Magnus Hagander <magnus@hagander.net> writes: > Yeah, there's clearly a second problem here.
I think this test script is broken in many ways.
It's scribbling on the source cluster's disk files and assuming that that translates one-for-one to what gets sent to the slave server --- but what if some of the blocks that it modifies on-disk are resident in the source's shared buffers? I think you'd have to shut down the source and then apply the corruption if you want stable results.
It doesn't actually use a slave server as part of the tests.
And basebackups don't read from the sources shared buffers, but it *does* read from the kernel buffers.
I'd bet a good lunch that nondefault BLCKSZ would break it, as well, since the way in which the corruption is induced is just guessing as to where page boundaries are.
Yeah, that might be a problem. Those should be calculated from the block size.
Also, scribbling on tables as sensitive as pg_class is just asking for trouble IMO. I don't see anything in this test, for example, that prevents autovacuum from running and causing a PANIC before the test can complete. Even with AV off, there's a good chance that clobber- cache-always animals will fall over because they do so many more physical accesses to the system catalogs. I'd suggest inducing the corruption in some user table(s) that we can more tightly constrain the source server's accesses to.
Yeah, that seems like a good idea. And probably also shut the server down while writing the corruption, just in case.
Will stick looking into that on my todo for when I'm back, unless beaten to it. Michael, you want a stab at it?