Re: Python/pytest test framework take two

On Wed, 28 Oct 2026 at 22:33:16 +0100, Jelte Fennema-Nio
<me(at)jeltef(dot)nl> wrote:
> --tb=short, only the last 30 server log
> lines, pg_ctl --silent, framework internals hidden from tracebacks,
> and a clearer error for unconsumed background query results.

Thanks, Jelte, for working on this. I'm happy to see this move to Python
happening and would like to help. Two follow-ups:

1. result(timeout=...) and exception(timeout=...) mark a future consumed
before waiting. A caught timeout can therefore allow connection reuse
while the query still runs. The first patch fixes this: eight baseline
failures become passes (20/20 total).

2. The bulk pg_visibility conversion (8a8f008f7596) misses
close_portal("") after begin / select txid_current(), retaining a snapshot
that psql releases. The second patch adds two passing comparison tests
for this documented behavior.

Apply order: the timeout patch, then the snapshot patch, to v2 (1d2adb67c551).
On macOS arm64, all 22 new cases, core regression and isolation pass.
The known Python 3.9 timestamp failures and optional-feature skips remain.
Reproduction steps, exact revisions and detailed results are in [1].

My Postgres hacking AI harness did the legwork here. I haven't yet
reviewed the findings or patches by hand, but I trust it enough to share
them for review. If anything is off, please tell me and I'll adjust or
stop this work.

[1] https://github.com/NikolayS/postgres/blob/d972a089cf7b640ee235489c71daac100d9f4bad/review/pytest-followup/README.md

Nik