議論： Long-running and idle-in-transaction sessions in PostgreSQL: what they block and how to bound them

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-16

翻訳がないため、原文を表示しています。 原文

Where the forgotten transaction comes from, and two settings around the ones named. Most drivers open a transaction implicitly: psycopg (2 and 3) runs with autocommit off by default, so the first statement issues `BEGIN` and the session stays `idle in transaction` until `commit()` or `rollback()`; JDBC defaults to autocommit on, so the same code path there holds nothing, which is why the symptom follows the language. PgBouncer has its own `idle_transaction_timeout` (off by default) that disconnects the client that is idle in a transaction, which surfaces the error at the client that caused it rather than at whichever client next receives the server connection, so on pooled setups it is the better place for the limit. PostgreSQL 14 added `idle_session_timeout` for sessions idle outside a transaction, which frees connections rather than the xmin horizon, and `vacuum_failsafe_age`, which lets a wraparound-driven vacuum skip index cleanup when the age gets critical; neither replaces the article's timeouts, but both belong on the same checklist.