On Wed, Jul 4, 2018 at 5:08 AM, Pavan Deolasee <pavan.deolasee@gmail.com> wrote:
> But AFAICS get_relation_info() tries to lock every index and since REINDEX
> will be holding a AEL on the index being reindexed, get_relation_info()
> blocks. Since get_relation_info() gets into every read path, wouldn't a
> concurrent REINDEX pretty much block every read access to the table, even if
> REINDEX not holding AEL on the table itself?
Not necessarily -- prepared statements may not block.
> I wonder if we just need fix the docs to or if we actually regressed at some
> point in the history or if we have a bug in the implementation? It mostly
> seems like a case of wrongly written docs even though in theory it might be
> possible to skip an index being rebuilt.
I still agree with this, though. The practical distinction between
getting an AEL on the table and what REINDEX does is pretty much
indistinguishable from zero.
--
Peter Geoghegan