wire: revive the connection between ops, so a dead pg is not forever #108

Merged

PerishFire merged 1 commit from wire-revive into main

2026-07-21 14:17:32 +00:00

PerishFire commented

2026-07-21 14:15:54 +00:00

Owner

No description provided.

PerishFire added 1 commit

2026-07-21 14:15:55 +00:00

wire: revive the connection between ops, so a dead pg is not forever

guard / guard (pull_request) Successful in 1m33s

Details

208098d5a9

A restarted postgres left the engine holding a corpse. adapt/pg.rs owns a
single PgConnection and never re-established it, so every op after the
restart failed and stayed failing — on the cluster this read as a login
that returned 401 forever, and only an api restart cleared it.

Wire grows a defaulted revive(), called from Core::seize() at the same
entry that rolls back a dirty seat. Reviving only BETWEEN ops means no
statement is re-issued and no write is replayed; the op that meets the
dead connection still fails, correctly. sqlite takes the default no-op.

Postgres tracks whether its last call was sound and pings only when it
was not: a ping on every op cost +0.2ms per read (500 reads 232ms ->
339ms, +43%) to save one failed request per outage, which is a bad
trade. Healthy paths now pay nothing (measured 208ms).

tests/pulse/pg::revived terminates the engine's own backend from a
second connection and pins the contract: one op fails, the next runs
revived. It fails without seize()'s revive call.

Law: txn.md X-9 plus a must-not; vocabulary.md registers revive.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

PerishFire merged commit eb24c8907b into main

2026-07-21 14:17:32 +00:00

PerishFire deleted branch wire-revive

2026-07-21 14:17:32 +00:00

PerishFire referenced this pull request from a commit

2026-07-21 14:17:34 +00:00

wire: revive the connection between ops, so a dead pg is not forever (#108)

PerishFire referenced this pull request from a commit

2026-07-21 15:21:57 +00:00

release: keel family 0.4.1