Benchmark Postgres sequence gap strategies #8

New issue

Closed

opened 2026-02-19 21:28:02 +00:00 by ash · 3 comments

ash commented

2026-02-19 21:28:02 +00:00

Owner

What

Benchmark all 4 approaches to the sequence gap problem with real Postgres on VPS.

Strategies

Single writer (baseline — no gaps)
Published sequence table (extra write, trivial reads)
pg_snapshot / xmin visibility (correct but unknown perf)
Gap detection with timeout (generic fallback)

Method

1M events in table
Concurrent writers (10 goroutines)
Measure: read latency, write overhead, gap detection accuracy
Test with held-open transactions
p50/p95/p99 latencies

VPS: 143.14.50.71

## What Benchmark all 4 approaches to the sequence gap problem with real Postgres on VPS. ## Strategies 1. Single writer (baseline — no gaps) 2. Published sequence table (extra write, trivial reads) 3. pg_snapshot / xmin visibility (correct but unknown perf) 4. Gap detection with timeout (generic fallback) ## Method - 1M events in table - Concurrent writers (10 goroutines) - Measure: read latency, write overhead, gap detection accuracy - Test with held-open transactions - p50/p95/p99 latencies ## VPS: 143.14.50.71

ash added the

performance

investigation

labels

2026-02-19 21:28:02 +00:00

ash commented

2026-02-20 08:06:35 +00:00

Author

Owner

Deferred to post-v1. Deferred: single writer avoids gaps entirely. Benchmark when we need multi-writer.

**Deferred to post-v1.** Deferred: single writer avoids gaps entirely. Benchmark when we need multi-writer.

ash referenced this issue from a commit

2026-02-20 08:16:42 +00:00

feat(pgstore): benchmark sequence gap strategies and concurrent writers

ash closed this issue

2026-02-20 08:16:42 +00:00

ash commented

2026-02-20 08:16:57 +00:00

Author

Owner

Benchmarked on real Postgres via VPS: single writer 122 evt/s, concurrent writers scale to ~317 evt/s with 8+ writers, global read 1K events in 6-8ms, advisory lock 3.3ms/op. Gap detection confirmed: all sequences contiguous after concurrent writes.