Gregory Stark <stark@enterprisedb.com> writes:
> The problem testing this patch is that the window for a committed transaction
> to not be synced is quite narrow, especially for the regression tests. For
> testing purposes I wonder if there are ways we can widen this window. Some
> ideas, some wackier than others, are:
> . Raise the default wal_writer_delay to 5s or so -- also temporary until
> release
I ran 100+ cycles of the parallel regression tests with this setting,
and didn't see any failures.
> . Add an ifdef USE_ASSERT_CHECKING which randomly omits setting hint
> bits even when it could.
I think this is better done by code inspection, ie, look for places that
assume HEAP_XMIN/XMAX_COMMITTED is or can be set.
I made a pass over CVS HEAD and found some apparent trouble spots:
heapam.c lines 1843-1852 presume previous xmax can be hinted
immediately, ditto lines 2167-2176, ditto lines 2716-2725.
I think probably we should just remove those lines --- they are only
trying to save work for future tqual.c calls.
regards, tom lane
