Dev-iL commented on code in PR #67800:
URL: https://github.com/apache/airflow/pull/67800#discussion_r3369000609
##########
airflow-core/docs/howto/set-up-database.rst:
##########
@@ -233,6 +233,13 @@ For more information regarding setup of the PostgreSQL
connection, see `PostgreS
See also :ref:`Helm Chart production guide <production-guide:pgbouncer>`
+ Some Airflow database routes use an async engine (the Execution API, for
example). The asyncpg
Review Comment:
Switched the docs to `statement_cache_size=0` +
`prepared_statement_cache_size=0`, both in `connect_args`. This is because the
dialect pops the latter before `connect()`, so no need to split it into the
URL. Verified on breeze/PG14: default leaves 6 `__asyncpg_stmt__` entries on
the backend, the recipe leaves 0. Defaulting to psycopg3 is tracked in #67801.
---
Regarding the translatability of the benchmark results to practice:
- The big gap in the benchmark is mostly an artifact of the tight localhost
loop: back-to-back queries w/o a network hop - so it isolates driver overhead.
- The benchmark mixes concurrency models: async drivers run as coroutines on
one event loop, sync ones via a thread pool, so the sync-vs-async rows aren't
directly comparable. asyncpg vs psycopg3-async (both async, dedicated
connection each) is the clean comparison, and the one that matters for choosing
the default.
- In real Airflow it largely won't translate: one query per request behind a
network/PgBouncer hop, where the round-trip dominates and the driver delta is a
small fraction.
Either way it's moot if we pick psycopg3 on pgbouncer-safety rather than
throughput.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
