Re: Improve search for missing parent downlinks in amcheck

On Sun, Apr 28, 2019 at 4:36 AM Peter Geoghegan <pg@bowt.ie> wrote:
> On Sat, Apr 27, 2019 at 5:13 PM Alexander Korotkov
> <a.korotkov@postgrespro.ru> wrote:
> > Yes, increasing of Bloom filter size also helps.  But my intention was
> > to make non-lossy check here.
>
> Why is that your intention? Do you want to do this as a feature for
> Postgres 13, or do you want to treat this as a bug that we need to
> backpatch a fix for?

I think this definitely not bug fix.  Bloom filter was designed to be
lossy, no way blaming it for that :)

> Can we avoid the problem you saw with the Bloom filter approach by
> using the real size of the index (i.e.
> smgrnblocks()/RelationGetNumberOfBlocks()) to size the Bloom filter,
> and/or by rethinking the work_mem cap? Maybe we can have a WARNING
> that advertises that work_mem is probably too low?
>
> The  state->downlinkfilter Bloom filter should be small in almost all
> cases, so I still don't fully understand your concern. With a 100GB
> index, we'll have ~13 million blocks. We only need a Bloom filter that
> is ~250MB to have less than a 1% chance of missing inconsistencies
> even with such a large index. I admit that its unfriendly that users
> are not warned about the shortage currently, but that is something we
> can probably find a simple (backpatchable) fix for.

Sounds reasonable.  I'll think about proposing backpatch of something like this.

------
Alexander Korotkov
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company