Dear Vignesh,
> I was unable to reproduce the same test failure on the PG17 branch,
> even after running the test around 500 times. However, on the master
> branch, the failure consistently reproduces approximately once in
> every 50 runs. I also noticed that while the buildfarm has reported
> multiple failures for this test for the master branch, none of them
> appear to be on the PG17 branch. I'm not yet sure why this discrepancy
> exists.
I was also not able to reproduce as-is. After analyzing bit more, I found on
PG17, the workload cannot generate an FPI_FOR_HINT. The type of WAL record
has longer length than the page there was a possibility that the WAL record
could be flushed partially in HEAD. But in PG17 it could not happen so that
OVERWRITE_CONTRECORD won't be appeared.
I modified the test code like [1] and confirmed that the same stuck could happen
on PG17. It generates a long record which can go across the page and can be
flushed partially.
[1]:
```
--- a/src/test/recovery/t/046_checkpoint_logical_slot.pl
+++ b/src/test/recovery/t/046_checkpoint_logical_slot.pl
@@ -123,6 +123,10 @@ $node->safe_psql('postgres',
$node->safe_psql('postgres',
q{select injection_points_wakeup('checkpoint-before-old-wal-removal')});
+# Generate a long WAL record
+$node->safe_psql('postgres',
+ q{select pg_logical_emit_message(false, '', repeat('123456789',
1000))});
```
Best regards,
Hayato Kuroda
FUJITSU LIMITED
