On Jan19, 2014, at 20:00 , David Rowley <dgrowleyml@gmail.com> wrote:
> I've applied that patch again and put in the sort operators.
I've push a new version to https://github.com/fgp/postgres/tree/invtrans
This branch includes the following changes
* A bunch of missing declaration for *_inv functions
* An assert that the frame end doesn't move backwards - I realized thatit is after all easy to do that, if it's done
afterthe loop which addsthe new values, not before.
* EXPLAIN VERBOSE ANALYZE now shows the max. number of forward aggregatetransitions per row and aggregate. It's a bit
imprecise,because it doesn'ttrack the count per aggregate, but it's still a good metric for how wellthe inverse
transitionfunctions work. If the number is close to one, youknow that very few rescans are happening.
* I've also renamed INVFUNC to INVSFUNC. That's a pretty invasive change, andit's the last commit, so if you object to
that,then you can merge up toeafa72330f23f7c970019156fcc26b18dd55be27 instead
ofde3d9148be9732c4870b76af96c309eaf1d613d7.
A few more things I noticed, all minor stuff
* do_numeric_discard()'s inverseTransValid flag is unnecessary. First, if theinverse transition function returns NULL
once,we never call it again, so theflag won't have any practical effect. And second, assume we ever called theforward
transitionfunction after the inverse fail, and then retried the inverse.In the case of do_numeric_discard(), that
actually*could* allow the inverseto suddenly succeed - if the call to the forward function increased the dscalebeyond
thatof the element we tried to remove, removal would suddenly bepossible again. We never do that, of course, and it
seemsunlikely we everwill. But it's still weird to have code which serves no other purpose than topessimize a case
whichwould otherwise just work fine.
* The state == NULL checks in all the strict inverse transition functions areredundant.
I haven't taken a close look at the documentation yet, I hope to be able to
do that tomorrow. Otherwise, things look good as far as I'm concerned.
best regards,
Florian Pflug