Re: Avoid overhead open-close indexes (catalog updates)

Em qui., 10 de nov. de 2022 às 05:16, Michael Paquier <michael(at)paquier(dot)xyz>
escreveu:

> On Thu, Sep 01, 2022 at 08:42:15AM -0300, Ranier Vilela wrote:
> > Let's wait for the patch to be accepted and committed, so we can try to
> > change it.
>
> FWIW, I think that this switch is a good idea for cases where we
> potentially update a bunch of tuples, especially based on what
> CatalogTupleInsert() tells in its top comment.

That's the idea.

> Each code path updated
> here needs a performance check to see if that's noticeable enough, but
> I can get behind the one of CopyStatistics(), at least.
>
For CopyStatistics() have performance checks.

>
> EnumValuesCreate() would matter less as this would require a large set
> of values in an enum, but perhaps ORMs would care and that should be
> measurable.

Have a list_length call, for a number of vals.
For 2 or more vals, it is already worth it, since
CatalogOpenIndexes/CatalogCloseIndexes will be called for each val.

> update_attstats() should lead to a measurable difference
> with a relation that has a bunch of attributes with few tuples.
>
Same here.
For 2 or more attributes, it is already worth it, since
CatalogOpenIndexes/CatalogCloseIndexes will be called for each.

DefineTSConfiguration() is less of an issue, still fine to change.
>
Ok.

AddRoleMems() should be equally measurable with a large DDL. As a
> whole, this looks pretty sane to me and a good idea to move on with.
>
One filter, only.

For all these functions, the only case that would possibly have no effect
would be in the case of changing a single tuple, in which case there would
be only one call CatalogOpenIndexes/CatalogCloseIndexes for both paths.

> I still need to check properly the code paths changed here, of
> course..
>
At least, the patch still applies properly.

regards,
Ranier Vilela