On Wed, Feb 14, 2024 at 04:12:54PM +0100, Han-Wen Nienhuys wrote:

Good catch!

Sorry for messing this up.

Nothing to be sorry about, bugs happen to all of us.

quoted

In the worst case,
this can lead to a compacted stack that is missing records.

Yeah, that would be an insidious corruption. Have you considered
writing a test to reproduce this (and thus verify that the fix really
fixes the problem?)

I think it wouldn't be too difficult: you could create a custom
blocksource wrapper that returns I/O error on the Nth read, and then
create a reftable with two ref blocks (could just be 2 records if you
use a small blocksize and a large refname) and two log blocks.  Merge
that with an empty table, and see if the compacted result is what you
got in. Loop over N to get coverage for all error paths.

Ah, that's a feasible way to write such a test indeed. I may come back
to it in the future and will add it to our backlog. Thanks!

Patrick