On Tue, Mar 3, 2015 at 7:49 AM, Andres Freund <and...@2ndquadrant.com>
wrote:
> Hi,
>
> I've regularly wished we had automated tests that setup HS and then
> compare primary/standby at the end to verify replay worked
> correctly.
>
> Heikki's page comparison tools deals with some of that verification, but
> it's really quite expensive and doesn't care about runtime only
> differences. I.e. it doesn't test HS at all.
>
> I every now and then run installcheck against a primary, verify that
> replay works without errors, and then compare pg_dumpall from both
> clusters. Unfortunately that currently requires hand inspection of
> dumps, there are differences like:
> -SELECT pg_catalog.setval('default_seq', 1, true);
> +SELECT pg_catalog.setval('default_seq', 33, true);
>
> The reason these differences is that the primary increases the
> sequence's last_value by 1, but temporarily sets it to +SEQ_LOG_VALS
> before XLogInsert(). So the two differ.
>
> Does anybody have a good idea how to get rid of that difference? One way
> to do that would be to log the value the standby is sure to have - but
> that's not entirely trivial.
>
> I'd very much like to add a automated test like this to the tree, but I
> don't see wa way to do that sanely without a comparison tool...
>
Couldn't we just arbitrarily exclude sequence internal states from the
comparison?
That wouldn't work where the standby has been promoted and then used in a
way that draws on the sequence (with the same workload being put through
the now-promoted standby and the original-master), though, but I don't
think that that was what you were asking about.
How many similar issues have you seen?
In the case where you have a promoted replica and put the same through
workflow through both it and the master, I've seen "pg_dump -s" dump
objects in different orders, for no apparent reason. That is kind of
annoying, but I never traced it back to the cause (nor have I excluded
PEBCAK as the real cause).
Cheers,
Jeff
