On 15/12/2025 10:00, Alexander Lakhin wrote:
> Hello Heikki,
>
> 09.12.2025 14:00, Heikki Linnakangas wrote:
>> Committed with that and some other minor cleanups. Thanks everyone!
>> This patch has been brewing for a while :-).
>
> I've spotted a couple of failures of new test 007_multixact_conversion at
> buildfarm:
> ...
>
> Thus, it looks like these animals can hit 180 seconds timeout with some
> external factors (concurrent load?) that make them run 2-3 times slower...
Hmm, so it's hitting the timeout while running the multixid generation
workload part of the test. The workload keeps 20 parallel connections
open, using them in a round-robin fashion, and apparently the timeout
spans the whole duration of each connection rather than the individual
queries run on them.
On my laptop with jit_above_cost=0 and jit_optimize_above_cost=1000,
like on those buildfarm animals, the multixid generation workload runs
in under 10 s. I suppose it could be 20x slower on a slow, busy system.
So the straightforward fix is to bump up the timeout.
I'm tempted to force "jit=off" here, as the multixid generation is not
really the thing that's being tested here. It's just used to generate
multixids in the cluster before the upgrade. Then again, the point of
forcing JIT in these buildfarm members is to test JITting as part of
everything else that wasn't originally written as a JIT test.
- Heikki
