Re: create index concurrently blocks on transactions in other databases

Simon Riggs <simon(at)2ndquadrant(dot)com> writes:
> On Thu, 2007-09-06 at 12:43 +0200, hubert depesz lubaczewski wrote:
>> i dont really see why this would be neccessary - i understand that we
>> need to wait for transactions in the same db, but waiting for
>> transaction in another database?

> Manual says "and in addition it must wait for all existing transactions
> to terminate." This is needed for MVCC correctness and its not a bug.

No, it isn't needed for MVCC correctness. We're just out-waiting
transactions that might try to use the index and not find what they
should in it. AFAICS Hubert is right: there is no need to wait for
xacts in other DBs. (At least for non-shared indexes, but we can't
support C.I.C. on shared tables anyway.)

This would have been painful to fix in 8.2 but it seems relatively
trivial in HEAD --- we just need another filter option in
GetCurrentVirtualXIDs.

regards, tom lane