Обсуждение: Query Planner not taking advantage of HASH PARTITION

Benjamin Tingle <ben@tingle.org> writes:
> I've recently started taking advantage of the PARTITION BY HASH feature for
> my database system. It's a really great fit since my tables can get quite
> large (900M+ rows for some) and splitting them up into manageable chunks
> should let me upload to them without having to update an enormous index
> every time. What's more, since each partition has a write lock independent
> of the parent table, it should theoretically be possible to perform a
> parallelized insert operation, provided the data to be added is partitioned
> beforehand.

> What has been disappointing is that the query planner doesn't seem to
> recognize this potential.

That's because there isn't any.  The hash partitioning rule has
basically nothing to do with any plausible WHERE condition.  If you're
hoping to see partition pruning happen, you need to be using list or
range partitions, with operators compatible with your likely WHERE
conditions.

(I'm of the opinion that the hash partitioning option is more in the
category of a dangerous nuisance than a useful feature.  There are some
around here who will argue otherwise, but they're wrong for exactly the
reason that it's impossible to prune hash partitions.)

            regards, tom lane

Benjamin Tingle <ben@tingle.org> writes:
> Interesting. Why is it impossible to prune hash partitions? Maybe prune
> isn’t the best word, more so use to advantage. At the very least, it should
> be possible to utilize a parallel insert against a table partitioned by
> hash. (Partition query rows, then distribute these rows to parallel workers)

Your plan-shape complaint had nothing to do with insertions; it had
to do with joining the partitioned table to another table.  That
join can't be optimized.

            regards, tom lane

On 2022-Apr-14, Benjamin Tingle wrote:

> It doesn't help if I partition temp_data by textfield beforehand either
> (using the same scheme as the target table). It still opts to concatenate
> all of temp_data, hash it, then perform a sequential scan against the
> target partitions.

Does it still do that if you set
  SET enable_partitionwise_join TO on;
?  If the partition strategies are identical, that might get you a
better plan.  (Actually, in pg13 and upwards the strategies don't need
to be exactly identical, just "compatible".)

-- 
Álvaro Herrera         PostgreSQL Developer  —  https://www.EnterpriseDB.com/
"Thou shalt not follow the NULL pointer, for chaos and madness await
thee at its end." (2nd Commandment for C programmers)

On Wed, Apr 20, 2022 at 07:11:37PM -0700, Benjamin Tingle wrote:
> @ the first point about write locks
> I think I had/have a misconception about how inserts work in postgres. It's
> my understanding that postgres will never draft a parallel insert plan for
> any query (except maybe CREATE TABLE AS?)

It's correct that DML (INSERT/UPDATE/DELETE) currently is not run in parallel.
https://www.postgresql.org/docs/current/when-can-parallel-query-be-used.html

> because the process needs to acquire an exclusive access write lock to the
> table it is inserting on.

But this is incorrect - DML acquires a relation lock, but not a strong one.
Multiple processes can insert into a table at once (because the row-excl lock
level is not self-conflicting, to be technical).
https://www.postgresql.org/docs/current/explicit-locking.html

In fact, that's a design requirement.  It's understood that many people would
be unhappy if only one client were able to run UPDATEs at once, and that only a
toy system would acquire a strong lock for DML.

-- 
Justin