Re: Skip partition tuple routing with constant partition key

Hou-san,

On Mon, May 24, 2021 at 10:31 AM houzj(dot)fnst(at)fujitsu(dot)com
<houzj(dot)fnst(at)fujitsu(dot)com> wrote:
> From: Amit Langote <amitlangote09(at)gmail(dot)com>
> Sent: Thursday, May 20, 2021 8:23 PM
> > This one seems bit tough. ExecPartitionCheck() uses the generic expression
> > evaluation machinery like a black box, which means execPartition.c can't really
> > tweal/control the time spent evaluating partition constraints. Given that, we
> > may have to disable the caching when key->partexprs != NIL, unless we can
> > reasonably do what you are suggesting.[]
>
> I did some research on the CHECK expression that ExecPartitionCheck() execute.

Thanks for looking into this and writing the patch. Your idea does
sound promising.

> Currently for a normal RANGE partition key it will first generate a CHECK expression
> like : [Keyexpression IS NOT NULL AND Keyexpression > lowboud AND Keyexpression < lowboud].
> In this case, Keyexpression will be re-executed which will bring some overhead.
>
> Instead, I think we can try to do the following step:
> 1)extract the Keyexpression from the CHECK expression
> 2)evaluate the key expression in advance
> 3)pass the result of key expression to do the partition CHECK.
> In this way ,we only execute the key expression once which looks more efficient.

I would have preferred this not to touch anything but
ExecPartitionCheck(), at least in the first version. Especially,
seeing that your patch touches partbounds.c makes me a bit nervous,
because the logic there is pretty complicated to begin with.

How about we start with something like the attached? It's the same
idea AFAICS, but implemented with a smaller footprint. We can
consider teaching relcache about this as the next step, if at all. I
haven't measured the performance, but maybe it's not as fast as yours,
so will need some fine-tuning. Can you please give it a read?

--
Amit Langote
EDB: http://www.enterprisedb.com