Re: Concurrent ALTER SEQUENCE RESTART Regression

On 10/05/17 07:09, Peter Eisentraut wrote:
> On 5/7/17 19:43, Andres Freund wrote:
>> 3. Keep the catalog, make ALTER properly transactional, blocking
>> concurrent nextval()s. This resolves the issue that nextval() can't
>> see the changed definition of the sequence.
>
> This was the intended choice.
>
> [...]
>
> 5. nextval() disregarding uncommitted ALTER SEQUENCE changes. In
> <PG10, it would read the uncommitted metadata and observe it.
> Currently, it goes ahead even if there is an uncommitted ALTER
> SEQUENCE pending that would prohibit what it's about to do (e.g.,
> MAXVALUE change).
>
> I think the correct fix is to have nextval() and ALTER SEQUENCE use
> sensible lock levels so that they block each other. Since
> nextval() currently uses AccessShareLock, the suggestion was for
> ALTER SEQUENCE to therefore use AccessExclusiveLock. But I think a
> better idea would be for nextval() to use RowExclusiveLock
> (analogous to UPDATE) and ALTER SEQUENCE to use
> ShareRowExclusiveLock, which would also satisfy issue #1.
>

When I proposed this upstream, Andres raised concern about performance
of nextval() if we do this, did you try to run any benchmark on this?

Looking at the changes to open_share_lock(), I wonder if we need to have
lock level as parameter so that lastval() is not blocked.

--
Petr Jelinek http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services