCompleted) -> Result<()>;
+}
+
+/// Both senders write to the single outbound stream, so they share one channel.
+#[derive(Clone)]
+pub struct StreamSender {
+    worker_id: String,
+    tx: mpsc::UnboundedSender<WorkerToAdminMessage>,
+}
+
+impl StreamSender {
+    pub fn new(worker_id: String, tx: mpsc::UnboundedSender<WorkerToAdminMessage>) -> Self {
+        Self { worker_id, tx }
+    }
+
+    pub fn send(&self, body: Body) -> Result<()> {
+        self.tx.send(WorkerToAdminMessage {
+            worker_id: self.worker_id.clone(),
+            sent_at: Some(std::time::SystemTime::now().into()),
+            body: Some(body),
+        })?;
+        Ok(())
+    }
+}
+
+impl DetectionSender for StreamSender {
+    fn send_proposals(&self, proposals: DetectionProposals) -> Result<()> {
+        self.send(Body::DetectionProposals(proposals))
+    }
+
+    fn send_complete(&self, complete: DetectionComplete) -> Result<()> {
+        self.send(Body::DetectionComplete(complete))
+    }
+
+    fn send_activity(&self, _activity: ActivityEvent) -> Result<()> {
+        // Activity rides inside progress and completion messages rather than
+        // being a body of its own, so there is nothing to send on its own here.
+        Ok(())
+    }
+
+    fn send_observations(&self, observations: WorkerObservations) -> Result<()> {
+        self.send(Body::Observations(observations))
+    }
+}
+
+impl ExecutionSender for StreamSender {
+    fn send_progress(&self, progress: JobProgressUpdate) -> Result<()> {
+        self.send(Body::JobProgressUpdate(progress))
+    }
+
+    fn send_completed(&self, completed: JobCompleted) -> Result<()> {
+        self.send(Body::JobCompleted(completed))
+    }
+}
+
+/// Wraps a sender so what passes through it is counted. The handlers report
+/// their results to admin and nowhere else, so this is where a scrape can learn
+/// what happened without every handler having to know about metrics.
+pub struct MeteredSender<'a> {
+    inner: &'a StreamSender,
+    proposals: std::sync::atomic::AtomicUsize,
+    failed: std::sync::atomic::AtomicBool,
+}
+
+impl<'a> MeteredSender<'a> {
+    pub fn new(inner: &'a StreamSender) -> Self {
+        Self {
+            inner,
+            proposals: std::sync::atomic::AtomicUsize::new(0),
+            failed: std::sync::atomic::AtomicBool::new(false),
+        }
+    }
+
+    /// How many proposals went out, for the detection counter.
+    pub fn proposals(&self) -> usize {
+        self.proposals.load(std::sync::atomic::Ordering::Relaxed)
+    }
+
+    /// Whether the handler reported a failure of its own. A handler that fails
+    /// by returning an error is counted by the caller; this catches the one that
+    /// reports failure and returns Ok.
+    pub fn reported_failure(&self) -> bool {
+        self.failed.load(std::sync::atomic::Ordering::Relaxed)
+    }
+}
+
+impl DetectionSender for MeteredSender<'_> {
+    fn send_proposals(&self, proposals: DetectionProposals) -> Result<()> {
+        // Counted after the send, not before: a stream that closed mid-sweep
+        // would otherwise leave proposals_total claiming work admin never saw.
+        let count = proposals.proposals.len();
+        self.inner.send_proposals(proposals)?;
+        self.proposals
+            .fetch_add(count, std::sync::atomic::Ordering::Relaxed);
+        Ok(())
+    }
+
+    fn send_complete(&self, complete: DetectionComplete) -> Result<()> {
+        if !complete.success {
+            self.failed
+                .store(true, std::sync::atomic::Ordering::Relaxed);
+        }
+        self.inner.send_complete(complete)
+    }
+
+    fn send_activity(&self, activity: ActivityEvent) -> Result<()> {
+        self.inner.send_activity(activity)
+    }
+
+    fn send_observations(&self, observations: WorkerObservations) -> Result<()> {
+        self.inner.send_observations(observations)
+    }
+}
+
+impl ExecutionSender for MeteredSender<'_> {
+    fn send_progress(&self, progress: JobProgressUpdate) -> Result<()> {
+        self.inner.send_progress(progress)
+    }
+
+    fn send_completed(&self, completed: JobCompleted) -> Result<()> {
+        if !completed.success {
+            self.failed
+                .store(true, std::sync::atomic::Ordering::Relaxed);
+        }
+        self.inner.send_completed(completed)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::pb::JobProposal;
+
+    fn proposals(n: usize) -> DetectionProposals {
+        DetectionProposals {
+            proposals: (0..n).map(|_| JobProposal::default()).collect(),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn proposals_are_counted_once_they_are_sent() {
+        let (tx, mut rx) = mpsc::unbounded_channel();
+        let stream = StreamSender::new("worker-1".to_string(), tx);
+        let metered = MeteredSender::new(&stream);
+
+        metered.send_proposals(proposals(3)).expect("send");
+        assert_eq!(metered.proposals(), 3);
+        assert!(rx.try_recv().is_ok(), "the proposals reached the stream");
+    }
+
+    // Admin never saw these, so counting them would report work that was not
+    // handed over.
+    #[test]
+    fn proposals_are_not_counted_when_the_stream_is_gone() {
+        let (tx, rx) = mpsc::unbounded_channel();
+        let stream = StreamSender::new("worker-1".to_string(), tx);
+        let metered = MeteredSender::new(&stream);
+        drop(rx);
+
+        assert!(metered.send_proposals(proposals(3)).is_err());
+        assert_eq!(metered.proposals(), 0);
+    }
+
+    // A handler can report failure and still return Ok; the outcome has to come
+    // from what it said, not only from what it returned.
+    #[test]
+    fn a_reported_failure_is_remembered() {
+        let (tx, _rx) = mpsc::unbounded_channel();
+        let stream = StreamSender::new("worker-1".to_string(), tx);
+        let metered = MeteredSender::new(&stream);
+        assert!(!metered.reported_failure());
+
+        metered
+            .send_complete(DetectionComplete {
+                success: false,
+                ..Default::default()
+            })
+            .expect("send");
+        assert!(metered.reported_failure());
+    }
+}

seaweed-worker/crates/core/src/stream.rs

@@ -0,0 +1,491 @@
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+use anyhow::{Context, Result, anyhow};
+use tokio::sync::{Semaphore, mpsc};
+use tokio_stream::wrappers::UnboundedReceiverStream;
+use tonic::transport::{Certificate, Channel, ClientTlsConfig, Identity};
+use tracing::{info, warn};
+
+use crate::config::WorkerOptions;
+use crate::metrics::Metrics;
+use crate::pb::{
+    ConfigSchemaResponse, ExecuteJobRequest, JobCompleted, ObjectPreviewResponse, PreviewRow,
+    RequestObjectPreview, RunDetectionRequest, RunningWork, WorkerHeartbeat, WorkerHello,
+    admin_to_worker_message::Body as AdminBody,
+    plugin_control_service_client::PluginControlServiceClient,
+    worker_to_admin_message::Body as WorkerBody,
+};
+use crate::registry::Registry;
+use crate::senders::{MeteredSender, StreamSender};
+
+/// The protocol version this worker speaks, sent in WorkerHello.
+const PROTOCOL_VERSION: &str = "1";
+
+/// Connect to admin and serve the registry until the context is cancelled,
+/// reconnecting on failure. The stream is the only channel: everything admin
+/// asks for and everything the worker reports flows through it.
+pub async fn run(options: WorkerOptions, registry: Registry) -> Result<()> {
+    let metrics = Metrics::new(&options.worker_id, &options.worker_version)?;
+    run_with_metrics(options, registry, metrics).await
+}
+
+/// Runs with metrics a caller has already made, so it can serve them and add
+/// collectors of its own before the stream starts.
+pub async fn run_with_metrics(
+    options: WorkerOptions,
+    registry: Registry,
+    metrics: Metrics,
+) -> Result<()> {
+    if registry.is_empty() {
+        return Err(anyhow!("no job handlers registered"));
+    }
+    if options.max_detection_concurrency < 1 || options.max_execution_concurrency < 1 {
+        return Err(anyhow!(
+            "concurrency limits must be at least 1, got detection={} execution={}",
+            options.max_detection_concurrency,
+            options.max_execution_concurrency
+        ));
+    }
+    let slots = Slots::new(&options);
+    metrics.set_slots("detection", 0, slots.detection_total as i64);
+    metrics.set_slots("execution", 0, slots.execution_total as i64);
+    loop {
+        match serve_once(&options, &registry, &slots, &metrics).await {
+            Err(err) => {
+                metrics.stream_ended("failed");
+                warn!("worker stream ended: {err:#}")
+            }
+            // Admin asked this worker to stop, so stop. Reconnecting here would
+            // make shutdown impossible: the worker would log back in.
+            Ok(Outcome::ShutdownRequested) => {
+                metrics.stream_ended("shutdown");
+                return Ok(());
+            }
+            // Admin closing a healthy stream is not an error, but reconnecting
+            // in silence hides the reason - two workers sharing an id evict
+            // each other and produce nothing but a login every few seconds.
+            Ok(Outcome::StreamClosed) => {
+                metrics.stream_ended("closed");
+                warn!(
+                    "admin closed the stream; reconnecting in {:?}. If this repeats, check for \
+                     another worker using the id {}",
+                    options.reconnect_delay, options.worker_id
+                )
+            }
+        }
+        tokio::time::sleep(options.reconnect_delay).await;
+    }
+}
+
+/// Why a stream ended. Only one of these means "do not come back".
+enum Outcome {
+    StreamClosed,
+    ShutdownRequested,
+}
+
+/// The capacity this worker advertises in WorkerHello. Admin schedules against
+/// those numbers, so the worker has to actually hold to them - and the heartbeat
+/// has to report what is in use, or admin is scheduling blind.
+#[derive(Clone)]
+struct Slots {
+    detection: Arc<Semaphore>,
+    execution: Arc<Semaphore>,
+    detection_total: i32,
+    execution_total: i32,
+}
+
+impl Slots {
+    fn new(options: &WorkerOptions) -> Self {
+        Self {
+            detection: Arc::new(Semaphore::new(options.max_detection_concurrency as usize)),
+            execution: Arc::new(Semaphore::new(options.max_execution_concurrency as usize)),
+            detection_total: options.max_detection_concurrency,
+            execution_total: options.max_execution_concurrency,
+        }
+    }
+
+    fn detection_used(&self) -> i32 {
+        self.detection_total - self.detection.available_permits() as i32
+    }
+
+    fn execution_used(&self) -> i32 {
+        self.execution_total - self.execution.available_permits() as i32
+    }
+}
+
+/// Dials admin, over mTLS when certificates are configured. Plaintext is the
+/// default and is fine over loopback; anything else carries preview rows and
+/// execution commands in the clear, and a cluster with grpc TLS on refuses the
+/// connection anyway.
+async fn connect(options: &WorkerOptions, grpc_address: &str) -> Result<Channel> {
+    let Some(tls) = options.tls.as_ref() else {
+        return Ok(Channel::from_shared(format!("http://{grpc_address}"))?
+            .connect_timeout(Duration::from_secs(10))
+            .connect()
+            .await?);
+    };
+
+    let ca = tokio::fs::read(&tls.ca_path)
+        .await
+        .with_context(|| format!("read CA certificate {}", tls.ca_path))?;
+    let cert = tokio::fs::read(&tls.client_cert_path)
+        .await
+        .with_context(|| format!("read client certificate {}", tls.client_cert_path))?;
+    let key = tokio::fs::read(&tls.client_key_path)
+        .await
+        .with_context(|| format!("read client key {}", tls.client_key_path))?;
+
+    let mut config = ClientTlsConfig::new()
+        .ca_certificate(Certificate::from_pem(ca))
+        .identity(Identity::from_pem(cert, key));
+    if let Some(server_name) = tls.server_name.as_ref() {
+        config = config.domain_name(server_name.clone());
+    }
+
+    Ok(Channel::from_shared(format!("https://{grpc_address}"))?
+        .tls_config(config)?
+        .connect_timeout(Duration::from_secs(10))
+        .connect()
+        .await?)
+}
+
+async fn serve_once(
+    options: &WorkerOptions,
+    registry: &Registry,
+    slots: &Slots,
+    metrics: &Metrics,
+) -> Result<Outcome> {
+    // Operators give the admin's HTTP address; the gRPC port is derived, the
+    // same way the Go worker does it.
+    let grpc_address = crate::address::server_to_grpc_address(&options.admin_address)
+        .ok_or_else(|| anyhow!("cannot parse admin address {}", options.admin_address))?;
+    let channel = connect(options, &grpc_address).await?;
+    let mut client = PluginControlServiceClient::new(channel);
+
+    let (tx, rx) = mpsc::unbounded_channel();
+    let sender = StreamSender::new(options.worker_id.clone(), tx);
+
+    sender.send(WorkerBody::Hello(WorkerHello {
+        worker_id: options.worker_id.clone(),
+        worker_instance_id: options.worker_id.clone(),
+        address: options.worker_address.clone(),
+        worker_version: options.worker_version.clone(),
+        protocol_version: PROTOCOL_VERSION.to_string(),
+        capabilities: registry.capabilities(),
+        metadata: Default::default(),
+    }))?;
+
+    let mut inbound = client
+        .worker_stream(UnboundedReceiverStream::new(rx))
+        .await?
+        .into_inner();
+
+    let heartbeat = spawn_heartbeat(
+        sender.clone(),
+        options.clone(),
+        slots.clone(),
+        metrics.clone(),
+    );
+
+    while let Some(message) = inbound.message().await? {
+        let request_id = message.request_id.clone();
+        match message.body {
+            Some(AdminBody::Hello(hello)) => {
+                if !hello.accepted {
+                    return Err(anyhow!("admin rejected this worker: {}", hello.message));
+                }
+                metrics.stream_connected();
+                info!(
+                    "connected to admin at {} ({})",
+                    options.admin_address, grpc_address
+                );
+            }
+            Some(AdminBody::RequestConfigSchema(request)) => {
+                let response = match registry.get(&request.job_type) {
+                    Some(handler) => ConfigSchemaResponse {
+                        request_id: request_id.clone(),
+                        job_type: request.job_type.clone(),
+                        success: true,
+                        error_message: String::new(),
+                        job_type_descriptor: Some(handler.descriptor()),
+                    },
+                    None => ConfigSchemaResponse {
+                        request_id: request_id.clone(),
+                        job_type: request.job_type.clone(),
+                        success: false,
+                        error_message: format!("unknown job type {}", request.job_type),
+                        job_type_descriptor: None,
+                    },
+                };
+                sender.send(WorkerBody::ConfigSchemaResponse(response))?;
+            }
+            Some(AdminBody::RequestObjectPreview(request)) => {
+                spawn_preview(
+                    registry.clone(),
+                    sender.clone(),
+                    metrics.clone(),
+                    request_id.clone(),
+                    request,
+                );
+            }
+            Some(AdminBody::RunDetectionRequest(request)) => {
+                spawn_detection(
+                    registry.clone(),
+                    sender.clone(),
+                    slots.clone(),
+                    metrics.clone(),
+                    request,
+                );
+            }
+            Some(AdminBody::ExecuteJobRequest(request)) => {
+                spawn_execution(
+                    registry.clone(),
+                    sender.clone(),
+                    slots.clone(),
+                    metrics.clone(),
+                    request,
+                );
+            }
+            Some(AdminBody::CancelRequest(request)) => {
+                // Cancellation needs a per-request handle to be honoured; until
+                // then say so rather than silently continuing to run the job.
+                warn!(
+                    "cancel requested for {} ({}) but is not implemented",
+                    request.target_id, request.reason
+                );
+            }
+            Some(AdminBody::Shutdown(shutdown)) => {
+                info!("admin asked this worker to stop: {}", shutdown.reason);
+                heartbeat.abort();
+                return Ok(Outcome::ShutdownRequested);
+            }
+            None => {}
+        }
+    }
+
+    heartbeat.abort();
+    Ok(Outcome::StreamClosed)
+}
+
+/// Answers one preview request off the stream loop. Reading rows takes as long
+/// as it takes, and the stream has heartbeats to keep up meanwhile.
+fn spawn_preview(
+    registry: Registry,
+    sender: StreamSender,
+    metrics: Metrics,
+    request_id: String,
+    request: RequestObjectPreview,
+) {
+    tokio::spawn(async move {
+        let limit = request.row_limit.max(1) as usize;
+        let response = match registry.preview_provider(&request.format) {
+            None => ObjectPreviewResponse {
+                request_id,
+                success: false,
+                error_message: format!("this worker does not read {} objects", request.format),
+                ..Default::default()
+            },
+            Some(provider) => match provider.preview(&request.object_id, limit).await {
+                Ok(preview) => ObjectPreviewResponse {
+                    request_id,
+                    success: true,
+                    error_message: String::new(),
+                    columns: preview.columns,
+                    rows: preview
+                        .rows
+                        .into_iter()
+                        .map(|values| PreviewRow { values })
+                        .collect(),
+                    total_rows: preview.total_rows,
+                },
+                Err(err) => ObjectPreviewResponse {
+                    request_id,
+                    success: false,
+                    error_message: format!("{err:#}"),
+                    ..Default::default()
+                },
+            },
+        };
+        metrics.preview_finished(if response.success { "ok" } else { "failed" });
+        let _ = sender.send(WorkerBody::ObjectPreviewResponse(response));
+    });
+}
+
+fn spawn_heartbeat(
+    sender: StreamSender,
+    options: WorkerOptions,
+    slots: Slots,
+    metrics: Metrics,
+) -> tokio::task::JoinHandle<()> {
+    // The handle has to be the heartbeat's own, or aborting it aborts nothing
+    // and every reconnect leaves another ticker running.
+    tokio::spawn(async move {
+        let mut ticker = tokio::time::interval(options.heartbeat_interval);
+        loop {
+            ticker.tick().await;
+            // The heartbeat already computes this for admin; publish the same
+            // numbers so a scrape and the admin UI cannot disagree.
+            metrics.set_slots(
+                "detection",
+                slots.detection_used() as i64,
+                slots.detection_total as i64,
+            );
+            metrics.set_slots(
+                "execution",
+                slots.execution_used() as i64,
+                slots.execution_total as i64,
+            );
+            let beat = WorkerHeartbeat {
+                worker_id: options.worker_id.clone(),
+                running_work: Vec::<RunningWork>::new(),
+                detection_slots_used: slots.detection_used(),
+                detection_slots_total: slots.detection_total,
+                execution_slots_used: slots.execution_used(),
+                execution_slots_total: slots.execution_total,
+                queued_jobs_by_type: Default::default(),
+                metadata: Default::default(),
+            };
+            if sender.send(WorkerBody::Heartbeat(beat)).is_err() {
+                return;
+            }
+        }
+    })
+}
+
+fn spawn_detection(
+    registry: Registry,
+    sender: StreamSender,
+    slots: Slots,
+    metrics: Metrics,
+    request: RunDetectionRequest,
+) {
+    tokio::spawn(async move {
+        let Some(handler) = registry.get(&request.job_type) else {
+            return;
+        };
+        // Held until the sweep finishes, so the worker keeps to the capacity it
+        // advertised and the heartbeat reports the truth while it works.
+        let _permit = slots.detection.acquire().await;
+        let metered = MeteredSender::new(&sender);
+        let started = Instant::now();
+        let outcome = handler.detect(&request, &metered).await;
+        let result = if let Err(err) = &outcome {
+            warn!("detection for {} failed: {err:#}", request.job_type);
+            let _ = sender.send(WorkerBody::DetectionComplete(
+                crate::pb::DetectionComplete {
+                    request_id: request.request_id.clone(),
+                    job_type: request.job_type.clone(),
+                    success: false,
+                    error_message: format!("{err:#}"),
+                    total_proposals: 0,
+                },
+            ));
+            "failed"
+        } else if metered.reported_failure() {
+            "failed"
+        } else {
+            "ok"
+        };
+        metrics.detection_finished(
+            &request.job_type,
+            result,
+            started.elapsed().as_secs_f64(),
+            metered.proposals(),
+        );
+    });
+}
+
+fn spawn_execution(
+    registry: Registry,
+    sender: StreamSender,
+    slots: Slots,
+    metrics: Metrics,
+    request: ExecuteJobRequest,
+) {
+    tokio::spawn(async move {
+        let _permit = slots.execution.acquire().await;
+        let job_type = request
+            .job
+            .as_ref()
+            .map(|job| job.job_type.clone())
+            .unwrap_or_default();
+        let job_id = request
+            .job
+            .as_ref()
+            .map(|job| job.job_id.clone())
+            .unwrap_or_default();
+        let Some(handler) = registry.get(&job_type) else {
+            return;
+        };
+        let metered = MeteredSender::new(&sender);
+        let started = Instant::now();
+        let outcome = handler.execute(&request, &metered).await;
+        let result = if let Err(err) = &outcome {
+            warn!("job {job_id} failed: {err:#}");
+            let _ = sender.send(WorkerBody::JobCompleted(JobCompleted {
+                request_id: request.request_id.clone(),
+                job_id,
+                job_type: job_type.clone(),
+                success: false,
+                error_message: format!("{err:#}"),
+                ..Default::default()
+            }));
+            "failed"
+        } else if metered.reported_failure() {
+            "failed"
+        } else {
+            "ok"
+        };
+        metrics.job_finished(&job_type, result, started.elapsed().as_secs_f64());
+    });
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn options(detection: i32, execution: i32) -> WorkerOptions {
+        WorkerOptions {
+            max_detection_concurrency: detection,
+            max_execution_concurrency: execution,
+            ..Default::default()
+        }
+    }
+
+    // The heartbeat is admin's only view of how busy this worker is; before the
+    // permits existed it reported zero however much was running.
+    #[tokio::test]
+    async fn slots_report_what_is_held() {
+        let slots = Slots::new(&options(2, 3));
+        assert_eq!(slots.detection_used(), 0);
+        assert_eq!(slots.execution_used(), 0);
+
+        let held = slots.detection.acquire().await.unwrap();
+        assert_eq!(slots.detection_used(), 1);
+        assert_eq!(slots.execution_used(), 0, "the two do not share capacity");
+
+        drop(held);
+        assert_eq!(slots.detection_used(), 0);
+    }
+
+    // A limit of one means the second request waits, rather than running anyway
+    // as it did when every request simply spawned a task.
+    #[tokio::test]
+    async fn a_full_lane_makes_the_next_request_wait() {
+        let slots = Slots::new(&options(1, 1));
+        let held = slots.execution.clone().acquire_owned().await.unwrap();
+        assert_eq!(slots.execution_used(), 1);
+
+        let waiter = tokio::spawn({
+            let execution = slots.execution.clone();
+            async move { execution.acquire_owned().await.unwrap() }
+        });
+        tokio::task::yield_now().await;
+        assert!(!waiter.is_finished(), "the second request must not start");
+
+        drop(held);
+        let _second = waiter.await.expect("the waiter should be handed the slot");
+        assert_eq!(slots.execution_used(), 1);
+    }
+}

