docs: ADR 0012, pool pre-ping is off and postgres_retry recovers dead connections

[lesnik512]

ADR 0012 records why pool_pre_ping is off (#54). The record also states the rule that pays for the decision: every query a retried use case runs must be inside its Transaction block.

Measurements behind the ADR

chat-app main at 31eee00 (db-retry 0.8.0). The app (granian, 1 worker), Postgres 17 and an httpx load generator ran as containers on one Docker network on Docker Desktop. Only DB_POOL_PRE_PING differed between runs. Each run had a 5 s warm-up and a 10 s window, the API was recreated at each switch, and on and off alternated. No run had errors.

Postgres's statement log for 500 GET /api/auth/me/ requests shows each request making two pool checkouts, one for the token-revocation check and one for the use case. With pre-ping on there were 2002 BEGINs and 1001 ROLLBACKs, against 1002 and 1 with it off. The ; ping itself isn't logged.

Local network, 3 rounds (req/s per round, median p50 and p99):

endpoint	clients	on req/s	off req/s	on p50	off p50	on p99	off p99
GET /api/auth/me/	1	543 / 612 / 553	748 / 759 / 707	1.80 ms	1.30 ms	2.27 ms	1.84 ms
GET /api/auth/me/	20	1026 / 977 / 1034	1210 / 1201 / 1219	18.89 ms	15.67 ms	39.50 ms	33.91 ms
GET /api/chats/{id}/messages/	1	543 / 474 / 550	583 / 580 / 580	1.80 ms	1.70 ms	2.50 ms	2.23 ms
GET /api/chats/{id}/messages/	20	727 / 616 / 725	805 / 801 / 812	27.10 ms	23.95 ms	50.59 ms	45.52 ms

tc qdisc add dev eth0 root netem delay 1ms on the db container, 2 rounds:

endpoint	clients	on req/s	off req/s	on p50	off p50	on p99	off p99
GET /api/auth/me/	1	24 / 24	40 / 41	41.96 ms	25.15 ms	51.96 ms	32.14 ms
GET /api/auth/me/	20	170 / 171	403 / 398	117.44 ms	50.25 ms	188.37 ms	82.83 ms
GET /api/chats/{id}/messages/	1	20 / 20	33 / 32	50.50 ms	30.33 ms	58.06 ms	42.82 ms
GET /api/chats/{id}/messages/	20	160 / 161	332 / 334	124.26 ms	60.02 ms	177.83 ms	89.71 ms