0713 — Sustainable throughput keeps both producer and consumer queues bounded

Context

The requested maximum is simultaneous durable production and consumption with no growing backlog. Short producer-only or eventual-drain numbers cannot establish it. The reliability harness currently supplies explicit idempotency keys, which bypass automatic batching.

Decision

Use this Mac, 1 KB messages, one group, a minimal handler, automatic batching, and default failure-only delivery logging as the initial workload. Caller-key deduplication, extra groups, and full delivery logging are separate measurements. Require p99 end-to-end <= 1s initially.

Tune configurations before proposing library/SQL optimizations. Compare Docker and native Postgres under matched versions/resources. Keep fsync, synchronous_commit, full_page_writes, durable tables, and autovacuum on.

Start with 10–30s probes, then 1–2m candidates. Validate finalists with three 15m windows after warmup, extending until checkpoints, vacuum, and retention cleanup occur. Both producer and consumer queues must stay bounded during production. Report each group’s completions separately.

Budget 20 GiB for all benchmark storage and preserve 40 GiB host free. Measure growth before fixing retention or starting long runs. Cloud and hardware purchases are outside this scope.

Consequences

Measurement corrections precede tuning. Configurations and environment must accompany results. An eventual drain is a correctness check, not proof of sustained throughput. Recording and checker storage are part of the budget and their overhead must be measured.