seaweed-worker/crates/lance/Cargo.toml

@@ -0,0 +1,55 @@
+[package]
+name = "weed-lance-worker"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+description = "SeaweedFS maintenance worker for Lance tables"
+
+[lib]
+name = "weed_lance_worker"
+
+# The binary is not named for lance: it is the Rust side of `weed worker`, and
+# the job families it registers will outgrow this crate. When a second one
+# arrives the bin target moves to a crate of its own under the same name.
+[[bin]]
+name = "weed-worker"
+path = "src/main.rs"
+
+[dependencies]
+seaweed-worker-core = { path = "../core" }
+seaweed-worker-sort = { path = "../sort" }
+prometheus.workspace = true
+# Only the S3 backend: the other object stores lance enables by default are
+# build time this worker never spends.
+lance = { version = "10", default-features = false, features = ["aws"] }
+lance-index = "10"
+# For the manifest's own fragment type, which the sort marker summarises.
+lance-table = "10"
+# For the execution options the sort needs; the scanner's own stream helper
+# takes the default, which has spilling off.
+lance-datafusion = "10"
+arrow-schema = "58"
+arrow-cast = "58"
+chrono = "0.4"
+futures = "0.3"
+anyhow.workspace = true
+async-trait.workspace = true
+clap = { version = "4", features = ["derive", "env"] }
+reqwest = { version = "0.12", default-features = false, features = ["json", "rustls-tls"] }
+rustls = "0.23"
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+tokio.workspace = true
+tracing.workspace = true
+tracing-subscriber.workspace = true
+
+[dev-dependencies]
+tokio = { workspace = true }
+tonic.workspace = true
+arrow-array = "58"
+arrow-schema = "58"
+arrow-cast = "58"
+lance-linalg = "10"
+
+[lints]
+workspace = true

seaweed-worker/crates/lance/src/catalog/mod.rs

@@ -0,0 +1,10 @@
+//! How the worker finds Lance tables and gets at their bytes.
+//!
+//! It goes through the Lance namespace rather than the filer: the namespace is
+//! the catalog of record, it already knows which tables are Lance, and asking it
+//! to describe a table with vend_credentials is how the worker gets storage
+//! credentials without holding any of its own.
+
+pub mod namespace;
+
+pub use namespace::{NamespaceClient, TableDescription, parse_id};

seaweed-worker/crates/lance/src/catalog/namespace.rs

@@ -0,0 +1,129 @@
+use std::collections::HashMap;
+
+use std::time::Duration;
+
+use anyhow::{Context, Result};
+use serde::Deserialize;
+
+/// The delimiter the Lance namespace joins identifier parts with.
+const DELIMITER: &str = "$";
+
+#[derive(Debug, Deserialize)]
+pub struct TableDescription {
+    pub location: String,
+    #[serde(default)]
+    pub storage_options: HashMap<String, String>,
+    #[serde(default)]
+    pub managed_versioning: bool,
+}
+
+#[derive(Debug, Deserialize)]
+struct ListNamespacesResponse {
+    #[serde(default)]
+    namespaces: Vec<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct ListTablesResponse {
+    #[serde(default)]
+    tables: Vec<String>,
+}
+
+/// A thin client for the operations this worker needs. It deliberately does not
+/// wrap the whole spec: a maintenance worker lists, describes, and commits.
+pub struct NamespaceClient {
+    base_url: String,
+    http: reqwest::Client,
+}
+
+/// A namespace call that has not answered by now is not going to. Without this
+/// a gateway that accepts the connection and then goes quiet holds a detection
+/// slot open forever, and the sweep never finishes.
+const REQUEST_TIMEOUT: Duration = Duration::from_secs(30);
+const CONNECT_TIMEOUT: Duration = Duration::from_secs(10);
+
+impl NamespaceClient {
+    pub fn new(base_url: impl Into<String>) -> Self {
+        let http = reqwest::Client::builder()
+            .timeout(REQUEST_TIMEOUT)
+            .connect_timeout(CONNECT_TIMEOUT)
+            .build()
+            // The builder only fails on a bad TLS backend, which would break
+            // every call anyway; a client with no timeouts is worse than a panic
+            // at startup, so keep the default only as a last resort.
+            .unwrap_or_else(|err| {
+                tracing::warn!("falling back to an untimed HTTP client: {err}");
+                reqwest::Client::new()
+            });
+        Self {
+            base_url: base_url.into().trim_end_matches('/').to_string(),
+            http,
+        }
+    }
+
+    /// Every table the namespace holds, as delimiter-joined identifiers.
+    pub async fn list_all_tables(&self) -> Result<Vec<String>> {
+        let url = format!("{}/v1/table", self.base_url);
+        let response: ListTablesResponse = self
+            .http
+            .get(&url)
+            .send()
+            .await
+            .context("list tables")?
+            .error_for_status()?
+            .json()
+            .await?;
+        Ok(response.tables)
+    }
+
+    /// Child namespaces of `id`; the root lists table buckets.
+    pub async fn list_namespaces(&self, id: &[String]) -> Result<Vec<String>> {
+        let url = format!("{}/v1/namespace/{}/list", self.base_url, encode_id(id));
+        let response: ListNamespacesResponse = self
+            .http
+            .get(&url)
+            .send()
+            .await
+            .context("list namespaces")?
+            .error_for_status()?
+            .json()
+            .await?;
+        Ok(response.namespaces)
+    }
+
+    /// Resolve a table to a location and the credentials to reach it. The
+    /// credentials expire, so a long compaction re-describes rather than
+    /// carrying one set for the whole job.
+    pub async fn describe_table(&self, id: &[String]) -> Result<TableDescription> {
+        let url = format!("{}/v1/table/{}/describe", self.base_url, encode_id(id));
+        let body = serde_json::json!({ "id": id, "vend_credentials": true });
+        let description: TableDescription = self
+            .http
+            .post(&url)
+            .json(&body)
+            .send()
+            .await
+            .context("describe table")?
+            .error_for_status()?
+            .json()
+            .await?;
+        Ok(description)
+    }
+}
+
+fn encode_id(id: &[String]) -> String {
+    if id.is_empty() {
+        DELIMITER.to_string()
+    } else {
+        id.join(DELIMITER)
+    }
+}
+
+/// Splits a delimiter-joined identifier back into parts.
+pub fn parse_id(encoded: &str) -> Vec<String> {
+    encoded
+        .split(DELIMITER)
+        .filter(|part| !part.is_empty())
+        .map(|part| part.to_string())
+        .collect()
+}

seaweed-worker/crates/lance/src/dataset.rs

@@ -0,0 +1,98 @@
+//! Opening a Lance dataset with credentials the namespace vended.
+//!
+//! The worker holds no storage credentials of its own: it asks the namespace to
+//! describe a table with `vend_credentials`, and the `storage_options` that come
+//! back are handed to lance as-is. They expire, so a job that runs longer than
+//! their lifetime re-describes rather than carrying one set throughout.
+
+use std::collections::HashMap;
+
+use anyhow::{Context, Result};
+use lance::dataset::Dataset;
+use lance::dataset::builder::DatasetBuilder;
+
+use crate::catalog::{NamespaceClient, TableDescription};
+
+/// A table the worker is about to work on.
+pub struct OpenTable {
+    pub id: Vec<String>,
+    pub location: String,
+    pub dataset: Dataset,
+}
+
+/// What detection needs to decide whether a table is worth a job. Reading it
+/// opens the dataset but touches no data files.
+pub struct TableStats {
+    pub fragments: usize,
+    pub version: u64,
+    pub total_versions: usize,
+    pub rows: usize,
+    /// The Arrow schema as JSON, which is the only description of this table
+    /// anything outside the format can produce.
+    pub schema: Option<String>,
+}
+
+/// Storage options an operator supplies for deployments that vend none.
+pub type FallbackOptions = HashMap<String, String>;
+
+/// Resolve a table through the namespace and open it.
+pub async fn open(
+    client: &NamespaceClient,
+    id: &[String],
+    fallback: &FallbackOptions,
+) -> Result<OpenTable> {
+    let description = client.describe_table(id).await?;
+    let dataset = open_at(&description, fallback).await?;
+    Ok(OpenTable {
+        id: id.to_vec(),
+        location: description.location,
+        dataset,
+    })
+}
+
+/// The namespace vends object_store's own option names, so they pass straight
+/// through. What it vends always wins: the fallback exists because a deployment
+/// without STS vends no credentials at all, and then the worker has no other way
+/// to reach the data.
+async fn open_at(description: &TableDescription, fallback: &FallbackOptions) -> Result<Dataset> {
+    let mut options: HashMap<String, String> = fallback.clone();
+    options.extend(description.storage_options.clone());
+    DatasetBuilder::from_uri(&description.location)
+        .with_storage_options(options)
+        .load()
+        .await
+        .with_context(|| format!("open lance dataset at {}", description.location))
+}
+
+impl OpenTable {
+    pub async fn stats(&self) -> Result<TableStats> {
+        let versions = self.dataset.versions().await?;
+        Ok(TableStats {
+            fragments: self.dataset.get_fragments().len(),
+            version: self.dataset.version().version,
+            total_versions: versions.len(),
+            rows: self.dataset.count_rows(None).await.unwrap_or(0),
+            schema: schema_json(&self.dataset),
+        })
+    }
+}
+
+/// Renders the dataset's schema as JSON. Best effort: a schema that will not
+/// serialise is not a reason to fail a maintenance sweep.
+fn schema_json(dataset: &Dataset) -> Option<String> {
+    let arrow: arrow_schema::Schema = dataset.schema().into();
+    serde_json::to_string(
+        &arrow
+            .fields()
+            .iter()
+            .map(|f| {
+                serde_json::json!({
+                    "name": f.name(),
+                    "type": f.data_type().to_string(),
+                    "nullable": f.is_nullable(),
+                })
+            })
+            .collect::<Vec<_>>(),
+    )
+    .ok()
+}

seaweed-worker/crates/lance/src/jobs/cleanup.rs

@@ -0,0 +1,336 @@
+use std::collections::HashMap;
+
+use anyhow::{Context, Result, anyhow};
+use async_trait::async_trait;
+use chrono::{Duration, Utc};
+use lance::dataset::cleanup::{CleanupPolicy, cleanup_old_versions};
+use seaweed_worker_core::config_form::{form, int_or, int_value, number_field};
+use seaweed_worker_core::pb::{
+    ConfigValue, DetectionComplete, DetectionProposals, ExecuteJobRequest, JobCompleted,
+    JobProgressUpdate, JobProposal, JobResult, JobTypeCapability, JobTypeDescriptor,
+    RunDetectionRequest,
+};
+use seaweed_worker_core::{DetectionSender, ExecutionSender, JobHandler};
+use tracing::warn;
+
+use crate::catalog::{NamespaceClient, parse_id};
+use crate::dataset;
+use crate::jobs::{clamp, string_list, table_id};
+
+pub const JOB_TYPE: &str = "lance_cleanup_versions";
+
+const DEFAULT_RETAIN_HOURS: i64 = 168;
+const DEFAULT_MIN_VERSIONS: i64 = 5;
+
+// The ranges the form offers. Duration::hours also panics far outside this one.
+const MAX_RETAIN_HOURS: i64 = 8760;
+const MAX_MIN_VERSIONS: i64 = 1000;
+
+/// Lance keeps every version until something removes it. Lance can also do this
+/// itself through auto-cleanup, so this job is for deployments that would rather
+/// the cluster owned the policy than each writer.
+pub struct CleanupVersionsHandler {
+    namespace_url: String,
+    fallback: dataset::FallbackOptions,
+    metrics: Option<crate::metrics::LanceMetrics>,
+}
+
+impl CleanupVersionsHandler {
+    pub fn new(namespace_url: String) -> Self {
+        Self {
+            namespace_url,
+            fallback: dataset::FallbackOptions::new(),
+            metrics: None,
+        }
+    }
+
+    pub fn with_metrics(mut self, metrics: Option<crate::metrics::LanceMetrics>) -> Self {
+        self.metrics = metrics;
+        self
+    }
+
+    pub fn with_fallback(mut self, fallback: dataset::FallbackOptions) -> Self {
+        self.fallback = fallback;
+        self
+    }
+
+    fn client(&self) -> NamespaceClient {
+        NamespaceClient::new(self.namespace_url.clone())
+    }
+}
+
+/// The oldest version that may be removed while still leaving `min_versions`
+/// behind, or None when the table has no more than the floor.
+async fn version_floor(table: &dataset::OpenTable, min_versions: usize) -> Result<Option<u64>> {
+    let mut versions: Vec<u64> = table
+        .dataset
+        .versions()
+        .await?
+        .iter()
+        .map(|version| version.version)
+        .collect();
+    if versions.len() <= min_versions {
+        return Ok(None);
+    }
+    versions.sort_unstable();
+    Ok(Some(versions[versions.len() - min_versions]))
+}
+
+#[async_trait]
+impl JobHandler for CleanupVersionsHandler {
+    fn capability(&self) -> JobTypeCapability {
+        JobTypeCapability {
+            job_type: JOB_TYPE.to_string(),
+            can_detect: true,
+            can_execute: true,
+            max_detection_concurrency: 1,
+            max_execution_concurrency: 1,
+            display_name: "Lance Version Cleanup".to_string(),
+            description: "Remove old Lance versions and the files only they referenced".to_string(),
+            weight: 10,
+        }
+    }
+
+    fn descriptor(&self) -> JobTypeDescriptor {
+        let mut defaults: HashMap<String, ConfigValue> = HashMap::new();
+        defaults.insert("retain_hours".to_string(), int_value(DEFAULT_RETAIN_HOURS));
+        defaults.insert(
+            "min_versions_to_keep".to_string(),
+            int_value(DEFAULT_MIN_VERSIONS),
+        );
+
+        JobTypeDescriptor {
+            job_type: JOB_TYPE.to_string(),
+            display_name: "Lance Version Cleanup".to_string(),
+            description: "Age out Lance versions the table no longer needs".to_string(),
+            icon: "fas fa-broom".to_string(),
+            descriptor_version: 1,
+            worker_config_form: Some(form(
+                "lance-cleanup-worker",
+                "Version cleanup",
+                vec![
+                    number_field(
+                        "retain_hours",
+                        "Retain versions for (hours)",
+                        "Versions younger than this are always kept",
+                        1,
+                        8760,
+                    ),
+                    number_field(
+                        "min_versions_to_keep",
+                        "Minimum versions",
+                        "Never leave a table with fewer versions than this, whatever their age",
+                        1,
+                        1000,
+                    ),
+                ],
+                defaults.clone(),
+            )),
+            worker_default_values: defaults,
+            ..Default::default()
+        }
+    }
+
+    async fn detect(
+        &self,
+        request: &RunDetectionRequest,
+        sender: &dyn DetectionSender,
+    ) -> Result<()> {
+        let min_versions = clamp(
+            int_or(
+                &request.worker_config_values,
+                "min_versions_to_keep",
+                DEFAULT_MIN_VERSIONS,
+            ),
+            1,
+            MAX_MIN_VERSIONS,
+        ) as usize;
+        let client = self.client();
+        let tables = client.list_all_tables().await?;
+
+        let mut proposals = Vec::new();
+        for encoded in &tables {
+            let id = parse_id(encoded);
+            let table = match dataset::open(&client, &id, &self.fallback).await {
+                Ok(table) => table,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "open");
+                    }
+                    warn!("skipping {encoded}: {err:#}");
+                    continue;
+                }
+            };
+            if let Some(counters) = &self.metrics {
+                counters.worker.object_seen(JOB_TYPE);
+            }
+            let stats = match table.stats().await {
+                Ok(stats) => stats,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "stats");
+                    }
+                    warn!("skipping {encoded}: reading its stats failed: {err:#}");
+                    continue;
+                }
+            };
+            // Age is decided at execution against the retention window; a table
+            // at or under the floor cannot lose a version whatever its age, so
+            // proposing one would only produce a job with nothing to do.
+            if stats.total_versions <= min_versions {
+                continue;
+            }
+            let mut parameters: HashMap<String, ConfigValue> = HashMap::new();
+            parameters.insert("table_id".to_string(), string_list(&id));
+            proposals.push(JobProposal {
+                proposal_id: format!("{JOB_TYPE}:{encoded}"),
+                dedupe_key: format!("{JOB_TYPE}:{encoded}"),
+                job_type: JOB_TYPE.to_string(),
+                summary: format!("Clean up {encoded} ({} versions)", stats.total_versions),
+                detail: format!(
+                    "{} versions retained, above the {min_versions} floor; \
+                     those outside the retention window can go",
+                    stats.total_versions
+                ),
+                parameters,
+                ..Default::default()
+            });
+        }
+
+        let total = proposals.len() as i32;
+        sender.send_proposals(DetectionProposals {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            proposals,
+            has_more: false,
+        })?;
+        sender.send_complete(DetectionComplete {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            error_message: String::new(),
+            total_proposals: total,
+        })?;
+        Ok(())
+    }
+
+    async fn execute(
+        &self,
+        request: &ExecuteJobRequest,
+        sender: &dyn ExecutionSender,
+    ) -> Result<()> {
+        let job = request
+            .job
+            .as_ref()
+            .ok_or_else(|| anyhow!("execute request carried no job"))?;
+        let id = table_id(&job.parameters)
+            .ok_or_else(|| anyhow!("job {} carried no table_id", job.job_id))?;
+        let retain_hours = clamp(
+            int_or(
+                &request.worker_config_values,
+                "retain_hours",
+                DEFAULT_RETAIN_HOURS,
+            ),
+            0,
+            MAX_RETAIN_HOURS,
+        );
+        let min_versions = clamp(
+            int_or(
+                &request.worker_config_values,
+                "min_versions_to_keep",
+                DEFAULT_MIN_VERSIONS,
+            ),
+            1,
+            MAX_MIN_VERSIONS,
+        ) as usize;
+
+        let client = self.client();
+        let table = dataset::open(&client, &id, &self.fallback).await?;
+        let before = table.stats().await?;
+
+        // The floor is a promise about how much history survives, so it has to
+        // be applied here and not only when the job was proposed: by the time it
+        // runs, versions may have aged past the retention window, and age alone
+        // would take the table below what the operator asked to keep.
+        let Some(floor) = version_floor(&table, min_versions).await? else {
+            sender.send_completed(JobCompleted {
+                request_id: request.request_id.clone(),
+                job_id: job.job_id.clone(),
+                job_type: JOB_TYPE.to_string(),
+                success: true,
+                result: Some(JobResult {
+                    summary: format!(
+                        "kept all {} versions; the {min_versions} version floor leaves none to remove",
+                        before.total_versions
+                    ),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            })?;
+            return Ok(());
+        };
+
+        sender.send_progress(JobProgressUpdate {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            progress_percent: 10.0,
+            stage: format!("cleaning up {} versions", before.total_versions),
+            ..Default::default()
+        })?;
+
+        let policy = CleanupPolicy {
+            before_timestamp: Some(Utc::now() - Duration::hours(retain_hours)),
+            // should_clean() ANDs its clauses, so a version has to be both older
+            // than the window and below the floor to go.
+            before_version: Some(floor),
+            // Files this dataset cannot account for are left alone: they may
+            // belong to a writer that has not committed yet, and deleting them
+            // would corrupt a commit in flight.
+            delete_unverified: false,
+            // A tagged version is pinned on purpose, so refuse rather than
+            // silently dropping what someone named.
+            error_if_tagged_old_versions: true,
+            ..Default::default()
+        };
+        let stats = cleanup_old_versions(&table.dataset, policy)
+            .await
+            .with_context(|| format!("clean up versions of {}", table.location))?;
+
+        if let Some(counters) = &self.metrics {
+            counters.versions_removed.inc_by(stats.old_versions as u64);
+            counters.bytes_reclaimed.inc_by(stats.bytes_removed as u64);
+        }
+
+        let mut output: HashMap<String, ConfigValue> = HashMap::new();
+        output.insert(
+            "old_versions_removed".to_string(),
+            int_value(stats.old_versions as i64),
+        );
+        output.insert(
+            "bytes_removed".to_string(),
+            int_value(stats.bytes_removed as i64),
+        );
+        output.insert(
+            "data_files_removed".to_string(),
+            int_value(stats.data_files_removed as i64),
+        );
+
+        sender.send_completed(JobCompleted {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            result: Some(JobResult {
+                output_values: output,
+                summary: format!(
+                    "removed {} versions and {} bytes",
+                    stats.old_versions, stats.bytes_removed
+                ),
+                ..Default::default()
+            }),
+            ..Default::default()
+        })?;
+        Ok(())
+    }
+}

