[jthingelstad]

Follow-up to #622, for #630.

#622 raised the door pools' idle timeout from 30 s to 240 s. At 01:33 CDT the same morning RDS reported the db.t4g.micro "critically low on memory" and lowered shared_buffers from 23,081 to 11,295 pages for the next restart (#630). DatabaseConnections went from 1 to 3 to about 5.

This weighs what 240 s bought against what it held, and settles on 120 s.

Method. A week of real requests, replayed per environment through pg-pool's own rules (LIFO reuse, the idle timer, min, the 300 s lifetime and each door's max). The requests were 2,833 web-api reads that took a lease, 2,539 MCP tool calls and the collector's leases. The replay matches what was observed:

• At 30 s it gives web-api 61.6% fresh connections, against 61.8% observed.
• After the deploy it gives 3.73 backends, against 3.86 in Enhanced Monitoring.

Benefit and cost:

idle	week fresh (web-api + MCP)	share of 240 s's saving	extra backends, week avg	quiet half-day avg	under acceptance load
30 s	2,665 of 5,372	0%	0	1.21	2.65
60 s	2,375	51%	+0.12	1.32	2.93
120 s	2,195	82%	+0.34	1.54	3.31
240 s	2,091	100%	+0.72	1.93	3.82

• What it buys. Web-api's fresh share falls from 62% to about 45%, and connect_ms p50 falls from 33 ms to 0. MCP barely moves. The collector was already warm.
• What it costs. About 6 to 7 MB per idle backend (free plus cached fell 16 MB for 2.4 more backends).
• Swap. It follows acceptance's heavy reads more than the timeout.
• Ruled out. min: 1 with a short idle holds nearly as much as 240 s, because the kept connection lives its full 300 s.
• Pool max stays. Bursts are what need it.

[engineering] default changed: DECISIONS' door-pool line now says 120 s, with the evidence in docs/notes/2026-10-21-pool-idle-120.md. No contract moves.

npm run verify green. e2e-needed.sh prints false.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	🔄 Running since 2026-10-21T08:35:14.626619Z	5e78ac0	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[jthingelstad]

Deployed to production 2026-10-21T07:11:58Z (2026-10-21 02:11:58 CDT).

• Commit: 5e78ac0e (platform lane); this PR's: 5e78ac0e
• MCP contract: 11.13.3
• JSON API: 3.2.1
• Acceptance: passed (--acceptance=game)

[jthingelstad]

@codex review

[jthingelstad]

Read back live (02:12 to 04:07 CDT): app backends averaged 2.20 (Enhanced Monitoring) against 3.57 at 240 s and 1.36 at 30 s the same hours yesterday. DatabaseConnections averaged 2.3 against 3.5. web-api connect_ms p50 stayed at 0 (23% fresh, against 57% at 30 s yesterday). Full table on #630. The deploy ran --acceptance=game rather than the whole suite: the change moves no tool output, and the database's EBS byte balance was at 68% and falling under other sessions' whole-suite runs.