Preserve a statement cache size of 0

[ckarnell]

AsyncPostgresConfiguration.__init__ resolves both cache-size arguments with or:

self.statement_cache_size = statement_cache_size or <setting>

0 is falsy, so it gets replaced by the setting, which defaults to None. The if ... is not None further down then skips the key entirely and asyncpg keeps its cache at the default.

Your own settings say 0 is the value that matters. statement_cache_size is documented as "Setting this to 0 is required when using PgBouncer in transaction mode", and prepared_statement_cache_size as disabling the cache "when set to 0". The or discards exactly the value both descriptions name as the reason to reach for the setting.

Switched to is not None at both sites. Two lines.

The scope is narrow, and saying so up front seems better than letting it read bigger than it is: passing 0 through settings already worked, because the constructor argument is None on that path, and prefect itself only ever builds this class from a connection URL. So the person affected is someone wiring up AsyncPostgresConfiguration themselves to sit behind PgBouncer, which is who the docstring is talking to.

Verified with no database and no network, just the constructor. On main, statement_cache_size=0 comes back as None while statement_cache_size=100 comes back as 100, and the same split happens for the prepared cache. The same call keeps the non-zero value and drops the zero.

Reverting the two lines with the new tests in place fails them both, and restoring passes all four. tests/server/database is 97 passed and 7 skipped with the change, against 93 before it, no regressions either way.

No PgBouncer here. The claim that 0 reaches asyncpg is read off connect_args, not observed on a live connection.

[codspeed-hq]

Merging this PR will not alter performance

✅ 8 untouched benchmarks
⏩ 1 skipped benchmark¹

---

_{Comparing ckarnell:fix/statement-cache-size-zero (ff1dd63) with main (fc2add9)}

Footnotes

1. 1 benchmark was skipped, so the baseline result was used instead. If it was deleted from the codebase, click here and archive it to remove it from the performance reports. ↩

[desertaxle]

Thanks @ckarnell! Looks like there's one small formatting issue, but once that's fixed, this should be good to merge!

[ckarnell]

Fixed, thanks. It was a trailing blank line at the end of the test file, which ruff-format strips at the hook pin.

Worth flagging since it caught me out: pyproject.toml pins ruff==0.16.0 but the pre-commit hook is v0.15.19, and the hook is what CI runs. I checked against 0.15.19.