On 09.01.2020 19:48, Tomas Vondra wrote:
The most complex and challenged task is to support GTT for all kind
of indexes. Unfortunately I can not proposed some good universal
solution for it.
Just patching all existed indexes implementation seems to be the only
choice.
I haven't looked at the indexing issue closely, but IMO we need to
ensure that every session sees/uses only indexes on GTT that were
defined before the seesion started using the table.
Why? It contradicts with behavior of normal tables.
Assume that you have active clients and at some point of time DBA
recognizes that them are spending to much time in scanning some GTT.
It cab create index for this GTT but if existed client will not be able
to use this index, then we need somehow make this clients to restart
their sessions?
In my patch I have implemented building indexes for GTT on demand: if
accessed index on GTT is not yet initialized, then it is filled with
local data.
Can't we track which indexes a particular session sees, somehow?
Statistic is another important case.
But once again I do not completely understand why we want to address
all this issues with statistic in first version of the patch?
I think the question is which "issues with statistic" you mean. I'm sure
we can ignore some of them, e.g. the one with parallel workers not
having any stats (assuming we consider functions using GTT to be
parallel restricted).
If we do not use shared buffers for GTT then parallel processing of GTT
is not possible at all, so there is no problem with statistic for
parallel workers.
I think someone pointed out pushing stuff directly into the cache is
rather problematic, but I don't recall the details.
I have not encountered any problems, so if you can point me on what is
wrong with this approach, I will think about alternative solution.
--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