seaweed-worker/crates/lance/src/jobs/compact.rs

@@ -0,0 +1,318 @@
+use std::collections::HashMap;
+
+use anyhow::{Context, Result, anyhow};
+use async_trait::async_trait;
+use lance::dataset::optimize::{CompactionOptions, compact_files};
+use seaweed_worker_core::config_form::{form, int_or, int_value, number_field, string_value};
+use seaweed_worker_core::pb::{
+    ConfigValue, DetectionComplete, DetectionProposals, ExecuteJobRequest, JobCompleted,
+    JobProgressUpdate, JobProposal, JobResult, JobTypeCapability, JobTypeDescriptor,
+    RunDetectionRequest, WorkerObservations,
+};
+use seaweed_worker_core::{DetectionSender, ExecutionSender, JobHandler};
+use tracing::warn;
+
+use crate::catalog::{NamespaceClient, parse_id};
+use crate::dataset;
+use crate::jobs::{FORMAT, clamp, observation, string_list, table_id};
+
+pub const JOB_TYPE: &str = "lance_compact";
+
+const DEFAULT_TARGET_ROWS: i64 = 1_048_576;
+const DEFAULT_MIN_FRAGMENTS: i64 = 8;
+
+// The ranges the descriptor's form offers. These are also the values that stay
+// meaningful: a table needs two fragments before merging them means anything,
+// and a fragment target below a thousand rows defeats the purpose of the job.
+// Both are cast to usize, where a negative would arrive as an enormous number.
+const TARGET_ROWS_FLOOR: i64 = 1024;
+const TARGET_ROWS_CEILING: i64 = 16_777_216;
+const MIN_FRAGMENTS_FLOOR: i64 = 2;
+const MIN_FRAGMENTS_CEILING: i64 = 4096;
+
+/// Lance writes one fragment per write batch, so a table fed by small appends
+/// accumulates small files the same way an Iceberg table does.
+pub struct CompactHandler {
+    namespace_url: String,
+    fallback: dataset::FallbackOptions,
+    metrics: Option<crate::metrics::LanceMetrics>,
+}
+
+impl CompactHandler {
+    pub fn new(namespace_url: String) -> Self {
+        Self {
+            namespace_url,
+            fallback: dataset::FallbackOptions::new(),
+            metrics: None,
+        }
+    }
+
+    pub fn with_metrics(mut self, metrics: Option<crate::metrics::LanceMetrics>) -> Self {
+        self.metrics = metrics;
+        self
+    }
+
+    /// Storage options to use where the namespace vends none.
+    pub fn with_fallback(mut self, fallback: dataset::FallbackOptions) -> Self {
+        self.fallback = fallback;
+        self
+    }
+
+    fn client(&self) -> NamespaceClient {
+        NamespaceClient::new(self.namespace_url.clone())
+    }
+}
+
+#[async_trait]
+impl JobHandler for CompactHandler {
+    fn capability(&self) -> JobTypeCapability {
+        JobTypeCapability {
+            job_type: JOB_TYPE.to_string(),
+            can_detect: true,
+            can_execute: true,
+            max_detection_concurrency: 1,
+            max_execution_concurrency: 1,
+            display_name: "Lance Compaction".to_string(),
+            description: "Merge small Lance fragments into fewer, larger ones".to_string(),
+            weight: 20,
+        }
+    }
+
+    fn descriptor(&self) -> JobTypeDescriptor {
+        let mut defaults: HashMap<String, ConfigValue> = HashMap::new();
+        defaults.insert(
+            "target_rows_per_fragment".to_string(),
+            int_value(DEFAULT_TARGET_ROWS),
+        );
+        defaults.insert(
+            "min_fragments".to_string(),
+            int_value(DEFAULT_MIN_FRAGMENTS),
+        );
+
+        JobTypeDescriptor {
+            job_type: JOB_TYPE.to_string(),
+            display_name: "Lance Compaction".to_string(),
+            description: "Compact fragments of Lance tables".to_string(),
+            icon: "fas fa-compress".to_string(),
+            descriptor_version: 1,
+            worker_config_form: Some(form(
+                "lance-compact-worker",
+                "Compaction",
+                vec![
+                    number_field(
+                        "target_rows_per_fragment",
+                        "Target rows per fragment",
+                        "Rows to aim for when rewriting fragments",
+                        1024,
+                        16_777_216,
+                    ),
+                    number_field(
+                        "min_fragments",
+                        "Minimum fragments",
+                        "Leave a table alone until it has at least this many fragments",
+                        2,
+                        4096,
+                    ),
+                ],
+                defaults.clone(),
+            )),
+            worker_default_values: defaults,
+            ..Default::default()
+        }
+    }
+
+    /// Propose a job for every table with more fragments than the operator is
+    /// willing to leave alone. Opening a dataset reads its manifest, not its
+    /// data, so this stays cheap across a catalog.
+    async fn detect(
+        &self,
+        request: &RunDetectionRequest,
+        sender: &dyn DetectionSender,
+    ) -> Result<()> {
+        let min_fragments = clamp(
+            int_or(
+                &request.worker_config_values,
+                "min_fragments",
+                DEFAULT_MIN_FRAGMENTS,
+            ),
+            MIN_FRAGMENTS_FLOOR,
+            MIN_FRAGMENTS_CEILING,
+        ) as usize;
+        let client = self.client();
+        let tables = client.list_all_tables().await?;
+
+        let mut proposals = Vec::new();
+        let mut observations = Vec::new();
+        for encoded in &tables {
+            let id = parse_id(encoded);
+            let table = match dataset::open(&client, &id, &self.fallback).await {
+                Ok(table) => table,
+                Err(err) => {
+                    // A table that cannot be opened is the next run's problem,
+                    // not a reason to abandon the whole sweep.
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "open");
+                    }
+                    warn!("skipping {encoded}: {err:#}");
+                    continue;
+                }
+            };
+            // One unreadable table must not end the sweep: the tables already
+            // read would lose their proposals, and admin would get no
+            // completion for this request at all.
+            if let Some(counters) = &self.metrics {
+                counters.worker.object_seen(JOB_TYPE);
+            }
+            let stats = match table.stats().await {
+                Ok(stats) => stats,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "stats");
+                    }
+                    warn!("skipping {encoded}: reading its stats failed: {err:#}");
+                    continue;
+                }
+            };
+            // Logged because "detection proposed nothing" is otherwise
+            // indistinguishable from a table the worker could not read.
+            tracing::info!(
+                "compaction detection: {encoded} has {} fragments, threshold {min_fragments}",
+                stats.fragments
+            );
+
+            let mut attributes: HashMap<String, ConfigValue> = HashMap::new();
+            attributes.insert("fragments".to_string(), int_value(stats.fragments as i64));
+            attributes.insert("version".to_string(), int_value(stats.version as i64));
+            attributes.insert(
+                "versions".to_string(),
+                int_value(stats.total_versions as i64),
+            );
+            attributes.insert("rows".to_string(), int_value(stats.rows as i64));
+            if let Some(schema) = stats.schema.clone() {
+                attributes.insert("schema".to_string(), string_value(schema));
+            }
+            observations.push(observation(&id, FORMAT, attributes));
+
+            if stats.fragments < min_fragments {
+                continue;
+            }
+            let mut parameters: HashMap<String, ConfigValue> = HashMap::new();
+            parameters.insert("table_id".to_string(), string_list(&id));
+            proposals.push(JobProposal {
+                proposal_id: format!("{JOB_TYPE}:{encoded}"),
+                dedupe_key: format!("{JOB_TYPE}:{encoded}"),
+                job_type: JOB_TYPE.to_string(),
+                summary: format!("Compact {encoded} ({} fragments)", stats.fragments),
+                detail: format!(
+                    "{} fragments at version {}, above the {min_fragments} the policy allows",
+                    stats.fragments, stats.version
+                ),
+                parameters,
+                ..Default::default()
+            });
+        }
+
+        if !observations.is_empty() {
+            sender.send_observations(WorkerObservations {
+                job_type: JOB_TYPE.to_string(),
+                observations,
+            })?;
+        }
+
+        let total = proposals.len() as i32;
+        sender.send_proposals(DetectionProposals {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            proposals,
+            has_more: false,
+        })?;
+        sender.send_complete(DetectionComplete {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            error_message: String::new(),
+            total_proposals: total,
+        })?;
+        Ok(())
+    }
+
+    async fn execute(
+        &self,
+        request: &ExecuteJobRequest,
+        sender: &dyn ExecutionSender,
+    ) -> Result<()> {
+        let job = request
+            .job
+            .as_ref()
+            .ok_or_else(|| anyhow!("execute request carried no job"))?;
+        let id = table_id(&job.parameters)
+            .ok_or_else(|| anyhow!("job {} carried no table_id", job.job_id))?;
+        let target_rows = clamp(
+            int_or(
+                &request.worker_config_values,
+                "target_rows_per_fragment",
+                DEFAULT_TARGET_ROWS,
+            ),
+            TARGET_ROWS_FLOOR,
+            TARGET_ROWS_CEILING,
+        ) as usize;
+
+        let client = self.client();
+        // Re-resolve rather than trusting the location detection saw: the table
+        // may have been repointed, and the vended credentials have expired.
+        let mut table = dataset::open(&client, &id, &self.fallback).await?;
+        let before = table.stats().await?;
+
+        sender.send_progress(JobProgressUpdate {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            progress_percent: 10.0,
+            stage: format!("compacting {} fragments", before.fragments),
+            ..Default::default()
+        })?;
+
+        let options = CompactionOptions {
+            target_rows_per_fragment: target_rows,
+            ..Default::default()
+        };
+        let metrics = compact_files(&mut table.dataset, options, None)
+            .await
+            .with_context(|| format!("compact {}", table.location))?;
+
+        if let Some(counters) = &self.metrics {
+            counters
+                .fragments_removed
+                .inc_by(metrics.fragments_removed as u64);
+        }
+
+        let after = table.stats().await?;
+        let mut output: HashMap<String, ConfigValue> = HashMap::new();
+        output.insert(
+            "fragments_removed".to_string(),
+            int_value(metrics.fragments_removed as i64),
+        );
+        output.insert(
+            "fragments_added".to_string(),
+            int_value(metrics.fragments_added as i64),
+        );
+        output.insert(
+            "files_removed".to_string(),
+            int_value(metrics.files_removed as i64),
+        );
+
+        sender.send_completed(JobCompleted {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            result: Some(JobResult {
+                output_values: output,
+                summary: format!("{} fragments became {}", before.fragments, after.fragments),
+                ..Default::default()
+            }),
+            ..Default::default()
+        })?;
+        Ok(())
+    }
+}

seaweed-worker/crates/lance/src/jobs/indices.rs

@@ -0,0 +1,275 @@
+use std::collections::HashMap;
+
+use anyhow::{Context, Result, anyhow};
+use async_trait::async_trait;
+use lance::index::DatasetIndexExt;
+use lance_index::optimize::OptimizeOptions;
+use seaweed_worker_core::config_form::{form, int_or, int_value, number_field};
+use seaweed_worker_core::pb::{
+    ConfigValue, DetectionComplete, DetectionProposals, ExecuteJobRequest, JobCompleted,
+    JobProgressUpdate, JobProposal, JobResult, JobTypeCapability, JobTypeDescriptor,
+    RunDetectionRequest,
+};
+use seaweed_worker_core::{DetectionSender, ExecutionSender, JobHandler};
+use tracing::warn;
+
+use crate::catalog::{NamespaceClient, parse_id};
+use crate::dataset::{self, OpenTable};
+use crate::jobs::{clamp, string_list, table_id};
+
+pub const JOB_TYPE: &str = "lance_optimize_indices";
+
+const DEFAULT_MAX_UNINDEXED_ROWS: i64 = 100_000;
+
+// Zero is a real setting - reindex as soon as any row is uncovered - so the
+// floor is what stays meaningful rather than what the form offers. The ceiling
+// is the form's, and the point of both is that this is cast to u64: a negative
+// would arrive as an enormous budget and mean "never reindex".
+const MAX_UNINDEXED_FLOOR: i64 = 0;
+const MAX_UNINDEXED_CEILING: i64 = 100_000_000;
+
+/// Rows written after an index was built are not covered by it, so a vector
+/// search quietly misses them. This is the job with no Iceberg equivalent, and
+/// the reason a neglected Lance table is a correctness problem rather than a
+/// slow one.
+pub struct OptimizeIndicesHandler {
+    namespace_url: String,
+    fallback: dataset::FallbackOptions,
+    metrics: Option<crate::metrics::LanceMetrics>,
+}
+
+impl OptimizeIndicesHandler {
+    pub fn new(namespace_url: String) -> Self {
+        Self {
+            namespace_url,
+            fallback: dataset::FallbackOptions::new(),
+            metrics: None,
+        }
+    }
+
+    pub fn with_metrics(mut self, metrics: Option<crate::metrics::LanceMetrics>) -> Self {
+        self.metrics = metrics;
+        self
+    }
+
+    pub fn with_fallback(mut self, fallback: dataset::FallbackOptions) -> Self {
+        self.fallback = fallback;
+        self
+    }
+
+    fn client(&self) -> NamespaceClient {
+        NamespaceClient::new(self.namespace_url.clone())
+    }
+}
+
+/// Rows no index covers, summed across a table's indices. A table with no
+/// indices at all has nothing to optimize, which is different from a table whose
+/// indices have fallen behind.
+async fn unindexed_rows(table: &OpenTable) -> Result<Option<u64>> {
+    let indices = table.dataset.load_indices().await?;
+    if indices.is_empty() {
+        return Ok(None);
+    }
+
+    let mut worst = 0u64;
+    let mut names: Vec<String> = indices.iter().map(|index| index.name.clone()).collect();
+    names.sort();
+    names.dedup();
+    for name in names {
+        let raw = table.dataset.index_statistics(&name).await?;
+        let stats: serde_json::Value = serde_json::from_str(&raw)
+            .with_context(|| format!("parse index statistics for {name}"))?;
+        let unindexed = stats
+            .get("num_unindexed_rows")
+            .and_then(|value| value.as_u64())
+            .unwrap_or(0);
+        worst = worst.max(unindexed);
+    }
+    Ok(Some(worst))
+}
+
+#[async_trait]
+impl JobHandler for OptimizeIndicesHandler {
+    fn capability(&self) -> JobTypeCapability {
+        JobTypeCapability {
+            job_type: JOB_TYPE.to_string(),
+            can_detect: true,
+            can_execute: true,
+            max_detection_concurrency: 1,
+            max_execution_concurrency: 1,
+            display_name: "Lance Index Optimization".to_string(),
+            description: "Extend indices to cover rows written since they were built".to_string(),
+            weight: 30,
+        }
+    }
+
+    fn descriptor(&self) -> JobTypeDescriptor {
+        let mut defaults: HashMap<String, ConfigValue> = HashMap::new();
+        defaults.insert(
+            "max_unindexed_rows".to_string(),
+            int_value(DEFAULT_MAX_UNINDEXED_ROWS),
+        );
+
+        JobTypeDescriptor {
+            job_type: JOB_TYPE.to_string(),
+            display_name: "Lance Index Optimization".to_string(),
+            description: "Keep vector and scalar indices covering the whole table".to_string(),
+            icon: "fas fa-magnifying-glass-chart".to_string(),
+            descriptor_version: 1,
+            worker_config_form: Some(form(
+                "lance-indices-worker",
+                "Index optimization",
+                vec![number_field(
+                    "max_unindexed_rows",
+                    "Unindexed row budget",
+                    "Reindex once a table has more rows than this outside its indices",
+                    1_000,
+                    100_000_000,
+                )],
+                defaults.clone(),
+            )),
+            worker_default_values: defaults,
+            ..Default::default()
+        }
+    }
+
+    async fn detect(
+        &self,
+        request: &RunDetectionRequest,
+        sender: &dyn DetectionSender,
+    ) -> Result<()> {
+        let budget = clamp(
+            int_or(
+                &request.worker_config_values,
+                "max_unindexed_rows",
+                DEFAULT_MAX_UNINDEXED_ROWS,
+            ),
+            MAX_UNINDEXED_FLOOR,
+            MAX_UNINDEXED_CEILING,
+        ) as u64;
+        let client = self.client();
+        let tables = client.list_all_tables().await?;
+
+        let mut proposals = Vec::new();
+        for encoded in &tables {
+            let id = parse_id(encoded);
+            let table = match dataset::open(&client, &id, &self.fallback).await {
+                Ok(table) => table,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "open");
+                    }
+                    warn!("skipping {encoded}: {err:#}");
+                    continue;
+                }
+            };
+            if let Some(counters) = &self.metrics {
+                counters.worker.object_seen(JOB_TYPE);
+            }
+            let unindexed = match unindexed_rows(&table).await {
+                Ok(Some(unindexed)) => unindexed,
+                Ok(None) => continue,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "index_stats");
+                    }
+                    warn!("skipping {encoded}: reading its index stats failed: {err:#}");
+                    continue;
+                }
+            };
+            if unindexed <= budget {
+                continue;
+            }
+            let mut parameters: HashMap<String, ConfigValue> = HashMap::new();
+            parameters.insert("table_id".to_string(), string_list(&id));
+            proposals.push(JobProposal {
+                proposal_id: format!("{JOB_TYPE}:{encoded}"),
+                dedupe_key: format!("{JOB_TYPE}:{encoded}"),
+                job_type: JOB_TYPE.to_string(),
+                summary: format!("Reindex {encoded} ({unindexed} rows uncovered)"),
+                detail: format!(
+                    "{unindexed} rows sit outside the indices, above the {budget} the policy allows; \
+                     a search of this table misses them"
+                ),
+                parameters,
+                ..Default::default()
+            });
+        }
+
+        let total = proposals.len() as i32;
+        sender.send_proposals(DetectionProposals {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            proposals,
+            has_more: false,
+        })?;
+        sender.send_complete(DetectionComplete {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            error_message: String::new(),
+            total_proposals: total,
+        })?;
+        Ok(())
+    }
+
+    async fn execute(
+        &self,
+        request: &ExecuteJobRequest,
+        sender: &dyn ExecutionSender,
+    ) -> Result<()> {
+        let job = request
+            .job
+            .as_ref()
+            .ok_or_else(|| anyhow!("execute request carried no job"))?;
+        let id = table_id(&job.parameters)
+            .ok_or_else(|| anyhow!("job {} carried no table_id", job.job_id))?;
+
+        let client = self.client();
+        let mut table = dataset::open(&client, &id, &self.fallback).await?;
+        let before = unindexed_rows(&table).await?.unwrap_or(0);
+
+        sender.send_progress(JobProgressUpdate {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            progress_percent: 10.0,
+            stage: format!("indexing {before} uncovered rows"),
+            ..Default::default()
+        })?;
+
+        // Merging every delta keeps read latency from drifting as the table is
+        // reindexed again and again; leaving them unmerged is how an index ends
+        // up fast to write and slow to search.
+        table
+            .dataset
+            .optimize_indices(&OptimizeOptions::default())
+            .await
+            .with_context(|| format!("optimize indices of {}", table.location))?;
+
+        let after = unindexed_rows(&table).await?.unwrap_or(0);
+        if let Some(counters) = &self.metrics {
+            counters.rows_indexed.inc_by(before.saturating_sub(after));
+        }
+        let mut output: HashMap<String, ConfigValue> = HashMap::new();
+        output.insert(
+            "unindexed_rows_before".to_string(),
+            int_value(before as i64),
+        );
+        output.insert("unindexed_rows_after".to_string(), int_value(after as i64));
+
+        sender.send_completed(JobCompleted {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            result: Some(JobResult {
+                output_values: output,
+                summary: format!("{before} uncovered rows became {after}"),
+                ..Default::default()
+            }),
+            ..Default::default()
+        })?;
+        Ok(())
+    }
+}

