On 2017-04-20 13:32:10 +0200, Petr Jelinek wrote:
> On 20/04/17 02:09, Andres Freund wrote:
> > On 2017-04-17 21:16:57 -0700, Andres Freund wrote:
> > I'm working on some infrastructure around this. Not sure if it needs to
> > be committed, but it's certainly useful for evaluation. Basically it's
> > a small UDF that:
> > 1) creates a slot via walsender protocol (to some dsn)
> > 2) imports that snapshot into yet another connection to that dsn
> > 3) runs some query over that new connection
> >
> > That makes it reasonably easy to run e.g. pgbench and continually create
> > slots, and use the snapshot to run queries "verifying" that things look
> > good. It's a bit shoestring-ed together, but everything else seems to
> > require more code. And it's just test.
> >
> > Unless somebody has a better idea?
> I don't. I mean it would be nice to have isolation tester support
> walsender protocol, but I don't know anything about isolation tester
> internals so no idea how much work that is.
Now that the replication protocol supports normal queries, it's actually
not much of an issue on its own. The problem is more that
isolationtester's client side language isn't powerfull enough - you
can't extract the snapshot name from one session and import it in
another. While that might be something we want to address, I certainly
don't want to tackle it for v10.
I'd started to develop a C toolkit as above, but after I got the basics
running I actually noticed it's pretty much unnecessary: You can just as
well do it with dblink and some plpgsql.
I can reliably reproduce several of the bugs in this thread in
relatively short amount of time before applying the patch, and so far
not after. Thats great!
> I guess you plan to make that as one of the test/modules or something
> similar (the UDF)?
I've a bunch of tests, but I don't quite know whether we can expose all
of them via classical tests. There are several easy ones that I
definitely want to add (import "empty" snapshot; import snapshot with
running xacts; create snapshot, perform some ddl, import snapshot,
perform some ddl, check things work reasonably crazy), but there's
enough others that are just probabilistic. I was wondering about adding
a loop that simply runs for like 30s and then quits or such, but who
knows.
Testing around this made me wonder whether we need to make bgwriter.c's
LOG_SNAPSHOT_INTERVAL_MS configurable - for efficient testing reducing
it is quite valuable, and on busier machines it'll also almost always be
a win to log more frequently... Opinions?
Greetings,
Andres Freund