RST-20 — Soak, maintenance and cost efficiency (30-minute pilot)

2026-10-10 correction: historical table size included indexes and index cost was counted again. Table/index cost slopes must be remeasured. Collectors now use pg_table_size and PG18 num_done and propagate errors. Import-blocked sampling, short duration and missing 24/48h maintenance evidence keep RST-20 INCOMPLETE; the old pilot is historical diagnostic evidence only.

Status: PILOT-MEASURED — isolated disposable PostGIS only; no migration, no product change, no new dependency. Date: 2026-10-10. Scope: RUST_STATION_BENCHMARK_PLAN.md RST-20 pilot only; the 24/48h campaigns stay separately scheduled (OWED) and this pilot is never labelled soak success. Language: English document; user communication in Portuguese.

B-BR/BUC: B-BR-RST-P02 (pinned 30 min, 20 rps, 2 min import cadence, hot-city weights, sampler cadence), B-BR-RST-P04 (fresh disposable database, dropped after; durability never weakened), B-BR-RST-P06 (per-window tails, slopes by least squares, no extrapolation). Serves BUC-RST-P01 (sustained import costs) and BUC-RST-P05 (trend baselines).

1. What was added

• read_soak_bench_test.go (integration): 20 rps reads with hot-city rotation (50% hot dense, 30% rotating sparse sweep, 20% nearby), one disjoint 20k-row edition every 2 minutes (atomic claim, oracle counts asserted per import), VACUUM ANALYZE at t=10/20, simulated retention (drop 3 oldest editions) at t=15/25, 30 s sampler (read window p95/errors, import lag, pool, WAL, sizes, dead tuples, backlog, checkpoints), slope + headroom + budget verdict at end. RST20_DURATION_MIN overrides duration for smoke checks and separately scheduled campaigns only.
• 16 disjoint 20k editions to scratch (offsets 1M–2.5M, zero cross-edition overlap); 9 consumed, 7 spare.

2. Pilot measurements (30 min + seed, 100k census, pool 10)

33 samples, 9 imports error-free, 6 editions retained, maintenance 1.5 s total, worst window p95 71.0 ms, 0 read errors, backlog slope 0.000 (bounded by worker count, never growing), pool headroom min 0.90, budget margin min 29.0 ms: verdict within budgets. End state: WAL 292 MB, tables 140.6 MB, indexes 61.6 MB, dead tuples 29,552, checkpoints unverified (counter reads ignored errors — instrumentation gap, not a zero claim).

Slopes per minute: WAL +9.28 MiB, tables +3.75 MiB (net of retention deleting ~147k rows — growth dominates), dead tuples +944, backlog +0.000. Retention by row-DELETE cost 1,086 s total for ~150k rows with no observed auto-vacuum relief: the strategy is expensive and dead-tuple heavy, owed a dedicated task (batched deletes or partition drops). Importer cadence slipped (9 vs 15 planned: blocking imports stretch the 2 min cadence) and sampler points landing inside imports are skipped — both reported, not smoothed.

Costs (explicit denominators, shared noisy host): WAL per staged row ~1.3 KB (292 MB over seed + 9×24,552 staged + maintenance + retention — traceable total, not purified); import wall ~0.5 s/1k rows single-threaded (RST-14 20k baseline, same shape); prep CPU 8.7 s/1M rows parse plus 2.2 s/1M emit (RST-13, release, i7-13620H). No currency math: no provider prices supplied, and free shared CPU is never costed. Workstation-vs-constrained comparison reuses the RST-18 VPS staging envelope (20 rps on the 1-CPU pod); re-running the pilot there needs a scheduled window, owed with the 24/48h campaigns.

3. Limits and next task

• 24/48h campaigns: OWED (provisioned isolated resources + CI_PLAN scheduling). Checkpoint/autovacuum counters need verified instrumentation before the long campaigns.
• Reproduce: generate 20k editions to /tmp/rst20, then go test -tags=integration -run NONE -bench BenchmarkSoakPilot -benchtime 1x (~35 min).
• Next (separately authorized): RST-21 decision, report and regression tiers over the complete matrix; environment prep (bundle index + collection checker) ships with this task.