seaweed-worker/crates/lance/src/jobs/mod.rs

@@ -0,0 +1,118 @@
+//! One module per job type. Each declares its capability and the settings form
+//! admin renders for it, then does the work.
+//!
+//! Detection opens each table and reads its manifest rather than its data, so a
+//! sweep across a catalog stays cheap. Execution re-resolves the table instead
+//! of trusting what detection saw: it may have been repointed, and the vended
+//! credentials expire.
+
+pub mod cleanup;
+pub mod compact;
+pub mod indices;
+pub mod sort;
+
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use seaweed_worker_core::JobHandler;
+use seaweed_worker_core::pb::{ConfigValue, ObjectObservation, StringList, config_value::Kind};
+
+use crate::catalog::parse_id;
+
+/// A table identifier travels in a proposal's parameters and comes back on the
+/// job, so both sides agree on one encoding.
+pub(crate) fn string_list(parts: &[String]) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::StringList(StringList {
+            values: parts.to_vec(),
+        })),
+    }
+}
+
+pub(crate) fn table_id(parameters: &HashMap<String, ConfigValue>) -> Option<Vec<String>> {
+    match parameters.get("table_id")?.kind.as_ref()? {
+        Kind::StringList(list) => Some(list.values.clone()),
+        Kind::StringValue(encoded) => Some(parse_id(encoded)),
+        _ => None,
+    }
+}
+
+/// Every handler this worker serves. A worker process may serve several job
+/// types, which is why WorkerHello carries a list.
+pub fn handlers(
+    namespace_url: String,
+    fallback: crate::dataset::FallbackOptions,
+    metrics: Option<crate::metrics::LanceMetrics>,
+) -> Vec<Arc<dyn JobHandler>> {
+    vec![
+        Arc::new(
+            compact::CompactHandler::new(namespace_url.clone())
+                .with_fallback(fallback.clone())
+                .with_metrics(metrics.clone()),
+        ),
+        Arc::new(
+            indices::OptimizeIndicesHandler::new(namespace_url.clone())
+                .with_fallback(fallback.clone())
+                .with_metrics(metrics.clone()),
+        ),
+        Arc::new(
+            sort::SortHandler::new(namespace_url.clone())
+                .with_fallback(fallback.clone())
+                .with_metrics(metrics.clone()),
+        ),
+        Arc::new(
+            cleanup::CleanupVersionsHandler::new(namespace_url)
+                .with_fallback(fallback)
+                .with_metrics(metrics),
+        ),
+    ]
+}
+
+/// Holds a configured value to the range its form offers. A value from outside
+/// it is one the UI could not have produced, and every one of these is cast to
+/// an unsigned type: a negative arrives as an enormous number, which silently
+/// turns a threshold into "never" rather than failing loudly.
+pub(crate) fn clamp(value: i64, low: i64, high: i64) -> i64 {
+    value.max(low).min(high)
+}
+
+/// The format the catalog records for the tables this worker maintains.
+pub const FORMAT: &str = "LANCE";
+
+/// Builds the observation a detection sweep reports for one table. Detection
+/// has already opened the dataset to decide whether it needs work, so saying
+/// what it saw costs nothing, and for a format the cluster cannot read this is
+/// the only description of the table anything can produce.
+pub(crate) fn observation(
+    id: &[String],
+    format: &str,
+    attributes: HashMap<String, ConfigValue>,
+) -> ObjectObservation {
+    ObjectObservation {
+        object_id: id.to_vec(),
+        object_kind: "table".to_string(),
+        format: format.to_string(),
+        attributes,
+        observed_at: Some(std::time::SystemTime::now().into()),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // Every configured threshold is cast to an unsigned type before use. A
+    // negative one would arrive as an enormous number and quietly mean "never",
+    // which looks exactly like a worker with nothing to do.
+    #[test]
+    fn clamp_keeps_a_negative_from_wrapping() {
+        assert_eq!(clamp(-1, 2, 4096) as usize, 2);
+        assert_eq!(clamp(i64::MIN, 0, 100_000_000) as u64, 0);
+    }
+
+    #[test]
+    fn clamp_holds_the_ceiling_and_passes_the_middle() {
+        assert_eq!(clamp(i64::MAX, 0, 8760), 8760);
+        assert_eq!(clamp(168, 0, 8760), 168);
+    }
+}

seaweed-worker/crates/lance/src/jobs/sort.rs

@@ -0,0 +1,488 @@
+use std::collections::HashMap;
+use std::sync::Arc;
+
+use anyhow::{Context, Result, anyhow};
+use async_trait::async_trait;
+use lance::dataset::scanner::ColumnOrdering;
+use lance::dataset::transaction::Operation;
+use lance::dataset::write::{CommitBuilder, InsertBuilder};
+use lance::dataset::{WriteDestination, WriteMode, WriteParams};
+use lance::index::DatasetIndexExt;
+use lance_datafusion::exec::{LanceExecutionOptions, execute_plan};
+use seaweed_worker_core::config_form::{form, int_or, int_value, string_or, string_value};
+use seaweed_worker_core::pb::{
+    ConfigValue, DetectionComplete, DetectionProposals, ExecuteJobRequest, JobCompleted,
+    JobProgressUpdate, JobProposal, JobResult, JobTypeCapability, JobTypeDescriptor,
+    RunDetectionRequest, WorkerObservations,
+};
+use seaweed_worker_core::{DetectionSender, ExecutionSender, JobHandler};
+use seaweed_worker_sort::{
+    CONFIG_MAX_ROWS_PER_FILE, CONFIG_MEMORY_BUDGET_MB, CONFIG_MIN_UNSORTED_ROWS,
+    CONFIG_SORT_FIELDS, DECLARED_FIELDS_KEY, FragmentSummary, SortSpec, SortState, config_fields,
+    resolve, verdict,
+};
+use tracing::warn;
+
+use crate::catalog::{NamespaceClient, parse_id};
+use crate::dataset;
+use crate::jobs::{FORMAT, clamp, observation, string_list, table_id};
+
+pub const JOB_TYPE: &str = "lance_sort";
+
+const DEFAULT_MIN_UNSORTED_ROWS: i64 = 1_048_576;
+const DEFAULT_MEMORY_BUDGET_MB: i64 = 512;
+const DEFAULT_MAX_ROWS_PER_FILE: i64 = 1_048_576;
+
+// The ranges the shared form offers, repeated here because a value from outside
+// one is a value the UI could not have produced and every one of these is cast
+// to an unsigned type on the way in.
+const MIN_UNSORTED_ROWS_FLOOR: i64 = 1;
+const MIN_UNSORTED_ROWS_CEILING: i64 = 1_000_000_000;
+const MEMORY_BUDGET_FLOOR: i64 = 64;
+const MEMORY_BUDGET_CEILING: i64 = 1_048_576;
+const MAX_ROWS_PER_FILE_FLOOR: i64 = 1024;
+const MAX_ROWS_PER_FILE_CEILING: i64 = 16_777_216;
+
+const BYTES_PER_MB: u64 = 1024 * 1024;
+
+/// Rewrites a table so its rows are stored in the order its fields declare,
+/// which is what lets a range scan read a few files instead of all of them.
+///
+/// Lance appends fragments in write order and has no notion of a sorted table,
+/// so nothing but a rewrite establishes that order, and nothing but another
+/// rewrite restores it once rows have been appended.
+pub struct SortHandler {
+    namespace_url: String,
+    fallback: dataset::FallbackOptions,
+    metrics: Option<crate::metrics::LanceMetrics>,
+}
+
+impl SortHandler {
+    pub fn new(namespace_url: String) -> Self {
+        Self {
+            namespace_url,
+            fallback: dataset::FallbackOptions::new(),
+            metrics: None,
+        }
+    }
+
+    pub fn with_metrics(mut self, metrics: Option<crate::metrics::LanceMetrics>) -> Self {
+        self.metrics = metrics;
+        self
+    }
+
+    /// Storage options to use where the namespace vends none.
+    pub fn with_fallback(mut self, fallback: dataset::FallbackOptions) -> Self {
+        self.fallback = fallback;
+        self
+    }
+
+    fn client(&self) -> NamespaceClient {
+        NamespaceClient::new(self.namespace_url.clone())
+    }
+}
+
+/// The fragments a dataset holds, in fragment order, as the sort crate reads
+/// them: the data files each one names, and the live rows it holds.
+///
+/// Appends only add fragments after the existing ones and lance hands out
+/// fragment ids monotonically even across an overwrite, so ordering by id makes
+/// "the fragments the sort wrote are still the first ones" a prefix test.
+fn fragment_summaries(dataset: &lance::Dataset) -> Vec<FragmentSummary> {
+    let mut fragments = dataset.get_fragments();
+    fragments.sort_by_key(|fragment| fragment.id());
+    fragments
+        .iter()
+        .map(|fragment| summarize(fragment.metadata()))
+        .collect()
+}
+
+/// The manifest's own account of one fragment. `num_rows` is physical rows less
+/// deletions, and None where the manifest does not record them.
+fn summarize(fragment: &lance_table::format::Fragment) -> FragmentSummary {
+    FragmentSummary {
+        files: fragment
+            .files
+            .iter()
+            .map(|file| file.path.clone())
+            .collect(),
+        rows: fragment.num_rows().map(|rows| rows as u64),
+    }
+}
+
+/// The scan ordering for a spec. Lance validates the column names against the
+/// schema when the ordering is set, so a spec naming a column the table does
+/// not have fails before anything is written.
+fn orderings(spec: &SortSpec) -> Vec<ColumnOrdering> {
+    spec.fields
+        .iter()
+        .map(|field| ColumnOrdering {
+            ascending: !field.descending,
+            nulls_first: field.nulls_first,
+            column_name: field.path.clone(),
+        })
+        .collect()
+}
+
+#[async_trait]
+impl JobHandler for SortHandler {
+    fn capability(&self) -> JobTypeCapability {
+        JobTypeCapability {
+            job_type: JOB_TYPE.to_string(),
+            can_detect: true,
+            can_execute: true,
+            max_detection_concurrency: 1,
+            max_execution_concurrency: 1,
+            display_name: "Lance Sort".to_string(),
+            description: "Rewrite Lance tables in the order their fields declare. The rewrite replaces every fragment, so indices do not survive it and want rebuilding afterwards.".to_string(),
+            weight: 40,
+        }
+    }
+
+    fn descriptor(&self) -> JobTypeDescriptor {
+        let mut defaults: HashMap<String, ConfigValue> = HashMap::new();
+        defaults.insert(CONFIG_SORT_FIELDS.to_string(), string_value(""));
+        defaults.insert(
+            CONFIG_MIN_UNSORTED_ROWS.to_string(),
+            int_value(DEFAULT_MIN_UNSORTED_ROWS),
+        );
+        defaults.insert(
+            CONFIG_MEMORY_BUDGET_MB.to_string(),
+            int_value(DEFAULT_MEMORY_BUDGET_MB),
+        );
+        defaults.insert(
+            CONFIG_MAX_ROWS_PER_FILE.to_string(),
+            int_value(DEFAULT_MAX_ROWS_PER_FILE),
+        );
+
+        JobTypeDescriptor {
+            job_type: JOB_TYPE.to_string(),
+            display_name: "Lance Sort".to_string(),
+            description: "Sort Lance tables by their declared fields".to_string(),
+            icon: "fas fa-sort".to_string(),
+            descriptor_version: 1,
+            worker_config_form: Some(form(
+                "lance-sort-worker",
+                "Sort",
+                config_fields(),
+                defaults.clone(),
+            )),
+            worker_default_values: defaults,
+            ..Default::default()
+        }
+    }
+
+    /// Propose a job for every table whose declared order the data no longer
+    /// follows. Opening a dataset reads its manifest and its configuration, not
+    /// its data, so this stays cheap across a catalog.
+    async fn detect(
+        &self,
+        request: &RunDetectionRequest,
+        sender: &dyn DetectionSender,
+    ) -> Result<()> {
+        let configured = string_or(&request.worker_config_values, CONFIG_SORT_FIELDS, "");
+        let min_unsorted_rows = clamp(
+            int_or(
+                &request.worker_config_values,
+                CONFIG_MIN_UNSORTED_ROWS,
+                DEFAULT_MIN_UNSORTED_ROWS,
+            ),
+            MIN_UNSORTED_ROWS_FLOOR,
+            MIN_UNSORTED_ROWS_CEILING,
+        ) as u64;
+
+        let client = self.client();
+        let tables = client.list_all_tables().await?;
+
+        let mut proposals = Vec::new();
+        let mut observations = Vec::new();
+        for encoded in &tables {
+            let id = parse_id(encoded);
+            let table = match dataset::open(&client, &id, &self.fallback).await {
+                Ok(table) => table,
+                Err(err) => {
+                    // One unreadable table must not end the sweep: the tables
+                    // already read would lose their proposals.
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "open");
+                    }
+                    warn!("skipping {encoded}: {err:#}");
+                    continue;
+                }
+            };
+            if let Some(counters) = &self.metrics {
+                counters.worker.object_seen(JOB_TYPE);
+            }
+            let stats = match table.stats().await {
+                Ok(stats) => stats,
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "stats");
+                    }
+                    warn!("skipping {encoded}: reading its stats failed: {err:#}");
+                    continue;
+                }
+            };
+
+            let config = table.dataset.config().clone();
+            let spec = match resolve(
+                config.get(DECLARED_FIELDS_KEY).map(String::as_str),
+                &configured,
+            ) {
+                Ok(Some(spec)) => spec,
+                Ok(None) => {
+                    // Not every table is one an operator wants sorted, and a
+                    // table with no order named anywhere is not a failure.
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "no_sort_order");
+                    }
+                    continue;
+                }
+                Err(err) => {
+                    if let Some(counters) = &self.metrics {
+                        counters.worker.object_skipped(JOB_TYPE, "sort_order");
+                    }
+                    warn!("skipping {encoded}: {err:#}");
+                    continue;
+                }
+            };
+
+            let state = SortState::from_config(&config);
+            let decision = verdict(
+                &spec,
+                &state,
+                &fragment_summaries(&table.dataset),
+                min_unsorted_rows,
+            );
+
+            let mut attributes: HashMap<String, ConfigValue> = HashMap::new();
+            attributes.insert("rows".to_string(), int_value(stats.rows as i64));
+            attributes.insert("version".to_string(), int_value(stats.version as i64));
+            attributes.insert("fragments".to_string(), int_value(stats.fragments as i64));
+            attributes.insert("sort_fields".to_string(), string_value(spec.to_string()));
+            attributes.insert("sort_state".to_string(), string_value(decision.reason()));
+            if let Some(schema) = stats.schema.clone() {
+                attributes.insert("schema".to_string(), string_value(schema));
+            }
+            observations.push(observation(&id, FORMAT, attributes));
+
+            // Logged because "detection proposed nothing" is otherwise
+            // indistinguishable from a table the worker could not read.
+            tracing::info!(
+                "sort detection: {encoded} has {} rows in {spec} order: {}",
+                stats.rows,
+                decision.reason()
+            );
+
+            if !decision.needs_sort() {
+                continue;
+            }
+            let mut parameters: HashMap<String, ConfigValue> = HashMap::new();
+            parameters.insert("table_id".to_string(), string_list(&id));
+            proposals.push(JobProposal {
+                proposal_id: format!("{JOB_TYPE}:{encoded}"),
+                dedupe_key: format!("{JOB_TYPE}:{encoded}"),
+                job_type: JOB_TYPE.to_string(),
+                summary: format!("Sort {encoded} by {spec}"),
+                detail: format!("{} rows, {}", stats.rows, decision.reason()),
+                parameters,
+                ..Default::default()
+            });
+        }
+
+        if !observations.is_empty() {
+            sender.send_observations(WorkerObservations {
+                job_type: JOB_TYPE.to_string(),
+                observations,
+            })?;
+        }
+
+        let total = proposals.len() as i32;
+        sender.send_proposals(DetectionProposals {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            proposals,
+            has_more: false,
+        })?;
+        sender.send_complete(DetectionComplete {
+            request_id: request.request_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            error_message: String::new(),
+            total_proposals: total,
+        })?;
+        Ok(())
+    }
+
+    async fn execute(
+        &self,
+        request: &ExecuteJobRequest,
+        sender: &dyn ExecutionSender,
+    ) -> Result<()> {
+        let job = request
+            .job
+            .as_ref()
+            .ok_or_else(|| anyhow!("execute request carried no job"))?;
+        let id = table_id(&job.parameters)
+            .ok_or_else(|| anyhow!("job {} carried no table_id", job.job_id))?;
+
+        let configured = string_or(&request.worker_config_values, CONFIG_SORT_FIELDS, "");
+        let memory_budget_mb = clamp(
+            int_or(
+                &request.worker_config_values,
+                CONFIG_MEMORY_BUDGET_MB,
+                DEFAULT_MEMORY_BUDGET_MB,
+            ),
+            MEMORY_BUDGET_FLOOR,
+            MEMORY_BUDGET_CEILING,
+        ) as u64;
+        let max_rows_per_file = clamp(
+            int_or(
+                &request.worker_config_values,
+                CONFIG_MAX_ROWS_PER_FILE,
+                DEFAULT_MAX_ROWS_PER_FILE,
+            ),
+            MAX_ROWS_PER_FILE_FLOOR,
+            MAX_ROWS_PER_FILE_CEILING,
+        ) as usize;
+
+        let client = self.client();
+        // Re-resolve rather than trusting what detection saw: the table may have
+        // been repointed, its declared order changed, and the vended credentials
+        // have expired.
+        let table = dataset::open(&client, &id, &self.fallback).await?;
+        let before = table.stats().await?;
+        let config = table.dataset.config().clone();
+        let spec = resolve(
+            config.get(DECLARED_FIELDS_KEY).map(String::as_str),
+            &configured,
+        )?
+        .ok_or_else(|| anyhow!("neither the table nor this worker names an order to sort by"))?;
+
+        sender.send_progress(JobProgressUpdate {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            progress_percent: 10.0,
+            stage: format!("sorting {} rows by {spec}", before.rows),
+            ..Default::default()
+        })?;
+
+        let mut scanner = table.dataset.scan();
+        scanner
+            .order_by(Some(orderings(&spec)))
+            .with_context(|| format!("order the scan of {} by {spec}", table.location))?;
+        let plan = scanner
+            .create_plan()
+            .await
+            .context("plan the ordered scan")?;
+
+        // Spilling is off in LanceExecutionOptions::default(), and the scanner's
+        // own stream helper takes that default, so the options are built here
+        // instead: a sort that cannot spill is one bounded by memory, which is
+        // the whole thing this job exists to avoid.
+        let stream = execute_plan(
+            plan,
+            LanceExecutionOptions {
+                use_spilling: true,
+                mem_pool_size: Some(memory_budget_mb * BYTES_PER_MB),
+                ..Default::default()
+            },
+        )
+        .context("run the ordered scan")?;
+
+        let params = WriteParams {
+            mode: WriteMode::Overwrite,
+            max_rows_per_file,
+            ..Default::default()
+        };
+        let destination = Arc::new(table.dataset.clone());
+        let mut transaction = InsertBuilder::new(WriteDestination::Dataset(destination.clone()))
+            .with_params(&params)
+            .execute_uncommitted_stream(stream)
+            .await
+            .context("write the sorted fragments")?;
+
+        // The marker rides in the same commit as the data it describes:
+        // Operation::Overwrite is the one place lance takes configuration
+        // values alongside fragments, and a marker written as a second commit
+        // would be lost if the worker died in between, re-sorting the whole
+        // table on the next sweep.
+        match &mut transaction.operation {
+            Operation::Overwrite {
+                config_upsert_values,
+                fragments,
+                ..
+            } => {
+                // The files these fragments name were written a moment ago and
+                // keep their names whatever version this commit becomes, which
+                // is what lets the marker travel inside the commit it describes.
+                let written: Vec<FragmentSummary> = fragments.iter().map(summarize).collect();
+                *config_upsert_values =
+                    Some(SortState::record(&spec, &written, before.rows as u64));
+            }
+            other => {
+                return Err(anyhow!(
+                    "a sorted rewrite produced {other} instead of an overwrite"
+                ));
+            }
+        }
+
+        let sorted = CommitBuilder::new(WriteDestination::Dataset(destination))
+            .execute(transaction)
+            .await
+            .context("commit the sorted rewrite")?;
+
+        if let Some(counters) = &self.metrics {
+            counters.rows_sorted.inc_by(before.rows as u64);
+        }
+
+        // An overwrite carries fragments, schema and configuration — not index
+        // metadata — so a table that had indices no longer does. Saying so is
+        // the difference between a slow table and a mystery.
+        let dropped_indices = match table.dataset.load_indices().await {
+            Ok(indices) => indices.len(),
+            Err(err) => {
+                warn!("could not read the indices of {}: {err:#}", table.location);
+                0
+            }
+        };
+        if dropped_indices > 0 {
+            warn!(
+                "{} lost {dropped_indices} index/indices to the sorted rewrite; run the index job to rebuild them",
+                table.location
+            );
+        }
+
+        let fragments_after = sorted.get_fragments().len();
+        let mut output: HashMap<String, ConfigValue> = HashMap::new();
+        output.insert("rows_sorted".to_string(), int_value(before.rows as i64));
+        output.insert("sort_fields".to_string(), string_value(spec.to_string()));
+        output.insert(
+            "fragments_before".to_string(),
+            int_value(before.fragments as i64),
+        );
+        output.insert(
+            "fragments_after".to_string(),
+            int_value(fragments_after as i64),
+        );
+        output.insert(
+            "indices_dropped".to_string(),
+            int_value(dropped_indices as i64),
+        );
+
+        sender.send_completed(JobCompleted {
+            request_id: request.request_id.clone(),
+            job_id: job.job_id.clone(),
+            job_type: JOB_TYPE.to_string(),
+            success: true,
+            result: Some(JobResult {
+                output_values: output,
+                summary: format!("{} rows rewritten in {spec} order", before.rows),
+                ..Default::default()
+            }),
+            ..Default::default()
+        })?;
+        Ok(())
+    }
+}

