On Sun, Sep 29, 2024 at 12:22 AM Alena Rybakina <a.rybak...@postgrespro.ru> wrote: > Hi! Thank you for your interesting for this patch! > > I took a very brief look at this and was wondering if it was worth > having a way to make the per-table vacuum statistics opt-in (like a > table storage parameter) in order to decrease the shared memory > footprint of storing the stats. > > I'm not sure how users can select tables that enable vacuum statistics > as I think they basically want to have statistics for all tables, but > I see your point. Since the size of PgStat_TableCounts approximately > tripled by this patch (112 bytes to 320 bytes), it might be worth > considering ways to reduce the number of entries or reducing the size > of vacuum statistics. > > The main purpose of these statistics is to see abnormal behavior of vacuum in > relation to a table or the database as a whole. > > For example, there may be a situation where vacuum has started to run more > often and spends a lot of resources on processing a certain index, but the > size of the index does not change significantly. Moreover, the table in which > this index is located can be much smaller in size. This may be because the > index is bloated and needs to be reindexed. > > This is exactly what vacuum statistics can show - we will see that compared > to other objects, vacuum processed more blocks and spent more time on this > index. > > Perhaps the vacuum parameters for the index should be set more aggressively > to avoid this in the future. > > I suppose that if we turn off statistics collection for a certain object, we > can miss it. In addition, the user may not enable the parameter for the > object in time, because he will forget about it.
I agree with this point. Additionally, in order to benefit from gatherting vacuum statistics only for some relations in terms of space, we need to handle variable-size stat entries. That would greatly increase the complexity. > As for the second option, now I cannot say which statistics can be removed, > to be honest. So far, they all seem necessary. Yes, but as Masahiko-san pointed out, PgStat_TableCounts is almost tripled in space. That a huge change from having no statistics on vacuum to have it in much more detail than everything else we currently have. I think the feasible way might be to introduce some most demanded statistics first then see how it goes. ------ Regards, Alexander Korotkov Supabase
