Hi,
On 1/13/23 10:17 AM, Drouvot, Bertrand wrote:
> Hi,
>
> On 1/11/23 9:27 PM, Andres Freund wrote:
>> Hi,
>>
>> On 2023-01-06 10:52:06 +0100, Drouvot, Bertrand wrote:
>>
>> The problem I have with that is that I saw a lot of flakiness in the tests due
>> to the race condition. So introducing them in that order just doesn't make a
>> whole lot of sense to me.
>
> You are right it does not make sense to introduce fixing the race condition after the TAP tests
> and after introducing the decoding logic. I'll reorder the sub-patches.
>
V40 attached is changing the sub-patches ordering.
>> The suggested path in earlier versions to avoid doing so was to make sure that
>> we pass down the Relation for the table into the necessary functions. Did you
>> explore that any further?
>
> So, for gistXLogPageReuse() and _bt_delitems_delete() this is "easy" to pass the Heap Relation.
> This is what was done in earlier versions of this patch series.
>
> But we would need to define a way to propagate the Heap Relation for those 2 functions:
>
> _bt_log_reuse_page()
> vacuumRedirectAndPlaceholder()
>
V40 is getting rid of the new indisusercatalog field in pg_index and is passing the
heap relation all the way down to _bt_log_reuse_page() and vacuumRedirectAndPlaceholder() instead
(and obviously to gistXLogPageReuse() and _bt_delitems_delete() too).
Remarks:
1) V40 adds the heap relation in the IndexVacuumInfo and ParallelVacuumState structs. It is used
for the _bt_log_reuse_page() and vacuumRedirectAndPlaceholder() cases where I did not find any place
where to get the heap relation from in the existing code path.
2) V40 adds a "real" heap relation to all the _bt_getbuf() calls. Another option could have been
to add it only for the code paths leading to _bt_log_reuse_page() but I thought it is cleaner to
do it for all of them.
> I will give it another look, also because I just realized that it could be beneficial
> for vacuumRedirectAndPlaceholder() too, as per this comment:
>
> "
> /* XXX: providing heap relation would allow more pruning */
> vistest = GlobalVisTestFor(NULL);
> "
Now, we could also pass the heap relation to GlobalVisTestFor() in vacuumRedirectAndPlaceholder().
Could be done in or independently of this patch series once committed (it's not part of V40).
Regards,
--
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com