seaweed-worker/crates/lance/src/lib.rs

@@ -0,0 +1,15 @@
+//! Maintenance jobs for Lance tables.
+//!
+//! A Lance dataset needs three things done to it over time: its fragments
+//! compacted, its indices extended to cover rows written after they were built,
+//! and its old versions removed. None can run in the Go worker, because all
+//! three read and rewrite Lance files. This crate is the worker that can.
+
+pub mod catalog;
+pub mod dataset;
+pub mod jobs;
+pub mod metrics;
+pub mod preview;
+pub mod tls;
+
+pub use jobs::handlers;

seaweed-worker/crates/lance/src/main.rs

@@ -0,0 +1,193 @@
+use std::time::Duration;
+
+use anyhow::{Context, Result};
+use clap::Parser;
+use std::net::{IpAddr, SocketAddr};
+use std::sync::Arc;
+
+use seaweed_worker_core::{Metrics, Registry, TlsOptions, WorkerOptions};
+use weed_lance_worker::handlers;
+use weed_lance_worker::metrics::LanceMetrics;
+
+/// Mirrors `weed worker`'s flags, because this is the same contract from another
+/// language and an operator should not have to learn a second set of names.
+#[derive(Parser, Debug)]
+#[command(name = "weed-worker", about = "SeaweedFS maintenance worker")]
+struct Args {
+    /// Admin server gRPC address.
+    #[arg(long, default_value = "localhost:23646", env = "WEED_ADMIN")]
+    admin: String,
+
+    /// Worker identity reported to admin.
+    #[arg(long, default_value = "lance-worker", env = "WEED_WORKER_ID")]
+    id: String,
+
+    /// Lance namespace the worker lists tables from.
+    #[arg(
+        long,
+        default_value = "http://localhost:9101",
+        env = "WEED_LANCE_NAMESPACE"
+    )]
+    namespace: String,
+
+    #[arg(long, default_value = "10", env = "WEED_HEARTBEAT_SECONDS")]
+    heartbeat_seconds: u64,
+
+    #[arg(long, default_value = "1")]
+    max_concurrency: i32,
+
+    /// Storage credentials to use where the namespace vends none. A gateway
+    /// with STS configured vends its own and these are ignored.
+    #[arg(long, env = "WEED_S3_ACCESS_KEY")]
+    access_key: Option<String>,
+
+    #[arg(long, env = "WEED_S3_SECRET_KEY")]
+    secret_key: Option<String>,
+
+    /// mTLS for the admin stream, the same certificates the Go worker reads
+    /// from the [grpc.worker] section of security.toml. All three together, or
+    /// none, in which case the stream is plaintext.
+    #[arg(long, env = "WEED_GRPC_CA")]
+    tls_ca: Option<String>,
+
+    #[arg(long, env = "WEED_GRPC_CLIENT_CERT")]
+    tls_cert: Option<String>,
+
+    #[arg(long, env = "WEED_GRPC_CLIENT_KEY")]
+    tls_key: Option<String>,
+
+    /// Name to verify admin's certificate against, when it is not the address
+    /// this worker dials.
+    #[arg(long)]
+    tls_server_name: Option<String>,
+
+    /// Serve /health, /ready and /metrics on this port, the way
+    /// `weed worker -metricsPort` does. Zero, the default, serves nothing.
+    #[arg(long, default_value = "0", env = "WEED_METRICS_PORT")]
+    metrics_port: u16,
+
+    /// Address the metrics server binds. Loopback by default, since the
+    /// endpoint is unauthenticated.
+    #[arg(long, default_value = "127.0.0.1", env = "WEED_METRICS_IP")]
+    metrics_ip: String,
+}
+
+impl Args {
+    /// The TLS configuration, or an error when the three certificate paths do
+    /// not arrive together: a CA on its own would silently give one-way TLS,
+    /// which the cluster's mutual setup refuses anyway.
+    fn tls(&self) -> Result<Option<TlsOptions>> {
+        match (
+            self.tls_ca.clone(),
+            self.tls_cert.clone(),
+            self.tls_key.clone(),
+        ) {
+            (None, None, None) => Ok(None),
+            (Some(ca_path), Some(client_cert_path), Some(client_key_path)) => {
+                Ok(Some(TlsOptions {
+                    ca_path,
+                    client_cert_path,
+                    client_key_path,
+                    server_name: self.tls_server_name.clone(),
+                }))
+            }
+            _ => Err(anyhow::anyhow!(
+                "--tls-ca, --tls-cert and --tls-key must be given together"
+            )),
+        }
+    }
+}
+
+/// Builds the metrics bind address. Parsing the host separately is what makes an
+/// IPv6 literal work: "::1" and 9327 joined with a colon is not an address, and
+/// formatting them that way turns `--metrics-ip ::1` into a startup failure.
+/// Brackets are accepted too, since that is how the same address is written in a
+/// URL and an operator will reasonably try it.
+fn metrics_address(ip: &str, port: u16) -> Result<SocketAddr> {
+    let host = ip.trim().trim_start_matches('[').trim_end_matches(']');
+    let parsed: IpAddr = host
+        .parse()
+        .with_context(|| format!("parse the metrics address {ip}"))?;
+    Ok(SocketAddr::new(parsed, port))
+}
+
+#[tokio::main]
+async fn main() -> Result<()> {
+    weed_lance_worker::tls::install_default_crypto_provider();
+
+    tracing_subscriber::fmt()
+        .with_env_filter(
+            tracing_subscriber::EnvFilter::try_from_default_env()
+                .unwrap_or_else(|_| tracing_subscriber::EnvFilter::new("info")),
+        )
+        .init();
+
+    let args = Args::parse();
+    let tls = args.tls()?;
+    let options = WorkerOptions {
+        admin_address: args.admin,
+        worker_id: args.id,
+        heartbeat_interval: Duration::from_secs(args.heartbeat_seconds),
+        max_detection_concurrency: args.max_concurrency,
+        max_execution_concurrency: args.max_concurrency,
+        tls,
+        ..Default::default()
+    };
+
+    let mut fallback = weed_lance_worker::dataset::FallbackOptions::new();
+    if let (Some(access), Some(secret)) = (args.access_key, args.secret_key) {
+        fallback.insert("aws_access_key_id".to_string(), access);
+        fallback.insert("aws_secret_access_key".to_string(), secret);
+    }
+
+    let metrics = Metrics::new(&options.worker_id, &options.worker_version)?;
+    let lance_metrics = LanceMetrics::new(&metrics)?;
+    if args.metrics_port > 0 {
+        seaweed_worker_core::metrics::spawn(
+            metrics.clone(),
+            metrics_address(&args.metrics_ip, args.metrics_port)?,
+        );
+    }
+
+    let mut registry = Registry::new().with_preview(Arc::new(
+        weed_lance_worker::preview::LancePreview::new(args.namespace.clone(), fallback.clone()),
+    ));
+    for handler in handlers(args.namespace, fallback, Some(lance_metrics)) {
+        registry = registry.register(handler);
+    }
+
+    seaweed_worker_core::stream::run_with_metrics(options, registry, metrics).await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::metrics_address;
+
+    #[test]
+    fn metrics_address_takes_ipv4_ipv6_and_brackets() {
+        assert_eq!(
+            metrics_address("127.0.0.1", 9327).unwrap().to_string(),
+            "127.0.0.1:9327"
+        );
+        // Joining these with a colon gives "::1:9327", which is not an address.
+        assert_eq!(
+            metrics_address("::1", 9327).unwrap().to_string(),
+            "[::1]:9327"
+        );
+        assert_eq!(
+            metrics_address("[::1]", 9327).unwrap().to_string(),
+            "[::1]:9327"
+        );
+        assert_eq!(
+            metrics_address("0.0.0.0", 9327).unwrap().to_string(),
+            "0.0.0.0:9327"
+        );
+    }
+
+    #[test]
+    fn metrics_address_rejects_a_hostname() {
+        // Binding takes an address, not a name; saying so beats a confusing
+        // failure inside the server.
+        assert!(metrics_address("localhost", 9327).is_err());
+    }
+}

seaweed-worker/crates/lance/src/metrics.rs

@@ -0,0 +1,50 @@
+//! What the Lance jobs reclaimed, in numbers a scrape can add up.
+//!
+//! The generic worker metrics say a job ran and how long it took. These say what
+//! it did: a compaction sweep that runs every minute and removes nothing is a
+//! different thing from one that never runs, and only these tell them apart.
+
+use anyhow::Result;
+use prometheus::IntCounter;
+use seaweed_worker_core::Metrics;
+
+/// Counters the Lance jobs raise, registered on the worker's own registry so
+/// they are served from the same endpoint.
+#[derive(Clone)]
+pub struct LanceMetrics {
+    /// The worker's own metrics, so a job can also record what it read and what
+    /// it had to skip - the pair that tells "nothing needed doing" apart from
+    /// "nothing could be read".
+    pub worker: Metrics,
+    pub fragments_removed: IntCounter,
+    pub rows_indexed: IntCounter,
+    pub rows_sorted: IntCounter,
+    pub versions_removed: IntCounter,
+    pub bytes_reclaimed: IntCounter,
+}
+
+impl LanceMetrics {
+    pub fn new(metrics: &Metrics) -> Result<Self> {
+        Ok(Self {
+            worker: metrics.clone(),
+            fragments_removed: metrics.counter(
+                "lance_fragments_removed_total",
+                "Fragments merged away by compaction.",
+            )?,
+            rows_indexed: metrics.counter(
+                "lance_rows_indexed_total",
+                "Rows brought under an index that did not cover them.",
+            )?,
+            rows_sorted: metrics.counter(
+                "lance_rows_sorted_total",
+                "Rows rewritten into their table's declared order.",
+            )?,
+            versions_removed: metrics.counter(
+                "lance_versions_removed_total",
+                "Dataset versions removed by cleanup.",
+            )?,
+            bytes_reclaimed: metrics
+                .counter("lance_bytes_reclaimed_total", "Bytes freed by cleanup.")?,
+        })
+    }
+}

seaweed-worker/crates/lance/src/preview.rs

@@ -0,0 +1,86 @@
+//! Sample rows of a Lance table, for a page that cannot read the format.
+//!
+//! Admin renders an Iceberg table by reading its Parquet files directly. There
+//! is no Go Lance reader, so for Lance the worker does the reading and hands
+//! back text.
+
+use anyhow::{Context, Result};
+use arrow_cast::display::{ArrayFormatter, FormatOptions};
+use async_trait::async_trait;
+use futures::TryStreamExt;
+use seaweed_worker_core::{Preview, PreviewProvider};
+
+use crate::catalog::NamespaceClient;
+use crate::dataset::{self, FallbackOptions};
+use crate::jobs::FORMAT;
+
+pub struct LancePreview {
+    namespace_url: String,
+    fallback: FallbackOptions,
+}
+
+impl LancePreview {
+    pub fn new(namespace_url: String, fallback: FallbackOptions) -> Self {
+        Self {
+            namespace_url,
+            fallback,
+        }
+    }
+}
+
+#[async_trait]
+impl PreviewProvider for LancePreview {
+    fn format(&self) -> &str {
+        FORMAT
+    }
+
+    async fn preview(&self, object_id: &[String], row_limit: usize) -> Result<Preview> {
+        let client = NamespaceClient::new(self.namespace_url.clone());
+        let table = dataset::open(&client, object_id, &self.fallback).await?;
+
+        let total_rows = table.dataset.count_rows(None).await.unwrap_or(0) as i64;
+        let mut scanner = table.dataset.scan();
+        scanner.limit(Some(row_limit as i64), None)?;
+
+        let batches: Vec<_> = scanner
+            .try_into_stream()
+            .await?
+            .try_collect()
+            .await
+            .with_context(|| format!("read rows from {}", table.location))?;
+
+        let columns = table
+            .dataset
+            .schema()
+            .fields
+            .iter()
+            .map(|field| field.name.clone())
+            .collect();
+
+        let mut rows = Vec::new();
+        for batch in &batches {
+            // ArrayFormatter renders each Arrow type the way its own tooling
+            // does, so a vector column reads as a vector rather than as bytes.
+            let formatters = batch
+                .columns()
+                .iter()
+                .map(|array| ArrayFormatter::try_new(array.as_ref(), &FormatOptions::default()))
+                .collect::<Result<Vec<_>, _>>()?;
+            for index in 0..batch.num_rows() {
+                rows.push(
+                    formatters
+                        .iter()
+                        .map(|formatter| formatter.value(index).to_string())
+                        .collect(),
+                );
+            }
+        }
+        rows.truncate(row_limit);
+
+        Ok(Preview {
+            columns,
+            rows,
+            total_rows,
+        })
+    }
+}

seaweed-worker/crates/lance/src/tls.rs

@@ -0,0 +1,35 @@
+use rustls::crypto::aws_lc_rs;
+
+// aws-lc-rs and ring both get linked transitively (lance's aws backend pulls
+// aws-lc-rs, reqwest's rustls-tls pulls ring), so rustls can't auto-select a
+// provider and tonic's client TLS panics on first use. Pin the default to
+// aws-lc-rs, matching the Rust volume server. Idempotent.
+pub fn install_default_crypto_provider() {
+    let _ = aws_lc_rs::default_provider().install_default();
+}
+
+#[cfg(test)]
+mod tests {
+    use tonic::transport::{Certificate, Channel, ClientTlsConfig, Identity};
+
+    use super::install_default_crypto_provider;
+
+    const TEST_CERT_PEM: &str = "-----BEGIN CERTIFICATE-----\nMIIBPDCB76ADAgECAhRuRPQgeAu43BT/M7EfAWSdapVdYDAFBgMrZXAwFDESMBAG\nA1UEAwwJbG9jYWxob3N0MB4XDTI2MDcwNTE2MTUyOVoXDTM2MDcwMjE2MTUyOVow\nFDESMBAGA1UEAwwJbG9jYWxob3N0MCowBQYDK2VwAyEAr/3bNIFI+8V32oCiY6y+\nXRFmZpdNQ2g//VtRkT+nQg+jUzBRMB0GA1UdDgQWBBTsy9tLf1zPiXCQfgci6zNi\ndEzRSjAfBgNVHSMEGDAWgBTsy9tLf1zPiXCQfgci6zNidEzRSjAPBgNVHRMBAf8E\nBTADAQH/MAUGAytlcANBAIvsdw0IbvOBBkb9cd7BfMJfIP9pQQrAL03pCRWJFnFh\nSysaLVgFXI4T078IiaM874oO+iB+5vNbWEpc7CkGow4=\n-----END CERTIFICATE-----\n";
+    const TEST_KEY_PEM: &str = "-----BEGIN PRIVATE KEY-----\nMC4CAQAwBQYDK2VwBCIEIHbyn71Kk+Y7KT3sBctit7uZpErpoH6qDbFj6P8qGaZH\n-----END PRIVATE KEY-----\n";
+
+    // Without install_default_crypto_provider this panics in the lance crate,
+    // where both aws-lc-rs and ring are linked. Mirrors the volume server's
+    // test_build_grpc_endpoint_with_tls_resolves_crypto_provider.
+    #[tokio::test]
+    async fn tls_channel_builds_after_crypto_provider_installed() {
+        install_default_crypto_provider();
+        let config = ClientTlsConfig::new()
+            .ca_certificate(Certificate::from_pem(TEST_CERT_PEM))
+            .identity(Identity::from_pem(TEST_CERT_PEM, TEST_KEY_PEM));
+        let endpoint = Channel::from_shared("https://127.0.0.1:9")
+            .unwrap()
+            .tls_config(config)
+            .unwrap();
+        assert_eq!(endpoint.uri().scheme_str(), Some("https"));
+    }
+}

seaweed-worker/crates/lance/tests/common/mod.rs

@@ -0,0 +1,97 @@
+//! What every integration test in this crate needs to drive a handler: a sender
+//! that keeps what the handler sent, and the two settings that make a handler
+//! usable against a test gateway.
+//!
+//! Each test binary compiles this module on its own and uses part of it.
+#![allow(dead_code)]
+
+use std::collections::HashMap;
+use std::sync::Mutex;
+
+use anyhow::Result;
+use seaweed_worker_core::pb::{
+    ActivityEvent, ConfigValue, DetectionComplete, DetectionProposals, JobCompleted,
+    JobProgressUpdate, JobProposal, ObjectObservation, WorkerObservations, config_value::Kind,
+};
+use seaweed_worker_core::{DetectionSender, ExecutionSender};
+
+/// Keeps everything a handler reports, so a test can assert on it afterwards.
+#[derive(Default)]
+pub struct Recorder {
+    pub proposals: Mutex<Vec<JobProposal>>,
+    pub observations: Mutex<Vec<ObjectObservation>>,
+    pub completed: Mutex<Vec<JobCompleted>>,
+}
+
+impl DetectionSender for Recorder {
+    fn send_proposals(&self, proposals: DetectionProposals) -> Result<()> {
+        self.proposals.lock().unwrap().extend(proposals.proposals);
+        Ok(())
+    }
+    fn send_complete(&self, _complete: DetectionComplete) -> Result<()> {
+        Ok(())
+    }
+    fn send_activity(&self, _activity: ActivityEvent) -> Result<()> {
+        Ok(())
+    }
+    fn send_observations(&self, observations: WorkerObservations) -> Result<()> {
+        self.observations
+            .lock()
+            .unwrap()
+            .extend(observations.observations);
+        Ok(())
+    }
+}
+
+impl ExecutionSender for Recorder {
+    fn send_progress(&self, _progress: JobProgressUpdate) -> Result<()> {
+        Ok(())
+    }
+    fn send_completed(&self, completed: JobCompleted) -> Result<()> {
+        self.completed.lock().unwrap().push(completed);
+        Ok(())
+    }
+}
+
+/// The gateway to run against, or nothing, in which case the caller skips:
+/// these handlers rewrite real files, and there is nothing to learn from them
+/// against a fake.
+pub fn namespace_url() -> Option<String> {
+    std::env::var("WEED_LANCE_NAMESPACE")
+        .ok()
+        .filter(|s| !s.is_empty())
+}
+
+pub fn int_config(name: &str, value: i64) -> HashMap<String, ConfigValue> {
+    let mut values = HashMap::new();
+    values.insert(
+        name.to_string(),
+        ConfigValue {
+            kind: Some(Kind::Int64Value(value)),
+        },
+    );
+    values
+}
+
+/// What to reach the data with where the namespace vends nothing of its own -
+/// a gateway without STS vends no credentials, and one bound to a wildcard
+/// address vends no endpoint either. Taken from the environment so a harness
+/// can point these tests at a gateway that does check what it is given.
+pub fn fallback() -> weed_lance_worker::dataset::FallbackOptions {
+    let mut options = weed_lance_worker::dataset::FallbackOptions::new();
+    let from_env =
+        |name: &str, default: &str| std::env::var(name).unwrap_or_else(|_| default.to_string());
+    options.insert(
+        "aws_access_key_id".to_string(),
+        from_env("AWS_ACCESS_KEY_ID", "any"),
+    );
+    options.insert(
+        "aws_secret_access_key".to_string(),
+        from_env("AWS_SECRET_ACCESS_KEY", "any"),
+    );
+    if let Ok(endpoint) = std::env::var("AWS_ENDPOINT_URL") {
+        options.insert("aws_endpoint".to_string(), endpoint);
+        options.insert("allow_http".to_string(), "true".to_string());
+    }
+    options
+}

