Re: [GENERAL] Using an SMP machine to make multiple indices on the same

"Hiroshi Inoue" <Inoue(at)tpf(dot)co(dot)jp> writes:
> In the end, I changed DefineIndex() to not call IndexesAreActive().

I saw that. But is it a good solution? If someone has deactivated
indexes on a user table (ie turned off relhasindex), then creating a
new index would activate them again, which would probably be bad.

I have realized that this code is wrong anyway, because it doesn't
acquire ShareLock on the relation until far too late; all the setup
processing is done with no lock at all :-(. LockClassinfoForUpdate
provided a little bit of security against concurrent schema changes,
though not enough.

Also, I'm now a little worried about whether concurrent index creations
will actually work. Both CREATE INDEX operations will try to update
the pg_class tuple to set relhasindex true. Since they use
simple_heap_update for that, the second one is likely to fail
because simple_heap_update doesn't handle concurrent updates.

I think what we probably want is

1. Acquire ShareLock at the very start.

2. Check for indexes present but relhasindex = false,
if so complain.

3. Build the index.

4. Update pg_class tuple, being prepared for concurrent
updates (ie, do NOT use simple_heap_update here).

I still don't see any value in LockClassinfoForUpdate, however.

regards, tom lane