> > This can work normally on range partitions. However, the constraint on hash > > partitions uses satisfies_hash_partition with the OID of the parent table, and > > the newly created constraint does not take effect. For example, in the following > > case, although there is a t_p1_a_check constraint on t_p1, it is still possible > > to perform an insert: > What I said here is wrong, the constraints on the hash partition will also take > effect. But this constraint depends on the oid of the parent partition.
We should definitely not have this constraint on hash-partition tables after the detach. However, I wonder if instead of adding it and later removing it as you propose, it wouldn't be better to just not add it in the first place. As a first step, I tried commenting out and found that no interesting test fails (only alter_table.sql fails but only because the constraint is not there when looking for it specifically.)
The current code does not explain *why* we have to add this constraint, and I had forgotten, so I went to look at the first patch submission in that thread [1] and saw this comment there:
+ /* + * Concurrent mode has to work harder; first we add a new constraint to the + * partition that matches the partition constraint. The reason for this is + * that the planner may have made optimizations that depend on the + * constraint. XXX Isn't it sufficient to invalidate the partition's + * relcache entry?
I'm trying to figure out whether it's possible that the planner would make optimizations based on the hashing function. Quite possibly it won't. If that's so, then we should just not make the constraint at all, which would make the fix even simpler. I also wonder, maybe that XXX comment (which I removed before committing the patch) is right and we don't actually need the constraint with _any_ partition strategy, not just hash.
The planner will do partition pruning based on partbound, if there is no the
new adding constraint when detaching concurrently, can we make sure that