seaweed-worker/crates/lance/tests/compaction.rs

@@ -0,0 +1,540 @@
+//! Drives the compaction handler against a live namespace.
+//!
+//! Skipped unless WEED_LANCE_NAMESPACE names one, the way the Go integration
+//! tests skip without Docker: compaction rewrites real files, and there is
+//! nothing to learn from it against a fake.
+
+use std::collections::HashMap;
+
+use anyhow::Result;
+use seaweed_worker_core::pb::{
+    ExecuteJobRequest, JobSpec, RunDetectionRequest, config_value::Kind,
+};
+use seaweed_worker_core::{JobHandler, PreviewProvider};
+use weed_lance_worker::catalog::NamespaceClient;
+use weed_lance_worker::jobs::cleanup::CleanupVersionsHandler;
+use weed_lance_worker::jobs::compact::{CompactHandler, JOB_TYPE};
+use weed_lance_worker::jobs::indices::OptimizeIndicesHandler;
+use weed_lance_worker::preview::LancePreview;
+
+mod common;
+use common::{Recorder, fallback, int_config, namespace_url};
+
+/// These tests drive one live gateway and one shared catalog: `list_all_tables`
+/// sweeps everything, so a table another test is writing shows up in this test's
+/// detection. Rust runs a binary's tests concurrently, so take a lock.
+static GATEWAY: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
+
+/// Declares a table through the namespace and writes `fragments` one-row
+/// appends into it, so a test brings its own state instead of depending on
+/// whatever a previous run left behind.
+async fn seed_fragmented_table(url: &str, name: &str, fragments: usize) -> Result<String> {
+    seed_table(url, name, fragments, 1, false).await
+}
+
+/// Writes `batches` appends of `rows_each` into a freshly declared table, and
+/// optionally builds a vector index after the first batch so the later ones are
+/// rows no index covers.
+async fn seed_table(
+    url: &str,
+    name: &str,
+    batches: usize,
+    rows_each: usize,
+    with_index: bool,
+) -> Result<String> {
+    use arrow_array::{
+        FixedSizeListArray, Float32Array, Int64Array, RecordBatch, RecordBatchIterator,
+    };
+    use arrow_schema::{DataType, Field, Schema};
+    use lance::dataset::{Dataset, WriteMode, WriteParams};
+    use lance::io::{ObjectStoreParams, StorageOptionsAccessor};
+    use std::sync::Arc;
+
+    // Declaring is the namespace's job, not the worker's, so the test asks for
+    // it directly rather than widening the client the worker uses. The bucket and
+    // namespace come first: a table cannot be declared under a parent that does
+    // not exist, and a test that assumes one is a test that only passes twice.
+    let http = reqwest::Client::new();
+    for parent in ["vec", "vec$ml"] {
+        http.post(format!("{url}/v1/namespace/{parent}/create"))
+            .json(&serde_json::json!({"mode": "EXIST_OK"}))
+            .send()
+            .await?
+            .error_for_status()?;
+    }
+    let encoded = format!("vec$ml${name}");
+    http.post(format!("{url}/v1/table/{encoded}/declare"))
+        .json(&serde_json::json!({}))
+        .send()
+        .await?
+        .error_for_status()?;
+
+    let client = NamespaceClient::new(url.to_string());
+    let id = vec!["vec".to_string(), "ml".to_string(), name.to_string()];
+    let description = client.describe_table(&id).await?;
+
+    let mut options = description.storage_options.clone();
+    options.extend(fallback());
+    const DIM: i32 = 16;
+    let schema = Arc::new(Schema::new(vec![
+        Field::new("id", DataType::Int64, false),
+        Field::new(
+            "vec",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), DIM),
+            false,
+        ),
+    ]));
+
+    for i in 0..batches {
+        let ids: Vec<i64> = (0..rows_each).map(|r| (i * rows_each + r) as i64).collect();
+        let values: Vec<f32> = ids
+            .iter()
+            .flat_map(|id| (0..DIM).map(move |d| (*id as f32) + d as f32))
+            .collect();
+        let vectors = FixedSizeListArray::new(
+            Arc::new(Field::new("item", DataType::Float32, true)),
+            DIM,
+            Arc::new(Float32Array::from(values)),
+            None,
+        );
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            vec![Arc::new(Int64Array::from(ids)), Arc::new(vectors)],
+        )?;
+        let reader = RecordBatchIterator::new(vec![Ok(batch)], schema.clone());
+        let params = WriteParams {
+            mode: if i == 0 {
+                WriteMode::Overwrite
+            } else {
+                WriteMode::Append
+            },
+            store_params: Some(ObjectStoreParams {
+                storage_options_accessor: Some(std::sync::Arc::new(
+                    StorageOptionsAccessor::with_static_options(options.clone()),
+                )),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let dataset = Dataset::write(reader, description.location.as_str(), Some(params)).await?;
+
+        // The index is built after the first batch, so everything appended
+        // afterwards is a row it does not cover.
+        if with_index && i == 0 {
+            use lance::index::DatasetIndexExt;
+            use lance::index::vector::VectorIndexParams;
+            use lance_index::IndexType;
+            use lance_index::vector::{ivf::IvfBuildParams, pq::PQBuildParams};
+
+            let mut dataset = dataset;
+            let params = VectorIndexParams::with_ivf_pq_params(
+                lance_linalg::distance::MetricType::L2,
+                IvfBuildParams::new(1),
+                PQBuildParams::new(4, 8),
+            );
+            dataset
+                .create_index(&["vec"], IndexType::Vector, None, &params, true)
+                .await?;
+        }
+    }
+    Ok(encoded)
+}
+
+/// A table with more fragments than the policy allows is proposed, and running
+/// the proposal leaves it with fewer than it started with.
+#[tokio::test]
+async fn compacts_a_fragmented_table() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    // Seeded here rather than by a script, so the test is repeatable: a previous
+    // run compacts the table it depended on.
+    let encoded = seed_fragmented_table(&url, "compactme", 12)
+        .await
+        .expect("seed a fragmented table");
+    let handler = CompactHandler::new(url).with_fallback(fallback());
+    let recorder = Recorder::default();
+
+    let request = RunDetectionRequest {
+        request_id: "detect-1".to_string(),
+        job_type: JOB_TYPE.to_string(),
+        worker_config_values: int_config("min_fragments", 4),
+        ..Default::default()
+    };
+    handler
+        .detect(&request, &recorder)
+        .await
+        .expect("detection failed");
+
+    let proposals = recorder.proposals.lock().unwrap().clone();
+    assert!(
+        !proposals.is_empty(),
+        "expected a proposal for the fragmented table"
+    );
+    let proposal = proposals
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .expect("no proposal for the seeded table");
+
+    // Detection opened the dataset to decide, so it reports what it saw. This is
+    // the only description of a Lance table anything outside the format can give.
+    let observations = recorder.observations.lock().unwrap().clone();
+    let observed = observations
+        .iter()
+        .find(|o| o.object_id.last().map(String::as_str) == Some("compactme"))
+        .expect("detection reported no observation for the seeded table");
+    assert_eq!(observed.format, "LANCE");
+    for attribute in ["fragments", "rows", "versions", "schema"] {
+        assert!(
+            observed.attributes.contains_key(attribute),
+            "observation is missing {attribute}: {:?}",
+            observed.attributes.keys().collect::<Vec<_>>()
+        );
+    }
+
+    let execute = ExecuteJobRequest {
+        request_id: "execute-1".to_string(),
+        job: Some(JobSpec {
+            job_id: "job-1".to_string(),
+            job_type: JOB_TYPE.to_string(),
+            parameters: proposal.parameters.clone(),
+            ..Default::default()
+        }),
+        worker_config_values: int_config("target_rows_per_fragment", 1_048_576),
+        ..Default::default()
+    };
+    handler
+        .execute(&execute, &recorder)
+        .await
+        .expect("execution failed");
+
+    let completed = recorder.completed.lock().unwrap().clone();
+    let result = completed.first().expect("no completion reported");
+    assert!(
+        result.success,
+        "compaction reported failure: {}",
+        result.error_message
+    );
+    let summary = result
+        .result
+        .as_ref()
+        .map(|r| r.summary.clone())
+        .unwrap_or_default();
+    assert!(
+        summary.contains("became"),
+        "completion carried no fragment counts: {summary}"
+    );
+    eprintln!("compaction result: {summary}");
+}
+
+/// A table with more versions than the floor is proposed, and running the job
+/// reports what it removed. The compaction test above leaves one behind.
+#[tokio::test]
+async fn cleans_up_old_versions() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_table(&url, "cleanme", 6, 4, false)
+        .await
+        .expect("seed a table with versions to clean");
+
+    let handler = CleanupVersionsHandler::new(url).with_fallback(fallback());
+    let recorder = Recorder::default();
+
+    let request = RunDetectionRequest {
+        request_id: "detect-cleanup".to_string(),
+        job_type: "lance_cleanup_versions".to_string(),
+        worker_config_values: int_config("min_versions_to_keep", 2),
+        ..Default::default()
+    };
+    handler
+        .detect(&request, &recorder)
+        .await
+        .expect("detection failed");
+    let proposals = recorder.proposals.lock().unwrap().clone();
+    let proposal = proposals
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("no cleanup proposal for the seeded table");
+
+    // Retain nothing, so every version outside the current one is fair game and
+    // the job has something to report rather than a no-op.
+    let execute = ExecuteJobRequest {
+        request_id: "execute-cleanup".to_string(),
+        job: Some(JobSpec {
+            job_id: "job-cleanup".to_string(),
+            job_type: "lance_cleanup_versions".to_string(),
+            parameters: proposal.parameters.clone(),
+            ..Default::default()
+        }),
+        worker_config_values: int_config("retain_hours", 0),
+        ..Default::default()
+    };
+    handler
+        .execute(&execute, &recorder)
+        .await
+        .expect("cleanup failed");
+
+    let completed = recorder.completed.lock().unwrap().clone();
+    let result = completed.last().expect("no completion reported");
+    assert!(
+        result.success,
+        "cleanup reported failure: {}",
+        result.error_message
+    );
+    eprintln!(
+        "cleanup result: {}",
+        result.result.as_ref().unwrap().summary
+    );
+}
+
+/// A table with no indices has nothing to optimize, so detection proposes
+/// nothing rather than queueing work that would do nothing.
+#[tokio::test]
+async fn skips_tables_without_indices() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_table(&url, "noindex", 2, 8, false)
+        .await
+        .expect("seed a table without an index");
+
+    let handler = OptimizeIndicesHandler::new(url).with_fallback(fallback());
+    let recorder = Recorder::default();
+
+    let request = RunDetectionRequest {
+        request_id: "detect-indices".to_string(),
+        job_type: "lance_optimize_indices".to_string(),
+        worker_config_values: int_config("max_unindexed_rows", 1),
+        ..Default::default()
+    };
+    handler
+        .detect(&request, &recorder)
+        .await
+        .expect("detection failed");
+    // Judge this table only: the catalog holds every other test's tables too,
+    // and an indexed one with uncovered rows is supposed to be proposed.
+    assert!(
+        !recorder
+            .proposals
+            .lock()
+            .unwrap()
+            .iter()
+            .any(|p| p.summary.contains(encoded.as_str())),
+        "a table with no indices must not be proposed for reindexing"
+    );
+}
+
+/// The job with no Iceberg equivalent: rows appended after an index was built
+/// are invisible to a search of it until this runs. Needs a table with an index
+/// and rows outside it, which `indexed.py` in the scratchpad seeds.
+#[tokio::test]
+async fn reindexes_rows_an_index_does_not_cover() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_table(&url, "reindexme", 2, 512, true)
+        .await
+        .expect("seed an indexed table with uncovered rows");
+    let handler = OptimizeIndicesHandler::new(url).with_fallback(fallback());
+    let recorder = Recorder::default();
+
+    let request = RunDetectionRequest {
+        request_id: "detect-reindex".to_string(),
+        job_type: "lance_optimize_indices".to_string(),
+        worker_config_values: int_config("max_unindexed_rows", 100),
+        ..Default::default()
+    };
+    handler
+        .detect(&request, &recorder)
+        .await
+        .expect("detection failed");
+    let proposals = recorder.proposals.lock().unwrap().clone();
+    let proposal = proposals
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("the seeded indexed table was not proposed");
+
+    let execute = ExecuteJobRequest {
+        request_id: "execute-reindex".to_string(),
+        job: Some(JobSpec {
+            job_id: "job-reindex".to_string(),
+            job_type: "lance_optimize_indices".to_string(),
+            parameters: proposal.parameters.clone(),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    handler
+        .execute(&execute, &recorder)
+        .await
+        .expect("reindex failed");
+
+    let completed = recorder.completed.lock().unwrap().clone();
+    let result = completed.last().expect("no completion reported");
+    assert!(
+        result.success,
+        "reindex reported failure: {}",
+        result.error_message
+    );
+    let output = &result.result.as_ref().unwrap().output_values;
+    let after = match output
+        .get("unindexed_rows_after")
+        .and_then(|v| v.kind.as_ref())
+    {
+        Some(Kind::Int64Value(value)) => *value,
+        other => panic!("no unindexed_rows_after in {other:?}"),
+    };
+    assert_eq!(
+        after, 0,
+        "rows are still outside the index after optimizing"
+    );
+    eprintln!(
+        "reindex result: {}",
+        result.result.as_ref().unwrap().summary
+    );
+}
+
+/// The UI's whole reason for asking a worker: admin cannot read a Lance table,
+/// so the rows have to come back already rendered.
+#[tokio::test]
+async fn previews_rows_of_a_table() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("set WEED_LANCE_NAMESPACE to run this test");
+        return;
+    };
+    seed_table(&url, "previewme", 2, 3, false)
+        .await
+        .expect("seed a table to preview");
+
+    let provider = LancePreview::new(url, fallback());
+    let id = vec!["vec".to_string(), "ml".to_string(), "previewme".to_string()];
+    let preview = provider.preview(&id, 4).await.expect("preview the table");
+
+    assert_eq!(preview.columns, vec!["id".to_string(), "vec".to_string()]);
+    assert_eq!(preview.total_rows, 6, "total is the table, not the sample");
+    assert_eq!(preview.rows.len(), 4, "row_limit bounds the sample");
+    assert!(
+        preview.rows[0][1].starts_with('['),
+        "a vector column should render as a list, got {:?}",
+        preview.rows[0][1]
+    );
+}
+
+/// The claim that removed managed versioning is not "one writer wins the
+/// conditional PUT" - that is only the mechanism. It is that concurrent writers
+/// lose nothing: the loser sees the conflict, rebases, and commits again. Eight
+/// writers appending at once must leave all eight batches in the table.
+#[tokio::test]
+async fn concurrent_writers_keep_every_commit() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("set WEED_LANCE_NAMESPACE to run this test");
+        return;
+    };
+    const WRITERS: i64 = 8;
+    const ROWS_EACH: i64 = 4;
+
+    seed_table(&url, "racers", 1, ROWS_EACH as usize, false)
+        .await
+        .expect("seed the table the writers will append to");
+
+    let client = NamespaceClient::new(url.clone());
+    let id = vec!["vec".to_string(), "ml".to_string(), "racers".to_string()];
+    let description = client.describe_table(&id).await.expect("describe");
+    let mut options = description.storage_options.clone();
+    options.extend(fallback());
+
+    let writes = (0..WRITERS).map(|writer| {
+        let location = description.location.clone();
+        let options = options.clone();
+        tokio::spawn(
+            async move { append_rows(&location, &options, writer * 1000, ROWS_EACH).await },
+        )
+    });
+    for (writer, handle) in writes.enumerate() {
+        handle
+            .await
+            .expect("writer panicked")
+            .unwrap_or_else(|err| panic!("writer {writer} failed to commit: {err:#}"));
+    }
+
+    let table = weed_lance_worker::dataset::open(&client, &id, &fallback())
+        .await
+        .expect("reopen the table");
+    let rows = table.dataset.count_rows(None).await.expect("count rows");
+    let expected = (ROWS_EACH + WRITERS * ROWS_EACH) as usize;
+    assert_eq!(
+        rows, expected,
+        "concurrent commits lost data: {rows} rows, want {expected}"
+    );
+}
+
+/// Appends one batch to an existing dataset, the way an independent writer would.
+async fn append_rows(
+    location: &str,
+    options: &HashMap<String, String>,
+    first_id: i64,
+    rows: i64,
+) -> Result<()> {
+    use arrow_array::{
+        FixedSizeListArray, Float32Array, Int64Array, RecordBatch, RecordBatchIterator,
+    };
+    use arrow_schema::{DataType, Field, Schema};
+    use lance::dataset::{Dataset, WriteMode, WriteParams};
+    use lance::io::{ObjectStoreParams, StorageOptionsAccessor};
+    use std::sync::Arc;
+
+    const DIM: i32 = 16;
+    let schema = Arc::new(Schema::new(vec![
+        Field::new("id", DataType::Int64, false),
+        Field::new(
+            "vec",
+            DataType::FixedSizeList(Arc::new(Field::new("item", DataType::Float32, true)), DIM),
+            false,
+        ),
+    ]));
+    let ids: Vec<i64> = (0..rows).map(|r| first_id + r).collect();
+    let values: Vec<f32> = ids
+        .iter()
+        .flat_map(|id| (0..DIM).map(move |d| (*id as f32) + d as f32))
+        .collect();
+    let vectors = FixedSizeListArray::new(
+        Arc::new(Field::new("item", DataType::Float32, true)),
+        DIM,
+        Arc::new(Float32Array::from(values)),
+        None,
+    );
+    let batch = RecordBatch::try_new(
+        schema.clone(),
+        vec![Arc::new(Int64Array::from(ids)), Arc::new(vectors)],
+    )?;
+    let params = WriteParams {
+        mode: WriteMode::Append,
+        store_params: Some(ObjectStoreParams {
+            storage_options_accessor: Some(Arc::new(StorageOptionsAccessor::with_static_options(
+                options.clone(),
+            ))),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    Dataset::write(
+        RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
+        location,
+        Some(params),
+    )
+    .await?;
+    Ok(())
+}

seaweed-worker/crates/lance/tests/lifecycle.rs

@@ -0,0 +1,123 @@
+//! Maintains one table someone else wrote, and nothing else.
+//!
+//! The Go suite in test/s3tables/lifecycle drives a table through its whole
+//! life - declared through the namespace, written by a Lance client, maintained,
+//! read again, dropped - and this is the maintenance step of it, so that step
+//! goes through the handlers a deployed worker runs rather than through the two
+//! lance calls they wrap. The table to work on comes in as WEED_LANCE_TABLE
+//! because the table is the Go test's, seeded and checked there.
+//!
+//! Skipped, like the rest of this crate's integration tests, unless
+//! WEED_LANCE_NAMESPACE names a live gateway.
+
+use std::collections::HashMap;
+
+use seaweed_worker_core::JobHandler;
+use seaweed_worker_core::pb::{ConfigValue, ExecuteJobRequest, JobSpec, RunDetectionRequest};
+use weed_lance_worker::jobs::cleanup::{self, CleanupVersionsHandler};
+use weed_lance_worker::jobs::compact::{CompactHandler, JOB_TYPE as COMPACT_JOB_TYPE};
+
+mod common;
+use common::{Recorder, fallback, int_config, namespace_url};
+
+fn table() -> Option<String> {
+    std::env::var("WEED_LANCE_TABLE")
+        .ok()
+        .filter(|s| !s.is_empty())
+}
+
+/// Compacts the named table's fragments, then drops the versions compaction
+/// left behind. Both go through the handler's own detect-then-execute path: a
+/// proposal the worker would not have made is not one worth running.
+#[tokio::test]
+async fn maintains_the_named_table() {
+    let (Some(url), Some(name)) = (namespace_url(), table()) else {
+        eprintln!("WEED_LANCE_NAMESPACE or WEED_LANCE_TABLE is unset, skipping");
+        return;
+    };
+
+    // A table written as a fragment per append is the case this exists for, so
+    // anything above one fragment is worth merging here.
+    let mut config = int_config("min_fragments", 2);
+    config.extend(int_config("target_rows_per_fragment", 1_048_576));
+    let compact = CompactHandler::new(url.clone()).with_fallback(fallback());
+    run(&compact, COMPACT_JOB_TYPE, &name, config, true).await;
+
+    // Keep the current version and retain nothing else, so the versions
+    // compaction superseded are actually removed rather than counted.
+    let mut config = int_config("min_versions_to_keep", 1);
+    config.extend(int_config("retain_hours", 0));
+    let cleanup = CleanupVersionsHandler::new(url).with_fallback(fallback());
+    run(&cleanup, cleanup::JOB_TYPE, &name, config, true).await;
+}
+
+/// Detects, finds the proposal for `name`, and executes it. The config carries
+/// every key both halves read; each ignores what it does not know.
+async fn run<H: JobHandler>(
+    handler: &H,
+    job_type: &str,
+    name: &str,
+    config: HashMap<String, ConfigValue>,
+    required: bool,
+) {
+    let recorder = Recorder::default();
+    handler
+        .detect(
+            &RunDetectionRequest {
+                request_id: format!("detect-{job_type}"),
+                job_type: job_type.to_string(),
+                worker_config_values: config.clone(),
+                ..Default::default()
+            },
+            &recorder,
+        )
+        .await
+        .unwrap_or_else(|err| panic!("{job_type} detection failed: {err}"));
+
+    let proposals = recorder.proposals.lock().unwrap().clone();
+    let Some(proposal) = proposals.iter().find(|p| p.summary.contains(name)) else {
+        assert!(
+            !required,
+            "{job_type} proposed nothing for {name}, out of {} proposals",
+            proposals.len()
+        );
+        eprintln!("{job_type}: nothing to do for {name}");
+        return;
+    };
+
+    handler
+        .execute(
+            &ExecuteJobRequest {
+                request_id: format!("execute-{job_type}"),
+                job: Some(JobSpec {
+                    job_id: format!("job-{job_type}"),
+                    job_type: job_type.to_string(),
+                    parameters: proposal.parameters.clone(),
+                    ..Default::default()
+                }),
+                worker_config_values: config,
+                ..Default::default()
+            },
+            &recorder,
+        )
+        .await
+        .unwrap_or_else(|err| panic!("{job_type} execution failed: {err}"));
+
+    let completed = recorder.completed.lock().unwrap().clone();
+    let result = completed
+        .last()
+        .unwrap_or_else(|| panic!("{job_type} reported no completion"));
+    assert!(
+        result.success,
+        "{job_type} reported failure: {}",
+        result.error_message
+    );
+    eprintln!(
+        "{job_type}: {}",
+        result
+            .result
+            .as_ref()
+            .map(|r| r.summary.clone())
+            .unwrap_or_default()
+    );
+}

