On 2018/05/18 6:14, David Rowley wrote:
> On 18 May 2018 at 02:13, Tom Lane <tgl@sss.pgh.pa.us> wrote:
>> Maybe what you need is a redesign. This convention seems impossibly
>> confusing and hence error-prone. What about using a separate bool to
>> indicate which list the index refers to?
>
> While I agree that the coding is a bit unusual, I think it's also good
> that we can get away without allocating yet another array nparts in
> size. ExecSetupPartitionTupleRouting is already a huge bottleneck with
> single-row INSERT into a partitioned table with a large number of
> partitions. Allocating yet another array nparts in size will just slow
> it down further.
I recall having considered the idea of adding an array of bools, but went
with the negative-indexes-for-partitioned-tables idea anyway, which I
remember was suggested by Robert back then [1]. I admit it's a bit
confusing, but it's nice not have one more array allocation in that path
as you say.
> I have patches locally that I'll be submitting during the v12 cycle to
> improve on this. Among other things, the patches go to lengths to not
> allocate these arrays when we don't have to.
That would be nice.
Thanks,
Amit
[1]
https://www.postgresql.org/message-id/CA%2BTgmobF2r%3Df-crrE-k7WM8iFpBKLz3dtBtEc%3DKmkudYViYcyQ%40mail.gmail.com
