On Tue, May 18, 2021 at 01:27:48PM +1200, David Rowley wrote:
> A slightly different optimization that I have considered and even
> written patches before was to have ExecFindPartition() cache the last
> routed to partition and have it check if the new row can go into that
> one on the next call. I imagined there might be a use case for
> speeding that up for RANGE partitioned tables since it seems fairly
> likely that most use cases, at least for time series ranges will
> always hit the same partition most of the time. Since RANGE requires
> a binary search there might be some savings there. I imagine that
> optimisation would never be useful for HASH partitioning since it
> seems most likely that we'll be routing to a different partition each
> time and wouldn't save much since routing to hash partitions are
> cheaper than other types. LIST partitioning I'm not so sure about. It
> seems much less likely than RANGE to hit the same partition twice in a
> row.
It depends a lot on the schema used and the load pattern, but I'd like
to think that a similar argument can be made in favor of LIST
partitioning here.
--
Michael