seaweed-worker/crates/lance/tests/sort.rs

@@ -0,0 +1,504 @@
+//! Drives the sort handler against a live namespace.
+//!
+//! Skipped unless WEED_LANCE_NAMESPACE names one, like the other integration
+//! tests here: the job rewrites every fragment of a real table and commits, and
+//! the commit is the half worth testing.
+
+use anyhow::Result;
+use seaweed_worker_core::JobHandler;
+use seaweed_worker_core::pb::{
+    ConfigValue, ExecuteJobRequest, JobSpec, RunDetectionRequest, config_value::Kind,
+};
+use weed_lance_worker::catalog::NamespaceClient;
+use weed_lance_worker::jobs::sort::{JOB_TYPE, SortHandler};
+
+mod common;
+use common::{Recorder, fallback, namespace_url};
+
+/// One live gateway and one shared catalog, and `list_all_tables` sweeps
+/// everything, so these tests take a lock the way the compaction ones do.
+static GATEWAY: tokio::sync::Mutex<()> = tokio::sync::Mutex::const_new(());
+
+fn config(values: &[(&str, ConfigValue)]) -> std::collections::HashMap<String, ConfigValue> {
+    values
+        .iter()
+        .map(|(name, value)| (name.to_string(), value.clone()))
+        .collect()
+}
+
+fn int(value: i64) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::Int64Value(value)),
+    }
+}
+
+fn text(value: &str) -> ConfigValue {
+    ConfigValue {
+        kind: Some(Kind::StringValue(value.to_string())),
+    }
+}
+
+/// Declares a table and writes `batches` appends whose ids descend, so the rows
+/// land in an order no sort would produce. `declared` is written into the
+/// dataset's own configuration, which is where a table says how it wants to be
+/// sorted.
+async fn seed_unsorted_table(
+    url: &str,
+    name: &str,
+    batches: i64,
+    rows_each: i64,
+    declared: Option<&str>,
+) -> Result<String> {
+    use arrow_array::{Int64Array, RecordBatch, RecordBatchIterator};
+    use arrow_schema::{DataType, Field, Schema};
+    use lance::dataset::{Dataset, WriteMode, WriteParams};
+    use lance::io::{ObjectStoreParams, StorageOptionsAccessor};
+    use std::sync::Arc;
+
+    let http = reqwest::Client::new();
+    for parent in ["vec", "vec$ml"] {
+        http.post(format!("{url}/v1/namespace/{parent}/create"))
+            .json(&serde_json::json!({"mode": "EXIST_OK"}))
+            .send()
+            .await?
+            .error_for_status()?;
+    }
+    let encoded = format!("vec$ml${name}");
+    http.post(format!("{url}/v1/table/{encoded}/declare"))
+        .json(&serde_json::json!({}))
+        .send()
+        .await?
+        .error_for_status()?;
+
+    let client = NamespaceClient::new(url.to_string());
+    let id = vec!["vec".to_string(), "ml".to_string(), name.to_string()];
+    let description = client.describe_table(&id).await?;
+    let mut options = description.storage_options.clone();
+    options.extend(fallback());
+
+    let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int64, false)]));
+    let total = batches * rows_each;
+    let mut dataset = None;
+    for batch_index in 0..batches {
+        // Ids descend across and within the batches, so the physical order is
+        // the reverse of the order the table declares.
+        let ids: Vec<i64> = (0..rows_each)
+            .map(|row| total - 1 - (batch_index * rows_each + row))
+            .collect();
+        let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int64Array::from(ids))])?;
+        let params = WriteParams {
+            mode: if batch_index == 0 {
+                WriteMode::Overwrite
+            } else {
+                WriteMode::Append
+            },
+            store_params: Some(ObjectStoreParams {
+                storage_options_accessor: Some(Arc::new(
+                    StorageOptionsAccessor::with_static_options(options.clone()),
+                )),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        dataset = Some(
+            Dataset::write(
+                RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
+                description.location.as_str(),
+                Some(params),
+            )
+            .await?,
+        );
+    }
+
+    if let Some(declared) = declared {
+        let mut dataset = dataset.expect("at least one batch was written");
+        dataset
+            .update_config([("seaweedfs.sort.fields", declared)])
+            .await?;
+    }
+    Ok(encoded)
+}
+
+/// Reads the ids back in the order they are stored, which is what a sort is
+/// supposed to change.
+async fn stored_ids(url: &str, name: &str) -> Result<Vec<i64>> {
+    use arrow_array::Int64Array;
+    use futures::TryStreamExt;
+
+    let client = NamespaceClient::new(url.to_string());
+    let id = vec!["vec".to_string(), "ml".to_string(), name.to_string()];
+    let table = weed_lance_worker::dataset::open(&client, &id, &fallback()).await?;
+    let stream = table.dataset.scan().try_into_stream().await?;
+    let batches: Vec<_> = stream.try_collect().await?;
+    let mut ids = Vec::new();
+    for batch in batches {
+        let column = batch
+            .column_by_name("id")
+            .expect("the seeded table has an id column")
+            .as_any()
+            .downcast_ref::<Int64Array>()
+            .expect("id is an Int64 column")
+            .clone();
+        ids.extend(column.values().iter().copied());
+    }
+    Ok(ids)
+}
+
+/// The whole loop: a table declaring an order its rows do not follow is
+/// proposed, sorting it stores the rows in that order, and the marker the same
+/// commit carries is what stops the next sweep proposing it again.
+#[tokio::test]
+async fn sorts_a_table_and_stops_proposing_it() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_unsorted_table(&url, "sortme", 4, 250, Some("id asc"))
+        .await
+        .expect("seed an unsorted table");
+
+    let before = stored_ids(&url, "sortme").await.expect("read the ids back");
+    assert!(
+        before.windows(2).any(|pair| pair[0] > pair[1]),
+        "the seeded table was already sorted, so the test proves nothing"
+    );
+
+    let handler = SortHandler::new(url.clone()).with_fallback(fallback());
+    let recorder = Recorder::default();
+    let detection = RunDetectionRequest {
+        request_id: "detect-sort".to_string(),
+        job_type: JOB_TYPE.to_string(),
+        worker_config_values: config(&[("min_unsorted_rows", int(1))]),
+        ..Default::default()
+    };
+    handler
+        .detect(&detection, &recorder)
+        .await
+        .expect("detection failed");
+
+    let proposal = recorder
+        .proposals
+        .lock()
+        .unwrap()
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("the unsorted table was not proposed");
+    assert!(
+        proposal.summary.contains("id asc"),
+        "the proposal does not say what order it would use: {}",
+        proposal.summary
+    );
+
+    let execute = ExecuteJobRequest {
+        request_id: "execute-sort".to_string(),
+        job: Some(JobSpec {
+            job_id: "job-sort".to_string(),
+            job_type: JOB_TYPE.to_string(),
+            parameters: proposal.parameters.clone(),
+            ..Default::default()
+        }),
+        worker_config_values: config(&[
+            ("memory_budget_mb", int(64)),
+            ("max_rows_per_file", int(1024)),
+        ]),
+        ..Default::default()
+    };
+    handler
+        .execute(&execute, &recorder)
+        .await
+        .expect("sorting failed");
+
+    let completed = recorder.completed.lock().unwrap().clone();
+    let result = completed.last().expect("no completion reported");
+    assert!(
+        result.success,
+        "the sort reported failure: {}",
+        result.error_message
+    );
+
+    let after = stored_ids(&url, "sortme")
+        .await
+        .expect("read the sorted ids back");
+    assert_eq!(after.len(), before.len(), "the sort lost or invented rows");
+    assert!(
+        after.windows(2).all(|pair| pair[0] <= pair[1]),
+        "the rows are not stored in ascending id order"
+    );
+
+    // The marker rode in the same commit as the data, so a second sweep finds
+    // the table up to date without anything else having written it.
+    let second = Recorder::default();
+    handler
+        .detect(&detection, &second)
+        .await
+        .expect("second detection failed");
+    assert!(
+        !second
+            .proposals
+            .lock()
+            .unwrap()
+            .iter()
+            .any(|p| p.summary.contains(encoded.as_str())),
+        "a table that was just sorted must not be proposed again"
+    );
+    eprintln!("sort result: {}", result.result.as_ref().unwrap().summary);
+}
+
+/// A table that declares no order, with a worker configuring none either, is
+/// not something to rewrite on a guess.
+#[tokio::test]
+async fn leaves_a_table_that_declares_no_order_alone() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_unsorted_table(&url, "unsorted", 2, 16, None)
+        .await
+        .expect("seed a table with no declared order");
+
+    let handler = SortHandler::new(url.clone()).with_fallback(fallback());
+    let recorder = Recorder::default();
+    handler
+        .detect(
+            &RunDetectionRequest {
+                request_id: "detect-none".to_string(),
+                job_type: JOB_TYPE.to_string(),
+                worker_config_values: config(&[
+                    ("min_unsorted_rows", int(1)),
+                    ("sort_fields", text("")),
+                ]),
+                ..Default::default()
+            },
+            &recorder,
+        )
+        .await
+        .expect("detection failed");
+
+    assert!(
+        !recorder
+            .proposals
+            .lock()
+            .unwrap()
+            .iter()
+            .any(|p| p.summary.contains(encoded.as_str())),
+        "a table with no order named anywhere must not be proposed"
+    );
+}
+
+/// The case a row-count threshold cannot see: the table replaced by a write of
+/// its own, growing by far less than `min_unsorted_rows`. Every row moved, so
+/// the table is unsorted, and detection has to say so however small the delta.
+#[tokio::test]
+async fn proposes_a_replacement_that_barely_changes_the_row_count() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_unsorted_table(&url, "replaceme", 4, 250, Some("id asc"))
+        .await
+        .expect("seed an unsorted table");
+
+    let handler = SortHandler::new(url.clone()).with_fallback(fallback());
+    // A threshold no realistic append would cross, so nothing but the identity
+    // of the data can be what proposes this table.
+    let detection = RunDetectionRequest {
+        request_id: "detect-replace".to_string(),
+        job_type: JOB_TYPE.to_string(),
+        worker_config_values: config(&[("min_unsorted_rows", int(1_000_000))]),
+        ..Default::default()
+    };
+
+    let first = Recorder::default();
+    handler.detect(&detection, &first).await.expect("detection");
+    let proposal = first
+        .proposals
+        .lock()
+        .unwrap()
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("a never-sorted table was not proposed");
+
+    handler
+        .execute(
+            &ExecuteJobRequest {
+                request_id: "execute-replace".to_string(),
+                job: Some(JobSpec {
+                    job_id: "job-replace".to_string(),
+                    job_type: JOB_TYPE.to_string(),
+                    parameters: proposal.parameters.clone(),
+                    ..Default::default()
+                }),
+                worker_config_values: config(&[("max_rows_per_file", int(1024))]),
+                ..Default::default()
+            },
+            &first,
+        )
+        .await
+        .expect("sorting failed");
+
+    // Sorted, and left alone by the next sweep.
+    let after_sort = Recorder::default();
+    handler
+        .detect(&detection, &after_sort)
+        .await
+        .expect("detection");
+    assert!(
+        !after_sort
+            .proposals
+            .lock()
+            .unwrap()
+            .iter()
+            .any(|p| p.summary.contains(encoded.as_str())),
+        "a table that was just sorted must not be proposed again"
+    );
+
+    // Now replace the data: 1002 rows where there were 1000, shuffled again.
+    seed_unsorted_table(&url, "replaceme", 3, 334, Some("id asc"))
+        .await
+        .expect("replace the table's data");
+
+    let after_replace = Recorder::default();
+    handler
+        .detect(&detection, &after_replace)
+        .await
+        .expect("detection");
+    assert!(
+        after_replace
+            .proposals
+            .lock()
+            .unwrap()
+            .iter()
+            .any(|p| p.summary.contains(encoded.as_str())),
+        "a replaced table must be proposed even when the row count barely moved"
+    );
+}
+
+/// Deletions must not hide appended rows. This is the arithmetic against a real
+/// manifest rather than a hand-built one: the sorted fragments keep their files
+/// and lose most of their rows, a smaller batch is appended after them, and the
+/// table has to be proposed *because rows were appended* — not because it looks
+/// replaced, and not left alone because it now holds fewer rows than when it
+/// was sorted.
+#[tokio::test]
+async fn deletions_do_not_hide_appended_rows() {
+    let _gateway = GATEWAY.lock().await;
+    let Some(url) = namespace_url() else {
+        eprintln!("WEED_LANCE_NAMESPACE is unset, skipping");
+        return;
+    };
+    let encoded = seed_unsorted_table(&url, "deleteme", 4, 250, Some("id asc"))
+        .await
+        .expect("seed an unsorted table");
+
+    let handler = SortHandler::new(url.clone()).with_fallback(fallback());
+    let detection = RunDetectionRequest {
+        request_id: "detect-delete".to_string(),
+        job_type: JOB_TYPE.to_string(),
+        worker_config_values: config(&[("min_unsorted_rows", int(100))]),
+        ..Default::default()
+    };
+
+    let recorder = Recorder::default();
+    handler
+        .detect(&detection, &recorder)
+        .await
+        .expect("detection");
+    let proposal = recorder
+        .proposals
+        .lock()
+        .unwrap()
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("a never-sorted table was not proposed");
+    handler
+        .execute(
+            &ExecuteJobRequest {
+                request_id: "execute-delete".to_string(),
+                job: Some(JobSpec {
+                    job_id: "job-delete".to_string(),
+                    job_type: JOB_TYPE.to_string(),
+                    parameters: proposal.parameters.clone(),
+                    ..Default::default()
+                }),
+                worker_config_values: config(&[("max_rows_per_file", int(1024))]),
+                ..Default::default()
+            },
+            &recorder,
+        )
+        .await
+        .expect("sorting failed");
+
+    // Delete 900 of the 1000 sorted rows, then append 200 unsorted ones. The
+    // table is smaller than when it was sorted, so net growth reads as zero.
+    let client = NamespaceClient::new(url.clone());
+    let id = vec!["vec".to_string(), "ml".to_string(), "deleteme".to_string()];
+    let mut table = weed_lance_worker::dataset::open(&client, &id, &fallback())
+        .await
+        .expect("reopen the sorted table");
+    table
+        .dataset
+        .delete("id < 900")
+        .await
+        .expect("delete most of the sorted rows");
+    append_ids(&url, "deleteme", 5_000, 200)
+        .await
+        .expect("append rows after the sorted ones");
+
+    let after = Recorder::default();
+    handler.detect(&detection, &after).await.expect("detection");
+    let proposal = after
+        .proposals
+        .lock()
+        .unwrap()
+        .iter()
+        .find(|p| p.summary.contains(encoded.as_str()))
+        .cloned()
+        .expect("a table with 200 unsorted rows appended was not proposed");
+    assert!(
+        proposal.detail.contains("200 rows appended"),
+        "expected the appended rows to be counted, got {:?}",
+        proposal.detail
+    );
+}
+
+/// Appends one batch of ascending ids to an existing table.
+async fn append_ids(url: &str, name: &str, first_id: i64, rows: i64) -> Result<()> {
+    use arrow_array::{Int64Array, RecordBatch, RecordBatchIterator};
+    use arrow_schema::{DataType, Field, Schema};
+    use lance::dataset::{Dataset, WriteMode, WriteParams};
+    use lance::io::{ObjectStoreParams, StorageOptionsAccessor};
+    use std::sync::Arc;
+
+    let client = NamespaceClient::new(url.to_string());
+    let id = vec!["vec".to_string(), "ml".to_string(), name.to_string()];
+    let description = client.describe_table(&id).await?;
+    let mut options = description.storage_options.clone();
+    options.extend(fallback());
+
+    let schema = Arc::new(Schema::new(vec![Field::new("id", DataType::Int64, false)]));
+    let ids: Vec<i64> = (0..rows).map(|row| first_id + row).collect();
+    let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(Int64Array::from(ids))])?;
+    let params = WriteParams {
+        mode: WriteMode::Append,
+        store_params: Some(ObjectStoreParams {
+            storage_options_accessor: Some(Arc::new(StorageOptionsAccessor::with_static_options(
+                options,
+            ))),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    Dataset::write(
+        RecordBatchIterator::new(vec![Ok(batch)], schema.clone()),
+        description.location.as_str(),
+        Some(params),
+    )
+    .await?;
+    Ok(())
+}

seaweed-worker/crates/sort/Cargo.toml

@@ -0,0 +1,16 @@
+[package]
+name = "seaweed-worker-sort"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+description = "The sort specification shared by SeaweedFS sorting jobs"
+
+[lib]
+name = "seaweed_worker_sort"
+
+[dependencies]
+seaweed-worker-core = { path = "../core" }
+anyhow.workspace = true
+
+[lints]
+workspace = true

