#4287 serve: step leases seem never to be reaped when serve restarts more often than once an hour

Setup

• swamp 20261031.063736.0-sha.e459d219, running as swamp serve with --hot-reload and the built-in filesystem datastore.
• .swamp.yaml has garbageCollection.automatic: true and bookkeeping: 2h.
• About 2,100 worker dispatches an hour.
• Serve often restarts. Our release process restarts it when a commit changes a manifest's additionalFiles, which is our workaround for #4282. On 2026-11-02 this happened every 15 to 30 minutes for a few hours.

What we observed

• Between 09:30Z and 11:45Z, our 15-minute datastore watch ran on 9 different serve instances (instanceId from /ready).
• Over the same window, the file count under .swamp/data rose at every sample and never dropped: 26,326, 30,303, 33,939, 38,590, 41,926, 47,765, 51,051, 56,449, 65,777, 72,141.
• Most of that is data/swamp/step-lease. It held 6,886 files at 09:00Z, 37,115 at 11:00Z and 70,676 at 12:15Z.
• swamp run gc and swamp data gc, run through serve at 12:15Z, removed nothing.

What we concluded (we may be wrong)

From reading the source, we think the step leases are only reaped by serve's worker GC:

• WorkerGcService.start() schedules its first sweep intervalMs after start.
• serve.ts passes DEFAULT_WORKER_GC_INTERVAL_MS (1 hour), and we found no flag or environment variable for it.
• reapBookkeeping runs inside that sweep.
• libswamp/retention/collectors.ts notes that serve bookkeeping is reaped by the worker GC, not by retention.

If that is right, a serve that lives less than an hour never sweeps, and each restart starts the hour again. RetentionGcService uses a 5-minute initialDelayMs, so outputs and runs still get reaped across restarts. That would explain why only the bookkeeping grew.

We may have missed another path that reaps these records, or misread how often the sweep runs. We would be glad to be corrected.

What might help

Any of these would keep bookkeeping bounded when serve restarts often:

• the worker GC uses a short initial delay, like retention GC does;
• it remembers when it last swept;
• it gets an interval flag.

Related

• #4282: why our serve restarts.
• #4285: output growth from set_status.