On Wed, Jun 17, 2020 at 10:22:28PM +0800, 李杰(慎追) wrote:
> However, I found a problem. If there are many partitions,
> we may need to handle too many missing index entries when
> validate_index(). Especially for the first partition, the time may
> have been long and many entries are missing. In this case, why
> don't we put the second and third phase together into a transaction
> for each partition?
Not sure I am following. In the case of REINDEX, it seems to me that
the calls to validate_index() and index_concurrently_build() can
happen in a separate transaction for each index, as long as all the
calls to index_concurrently_swap() are grouped together in the same
transaction to make sure that index partition trees are switched
consistently when all entries are swapped from an invalid state to a
valid state, because the swapping phase is also when we attach a fresh
index to a partition tree. See also index_concurrently_create_copy()
where we don't set parentIndexRelid for the lower call to
index_create(). It would be good of course to check that when
swapping we have the code to handle that for a lot of indexes at
once.
--
Michael