seaweed-worker/crates/sort/src/lib.rs

@@ -0,0 +1,612 @@
+//! The sort specification the sorting jobs share.
+//!
+//! A worker sorts a table by the fields the table itself declares, and by the
+//! fields its own configuration names for a table that declares none. Both
+//! halves of that rule, and the spelling of a field list, live here rather than
+//! in each job: a Lance job and an Iceberg job that disagreed about what
+//! "id desc nulls-first" means would be two features wearing one name.
+//!
+//! The spelling is deliberately the one `weed/worker/tasks/iceberg` already
+//! writes into an Iceberg snapshot's `sort-fields` summary, so the same order
+//! reads the same way whichever format holds the table and whichever language
+//! sorted it.
+
+use std::collections::{HashMap, HashSet};
+use std::fmt;
+
+use anyhow::{Context, Result, bail};
+use seaweed_worker_core::config_form::{number_field, text_field};
+use seaweed_worker_core::pb::ConfigField;
+
+/// Worker configuration keys. Shared so the Iceberg and Lance forms offer an
+/// operator the same names for the same settings.
+pub const CONFIG_SORT_FIELDS: &str = "sort_fields";
+pub const CONFIG_MIN_UNSORTED_ROWS: &str = "min_unsorted_rows";
+pub const CONFIG_MEMORY_BUDGET_MB: &str = "memory_budget_mb";
+pub const CONFIG_MAX_ROWS_PER_FILE: &str = "max_rows_per_file";
+
+/// The order an operator declares on a table. The worker only ever reads this
+/// one.
+pub const DECLARED_FIELDS_KEY: &str = "seaweedfs.sort.fields";
+
+/// What the worker recorded about the sort it last performed. Kept apart from
+/// the declaration above so that "the operator asked for this order" and "the
+/// worker achieved this order" stay distinguishable — comparing them is how a
+/// changed order is noticed.
+pub const SORTED_FIELDS_KEY: &str = "seaweedfs.sort.sorted_fields";
+pub const SORTED_ROWS_KEY: &str = "seaweedfs.sort.sorted_rows";
+/// How many fragments the sort wrote, and a digest of their data file names.
+/// Together they identify the data the sort produced — see [`verdict`] for why
+/// a version number cannot.
+pub const SORTED_FRAGMENTS_KEY: &str = "seaweedfs.sort.sorted_fragments";
+pub const SORTED_DIGEST_KEY: &str = "seaweedfs.sort.sorted_digest";
+
+/// One field of a sort order.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SortField {
+    pub path: String,
+    pub descending: bool,
+    pub nulls_first: bool,
+}
+
+/// A whole sort order, in the order the fields are compared.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct SortSpec {
+    pub fields: Vec<SortField>,
+}
+
+impl SortSpec {
+    /// Reads "user_id asc nulls-last, ts desc" into a spec. An empty string is
+    /// not an error: it is how an operator says nothing, and the caller decides
+    /// what that means.
+    ///
+    /// Direction defaults to ascending and null order to Iceberg's default for
+    /// the direction — nulls first ascending, nulls last descending — because
+    /// this spelling is shared with tables that already have that rule.
+    pub fn parse(text: &str) -> Result<Option<Self>> {
+        let mut fields = Vec::new();
+        let mut seen = HashSet::new();
+        for entry in text.split(',') {
+            let entry = entry.trim();
+            if entry.is_empty() {
+                continue;
+            }
+            let field =
+                parse_field(entry).with_context(|| format!("read the sort field {entry:?}"))?;
+            // Compared exactly, not case-folded: Arrow schemas are
+            // case-sensitive, so `id` and `ID` can be two real columns and
+            // folding them together would reject a valid order. The Iceberg
+            // job, whose format resolves names case-insensitively, does its own
+            // check against the table schema.
+            if !seen.insert(field.path.clone()) {
+                bail!("the sort field {:?} is named twice", field.path);
+            }
+            fields.push(field);
+        }
+        if fields.is_empty() {
+            return Ok(None);
+        }
+        Ok(Some(Self { fields }))
+    }
+}
+
+/// Renders a spec back, always spelling out direction and null order so that a
+/// recorded order round-trips to the same thing it was parsed from.
+impl fmt::Display for SortSpec {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        for (i, field) in self.fields.iter().enumerate() {
+            if i > 0 {
+                write!(f, ", ")?;
+            }
+            write!(
+                f,
+                "{} {} {}",
+                field.path,
+                if field.descending { "desc" } else { "asc" },
+                if field.nulls_first {
+                    "nulls-first"
+                } else {
+                    "nulls-last"
+                }
+            )?;
+        }
+        Ok(())
+    }
+}
+
+fn parse_field(entry: &str) -> Result<SortField> {
+    let mut tokens = entry.split_whitespace();
+    let path = tokens
+        .next()
+        .expect("split_whitespace yields one token for a non-empty entry")
+        .to_string();
+
+    let mut descending = None;
+    let mut nulls_first = None;
+    for token in tokens {
+        // Underscores are accepted because that is how the same words are
+        // spelled in configuration keys, and refusing them would only teach an
+        // operator that the two spellings are different things.
+        match token.to_ascii_lowercase().replace('_', "-").as_str() {
+            "asc" | "ascending" if descending.is_none() => descending = Some(false),
+            "desc" | "descending" if descending.is_none() => descending = Some(true),
+            "nulls-first" if nulls_first.is_none() => nulls_first = Some(true),
+            "nulls-last" if nulls_first.is_none() => nulls_first = Some(false),
+            other => bail!("{other:?} is not a direction or a null order"),
+        }
+    }
+
+    let descending = descending.unwrap_or(false);
+    Ok(SortField {
+        path,
+        descending,
+        nulls_first: nulls_first.unwrap_or(!descending),
+    })
+}
+
+/// The order to sort a table by: what the table declares wins, and the worker's
+/// configuration is the fallback for a table that declares nothing. Neither one
+/// is not an error — it means this is not a table the operator wants sorted.
+///
+/// A declaration that cannot be read is an error rather than a reason to fall
+/// back: sorting by the worker's default order instead of the one the table
+/// asked for would silently rewrite the table the wrong way.
+pub fn resolve(declared: Option<&str>, configured: &str) -> Result<Option<SortSpec>> {
+    if let Some(declared) = declared
+        && let Some(spec) = SortSpec::parse(declared).context("read the table's declared order")?
+    {
+        return Ok(Some(spec));
+    }
+    SortSpec::parse(configured).context("read the configured sort order")
+}
+
+/// One fragment as detection sees it.
+///
+/// `rows` is the live row count the manifest records — physical rows less the
+/// deletions — and is None when the manifest does not say, which older
+/// fragments do not.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct FragmentSummary {
+    pub files: Vec<String>,
+    pub rows: Option<u64>,
+}
+
+/// Every data file of these fragments, in order.
+pub fn data_files(fragments: &[FragmentSummary]) -> Vec<&str> {
+    fragments
+        .iter()
+        .flat_map(|fragment| fragment.files.iter().map(String::as_str))
+        .collect()
+}
+
+/// A digest of data file names, in the order the fragments hold them.
+///
+/// FNV-1a rather than the standard library's hasher, whose output is explicitly
+/// not stable across releases: a marker that hashed differently after a
+/// toolchain upgrade would re-sort every table once, silently.
+pub fn digest<S: AsRef<str>>(files: &[S]) -> String {
+    let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
+    for file in files {
+        for byte in file.as_ref().as_bytes().iter().chain(std::iter::once(&0u8)) {
+            hash ^= *byte as u64;
+            hash = hash.wrapping_mul(0x1000_0000_01b3);
+        }
+    }
+    format!("{hash:016x}")
+}
+
+/// What the worker recorded on a table the last time it sorted it.
+#[derive(Debug, Clone, Default, PartialEq, Eq)]
+pub struct SortState {
+    pub fields: Option<String>,
+    pub rows: Option<u64>,
+    /// The fragments the sort wrote: how many, and a digest of their data file
+    /// names. Data file names are chosen before the commit, which is what makes
+    /// them usable in a marker the same commit carries.
+    pub fragments: Option<usize>,
+    pub digest: Option<String>,
+}
+
+impl SortState {
+    /// Reads the marker out of a table's key/value configuration. A key that
+    /// will not parse is treated as absent, which asks for a sort rather than
+    /// skipping the table: the cost of a needless sort is time, and the cost of
+    /// skipping is a table that silently never gets sorted again.
+    pub fn from_config(config: &HashMap<String, String>) -> Self {
+        Self {
+            fields: config.get(SORTED_FIELDS_KEY).cloned(),
+            rows: config.get(SORTED_ROWS_KEY).and_then(|v| v.parse().ok()),
+            fragments: config
+                .get(SORTED_FRAGMENTS_KEY)
+                .and_then(|v| v.parse().ok()),
+            digest: config.get(SORTED_DIGEST_KEY).cloned(),
+        }
+    }
+
+    /// The marker to write for a sort that just finished.
+    pub fn record(
+        spec: &SortSpec,
+        fragments: &[FragmentSummary],
+        rows: u64,
+    ) -> HashMap<String, String> {
+        HashMap::from([
+            (SORTED_FIELDS_KEY.to_string(), spec.to_string()),
+            (SORTED_ROWS_KEY.to_string(), rows.to_string()),
+            (
+                SORTED_FRAGMENTS_KEY.to_string(),
+                fragments.len().to_string(),
+            ),
+            (
+                SORTED_DIGEST_KEY.to_string(),
+                digest(&data_files(fragments)),
+            ),
+        ])
+    }
+}
+
+/// Why a table does or does not need sorting.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum SortVerdict {
+    NeverSorted,
+    OrderChanged,
+    RowsAppended(u64),
+    /// The table was committed to since the sort, and the row count did not
+    /// grow — the data is not the data that was sorted.
+    Rewritten,
+    UpToDate,
+}
+
+impl SortVerdict {
+    pub fn needs_sort(&self) -> bool {
+        !matches!(self, Self::UpToDate)
+    }
+
+    pub fn reason(&self) -> String {
+        match self {
+            Self::NeverSorted => "never sorted".to_string(),
+            Self::OrderChanged => "the declared order changed since the last sort".to_string(),
+            Self::RowsAppended(rows) => format!("{rows} rows appended since the last sort"),
+            Self::Rewritten => "the data was replaced since the last sort".to_string(),
+            Self::UpToDate => "sorted".to_string(),
+        }
+    }
+}
+
+/// Decides whether a table is worth sorting.
+///
+/// The question detection actually has to answer is "is this still the data the
+/// sort wrote?", and neither the row count nor the dataset version can answer
+/// it. A rewrite that leaves the row count where it was is invisible to the
+/// first, and the second cannot be recorded at all: the commit carrying the
+/// marker is the commit that creates the version, and a commit that rebases
+/// past a conflict lands on a different number again.
+///
+/// Data file names can answer it. They are chosen before the commit, so the
+/// marker the same commit carries can name them, and they are stable no matter
+/// which version the commit ends up as. So:
+///
+/// - the same files → the table is exactly what the sort left;
+/// - the sorted files still there, with more after them → rows were appended,
+///   and only then is `min_unsorted_rows` consulted, which is the churn that
+///   threshold exists to damp;
+/// - anything else → the data was replaced, whatever the row count says.
+///
+/// Appends only ever add fragments after the existing ones, and lance hands out
+/// fragment ids monotonically even across an overwrite, so "the sorted files
+/// are still there" is a prefix test on the files in fragment order.
+///
+/// Deletes leave the data files alone and write a deletion file beside them, so
+/// they read as unchanged here — which is right: deleting rows does not unsort
+/// the ones that remain.
+pub fn verdict(
+    spec: &SortSpec,
+    state: &SortState,
+    fragments: &[FragmentSummary],
+    min_unsorted_rows: u64,
+) -> SortVerdict {
+    let (Some(recorded_digest), Some(recorded_fragments)) =
+        (state.digest.as_deref(), state.fragments)
+    else {
+        return SortVerdict::NeverSorted;
+    };
+    if state.fields.as_deref() != Some(spec.to_string().as_str()) {
+        return SortVerdict::OrderChanged;
+    }
+    if digest(&data_files(fragments)) == recorded_digest {
+        return SortVerdict::UpToDate;
+    }
+
+    // Not the same files. Only a prefix match means the sorted data survived
+    // and the rest was appended after it.
+    if recorded_fragments > fragments.len()
+        || digest(&data_files(&fragments[..recorded_fragments])) != recorded_digest
+    {
+        return SortVerdict::Rewritten;
+    }
+
+    // How many rows are in the fragments appended after the sorted ones.
+    //
+    // Not the table's net growth against the recorded row count: rows deleted
+    // from the sorted fragments hide appended rows one for one, and enough of
+    // them hide every appended row, leaving a table that reads as sorted with
+    // an unsorted tail. Counting the appended fragments themselves is the same
+    // arithmetic the threshold was always meant to do.
+    let mut appended = 0u64;
+    for fragment in &fragments[recorded_fragments..] {
+        // A fragment whose length the manifest does not record cannot be
+        // counted, and a table that cannot be judged is one to sort rather than
+        // one to leave alone forever.
+        let Some(rows) = fragment.rows else {
+            return SortVerdict::Rewritten;
+        };
+        appended += rows;
+    }
+    if appended >= min_unsorted_rows.max(1) {
+        return SortVerdict::RowsAppended(appended);
+    }
+    SortVerdict::UpToDate
+}
+
+/// The settings every sorting job offers, so admin renders one form for the
+/// same behaviour whichever format the job sorts.
+pub fn config_fields() -> Vec<ConfigField> {
+    vec![
+        text_field(
+            CONFIG_SORT_FIELDS,
+            "Sort fields",
+            "Order for tables that declare none of their own, as \"user_id asc, ts desc nulls-last\". Empty sorts only the tables that declare an order.",
+            "user_id asc, ts desc",
+        ),
+        number_field(
+            CONFIG_MIN_UNSORTED_ROWS,
+            "Minimum unsorted rows",
+            "Rows appended since the last sort before a table is worth sorting again",
+            1,
+            1_000_000_000,
+        ),
+        number_field(
+            CONFIG_MEMORY_BUDGET_MB,
+            "Memory budget (MB)",
+            "Memory the sort may hold before it spills runs to disk. A table larger than this sorts more slowly rather than failing.",
+            64,
+            1_048_576,
+        ),
+        number_field(
+            CONFIG_MAX_ROWS_PER_FILE,
+            "Maximum rows per file",
+            "Rows to write into each output file of a sorted rewrite",
+            1024,
+            16_777_216,
+        ),
+    ]
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parses_direction_and_null_order() {
+        let spec = SortSpec::parse("user_id, ts desc, name asc nulls-last")
+            .unwrap()
+            .unwrap();
+        assert_eq!(
+            spec.fields,
+            vec![
+                // Ascending defaults to nulls first, descending to nulls last,
+                // which is the rule the Iceberg tables sharing this spelling use.
+                SortField {
+                    path: "user_id".into(),
+                    descending: false,
+                    nulls_first: true
+                },
+                SortField {
+                    path: "ts".into(),
+                    descending: true,
+                    nulls_first: false
+                },
+                SortField {
+                    path: "name".into(),
+                    descending: false,
+                    nulls_first: false
+                },
+            ]
+        );
+    }
+
+    #[test]
+    fn renders_back_to_something_that_parses_the_same() {
+        let spec = SortSpec::parse("a, b desc nulls_first").unwrap().unwrap();
+        let rendered = spec.to_string();
+        assert_eq!(rendered, "a asc nulls-first, b desc nulls-first");
+        assert_eq!(SortSpec::parse(&rendered).unwrap().unwrap(), spec);
+    }
+
+    #[test]
+    fn empty_is_no_spec_rather_than_an_error() {
+        assert!(SortSpec::parse("").unwrap().is_none());
+        assert!(SortSpec::parse("  ,  ").unwrap().is_none());
+    }
+
+    #[test]
+    fn rejects_nonsense_and_repeats() {
+        assert!(SortSpec::parse("a sideways").is_err());
+        assert!(SortSpec::parse("a asc desc").is_err());
+        assert!(SortSpec::parse("a, a desc").is_err());
+    }
+
+    #[test]
+    fn the_table_wins_and_configuration_fills_in() {
+        let configured = "b desc";
+        let from_table = resolve(Some("a asc"), configured).unwrap().unwrap();
+        assert_eq!(from_table.fields[0].path, "a");
+
+        let from_config = resolve(None, configured).unwrap().unwrap();
+        assert_eq!(from_config.fields[0].path, "b");
+
+        let from_config_again = resolve(Some("  "), configured).unwrap().unwrap();
+        assert_eq!(from_config_again.fields[0].path, "b");
+
+        assert!(resolve(None, "").unwrap().is_none());
+    }
+
+    // Falling back would sort the table by an order nobody asked for.
+    #[test]
+    fn a_broken_declaration_is_an_error_not_a_fallback() {
+        assert!(resolve(Some("a sideways"), "b desc").is_err());
+    }
+
+    /// One fragment per name, one file each, with a known row count.
+    fn frags(names: &[(&str, u64)]) -> Vec<FragmentSummary> {
+        names
+            .iter()
+            .map(|(name, rows)| FragmentSummary {
+                files: vec![name.to_string()],
+                rows: Some(*rows),
+            })
+            .collect()
+    }
+
+    #[test]
+    fn verdicts_follow_the_marker() {
+        let spec = SortSpec::parse("a asc").unwrap().unwrap();
+        let sorted = frags(&[("a.lance", 500), ("b.lance", 500)]);
+
+        let fresh = SortState::default();
+        assert_eq!(
+            verdict(&spec, &fresh, &sorted, 100),
+            SortVerdict::NeverSorted
+        );
+
+        let state = SortState {
+            fields: Some(spec.to_string()),
+            rows: Some(1_000),
+            fragments: Some(sorted.len()),
+            digest: Some(digest(&data_files(&sorted))),
+        };
+
+        // The same files, whatever version the commit ended up as.
+        assert_eq!(verdict(&spec, &state, &sorted, 100), SortVerdict::UpToDate);
+
+        // Appended: the sorted fragments are still the first ones.
+        let appended_50 = frags(&[("a.lance", 500), ("b.lance", 500), ("c.lance", 50)]);
+        assert_eq!(
+            verdict(&spec, &state, &appended_50, 100),
+            SortVerdict::UpToDate
+        );
+        let appended_100 = frags(&[("a.lance", 500), ("b.lance", 500), ("c.lance", 100)]);
+        assert_eq!(
+            verdict(&spec, &state, &appended_100, 100),
+            SortVerdict::RowsAppended(100)
+        );
+
+        // Replaced: different files, and the row count is no defence — this is
+        // the case a row-count threshold lets through, whatever the delta.
+        let replaced = frags(&[("x.lance", 500), ("y.lance", 501)]);
+        assert_eq!(
+            verdict(&spec, &state, &replaced, 100),
+            SortVerdict::Rewritten
+        );
+        // Compacted into fewer files is a replacement too.
+        assert_eq!(
+            verdict(&spec, &state, &frags(&[("merged.lance", 1_000)]), 100),
+            SortVerdict::Rewritten
+        );
+
+        // Deleting rows rewrites no data file, so the table still reads as
+        // sorted — which it is.
+        let after_deletes = frags(&[("a.lance", 500), ("b.lance", 400)]);
+        assert_eq!(
+            verdict(&spec, &state, &after_deletes, 100),
+            SortVerdict::UpToDate
+        );
+
+        let other = SortSpec::parse("b desc").unwrap().unwrap();
+        assert_eq!(
+            verdict(&other, &state, &sorted, 100),
+            SortVerdict::OrderChanged
+        );
+    }
+
+    /// Deletions must not hide appended rows. Net growth against the recorded
+    /// row count would say this table shrank, and it would stay unsorted with
+    /// an unsorted tail for as long as the deletes kept pace with the appends.
+    #[test]
+    fn deletions_do_not_mask_appended_rows() {
+        let spec = SortSpec::parse("a asc").unwrap().unwrap();
+        let sorted = frags(&[("a.lance", 500), ("b.lance", 500)]);
+        let state = SortState {
+            fields: Some(spec.to_string()),
+            rows: Some(1_000),
+            fragments: Some(sorted.len()),
+            digest: Some(digest(&data_files(&sorted))),
+        };
+
+        // 800 rows deleted from the sorted fragments, 300 appended after them:
+        // the table is smaller than it was, and 300 rows of it are unsorted.
+        let mixed = frags(&[("a.lance", 100), ("b.lance", 100), ("c.lance", 300)]);
+        assert_eq!(
+            verdict(&spec, &state, &mixed, 100),
+            SortVerdict::RowsAppended(300)
+        );
+
+        // The threshold still damps a small append beside heavy deletion.
+        let small = frags(&[("a.lance", 100), ("b.lance", 100), ("c.lance", 20)]);
+        assert_eq!(verdict(&spec, &state, &small, 100), SortVerdict::UpToDate);
+
+        // And net growth is still detected where there are no deletions.
+        let grown = frags(&[("a.lance", 500), ("b.lance", 500), ("c.lance", 700)]);
+        assert_eq!(
+            verdict(&spec, &state, &grown, 100),
+            SortVerdict::RowsAppended(700)
+        );
+    }
+
+    // An appended fragment whose length the manifest does not record cannot be
+    // measured against the threshold, and a table that cannot be judged is one
+    // to sort rather than one to leave alone forever.
+    #[test]
+    fn an_uncountable_appended_fragment_is_stale() {
+        let spec = SortSpec::parse("a asc").unwrap().unwrap();
+        let sorted = frags(&[("a.lance", 500)]);
+        let state = SortState {
+            fields: Some(spec.to_string()),
+            rows: Some(500),
+            fragments: Some(1),
+            digest: Some(digest(&data_files(&sorted))),
+        };
+
+        let mut appended = sorted.clone();
+        appended.push(FragmentSummary {
+            files: vec!["b.lance".to_string()],
+            rows: None,
+        });
+        assert_eq!(
+            verdict(&spec, &state, &appended, 100),
+            SortVerdict::Rewritten
+        );
+        // The untouched table is still up to date: the files answer that
+        // without needing to count anything.
+        assert_eq!(verdict(&spec, &state, &sorted, 100), SortVerdict::UpToDate);
+    }
+
+    // Distinct columns whose names differ only in case are two columns in an
+    // Arrow schema, not one named twice.
+    #[test]
+    fn case_distinguishes_two_fields() {
+        let spec = SortSpec::parse("id asc, ID desc").unwrap().unwrap();
+        assert_eq!(spec.fields.len(), 2);
+        assert!(SortSpec::parse("id asc, id desc").is_err());
+    }
+
+    #[test]
+    fn the_marker_round_trips_through_a_config_map() {
+        let spec = SortSpec::parse("a asc").unwrap().unwrap();
+        let written = frags(&[("one.lance", 20), ("two.lance", 22)]);
+        let recorded = SortState::record(&spec, &written, 42);
+        let state = SortState::from_config(&recorded);
+        assert_eq!(state.rows, Some(42));
+        assert_eq!(state.fragments, Some(2));
+        assert_eq!(verdict(&spec, &state, &written, 1), SortVerdict::UpToDate);
+    }
+}