Re: Skip partition tuple routing with constant partition key

Hou-san,

On Tue, Jun 1, 2021 at 5:43 PM houzj.fnst@fujitsu.com
<houzj.fnst@fujitsu.com> wrote:
> From: Amit Langote <amitlangote09@gmail.com>
> > > > Where are you thinking to cache the partidx list?  Inside
> > > > PartitionDesc or some executor struct?
> > >
> > > I was thinking cache the partidx list in PartitionDescData which is in
> > > relcache, if possible, we can use the cached partition between statements.
> >
> > Ah, okay.  I thought you were talking about a different idea.
> > How and where would you determine that a cached partidx value is indeed the correct one for
> > a given row?
> > Anyway, do you want to try writing a patch to see how it might work?
>
> Yeah, the different idea here is to see if it is possible to share the cached
> partition info between statements efficiently.
>
> But, after some research, I found something not as expected:

Thanks for investigating this.

> Currently, we tried to use ExecPartitionCheck to check the if the cached
> partition is the correct one. And if we want to share the cached partition
> between statements, we need to Invoke ExecPartitionCheck for single-row INSERT,
> but the first time ExecPartitionCheck call will need to build expression state
> tree for the partition. From some simple performance tests, the cost to build
> the state tree could be more than the cached partition saved which could bring
> performance degradation.

Yeah, using the executor in the lower layer will defeat the whole
point of caching in that layer.

> So, If we want to share the cached partition between statements, we seems cannot
> use ExecPartitionCheck. Instead, I tried directly invoke the partition support
> function(partsupfunc) to check If the cached info is correct. In this approach I
> tried cache the *bound offset* in PartitionDescData, and we can use the bound offset
> to get the bound datum from PartitionBoundInfoData and invoke the partsupfunc
> to do the CHECK.
>
> Attach a POC patch about it. Just to share an idea about sharing cached partition info
> between statements.

I have not looked at your patch yet, but yeah that's what I would
imagine doing it.

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