Andres Freund <and...@anarazel.de> writes:
> On 2022-03-22 16:55:49 -0400, Tom Lane wrote:
>> BTW, I've had a bee in my bonnet for a long time about whether some of
>> nbtree's scan setup work could be done once during planning, rather than
>> over again during each indexscan start.
> It does show up in simple-index-lookup heavy workloads. Not as a major thing,
> but it's there. And it's just architecturally displeasing :)
> Are you thinking of just moving the setup stuff in nbtree (presumably parts of
> _bt_first() / _bt_preprocess_keys()) or also stuff in
> ExecIndexBuildScanKeys()?
Didn't really have specifics in mind. The key stumbling block is
that some (not all) of the work depends on knowing the specific
values of the indexqual comparison keys, so while you could do
that work in advance for constant keys, you'd still have to be
prepared to do work at scan start for non-constant keys. I don't
have a clear idea about how to factorize that effectively.
A couple of other random ideas in this space:
* I suspect that a lot of this work overlaps with the efforts that
btcostestimate makes along the way to getting a cost estimate.
So it's interesting to wonder whether we could refactor so that
btcostestimate is integrated with this hypothetical plan-time key
preprocessing and doesn't duplicate work.
* I think that we run through most or all of that preprocessing
logic even for internal catalog accesses, where we know darn well
how the keys are set up. We ought to think harder about how we
could short-circuit pointless work in those code paths.
I don't think any of this is an essential prerequisite to getting
something done for loose index scans, which ISTM ought to be the first
point of attack for v16. Loose index scans per se shouldn't add much
to the key preprocessing costs. But these ideas likely would be
useful to look into before anyone starts on the more complicated
preprocessing that would be needed for the ideas in the MDAM paper.
regards, tom lane
