Re: [Proposal] Make the optimiser aware of partitions ordering

On Mon, Mar 20, 2017 at 8:33 PM, Ronan Dunklau <ronan(dot)dunklau(at)dalibo(dot)com> wrote:
> On lundi 20 mars 2017 15:52:03 CET Robert Haas wrote:
>> On Mon, Mar 20, 2017 at 6:31 AM, Ronan Dunklau <ronan(dot)dunklau(at)dalibo(dot)com>
> wrote:
>> > With range partitioning, we guarantee that each partition contains non-
>> > overlapping values. Since we know the range allowed for each partition, it
>> > is possible to sort them according to the partition key (as is done
>> > already for looking up partitions).
>> >
>> > Thus, we ca generate sorted Append plans instead of MergeAppend when
>> > sorting occurs on the partition key.
>>
>> Great idea. This is too late for v10 at this point, but please add it
>> to the next CommitFest so we don't forget about it.
>
> I know it is too late, and thought that it was too early to add it to the
> commitfest properly since so many design decisions should be discussed. Thanks
> for the feedback, I added it.

Thanks for working on this. I was also thinking about the same.

I will try to get back to review/work with this for v11. Mean time, I
am working on partition-wise joins [1]. In those patches, I have added
a structure called PartitionScheme, which represents how a relation is
partitioned. For base relations it's mostly copy of PartitionDesc and
PartitionKey, but then it bubbles up across join, with each
partitioned join getting relevant partitioning scheme. If you could
base your design such that is uses PartitionScheme, it could be used
for joins and probably when Jeevan's patch for partition-wise
aggregate [2] comes along, it can be used with grouping.

[1] https://www.postgresql.org/message-id/CAFjFpRcMWwepj-Do1otxQ-GApGPSZ1FmH7YQvQTwzQOGczq_sw%40mail.gmail.com
[2] http://www.mail-archive.com/pgsql-hackers(at)postgresql(dot)org/msg308861.html

--
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company