On Fri, Sep 28, 2018 at 08:57:53AM +0200, Ævar Arnfjörð Bjarmason wrote: > > Thanks. I had ~400 runs of the tests I ran before and they were all > > OK. Now trying also with t1701 (which I hadn't noticed was a new > > test...). > > Ran that overnight with the same conditions as before. 2683 OK runs and > 0 failures (and counting). So it seems like the combination of the two > fixed the split index bugs. Yeah, I thought they would. If you look at the first loop of prepare_to_write_split_index() classifying which cache entries should be included in the new split index, you'll see only two code paths that could leave out an entry from the split index, i.e. where an entry could be left with a non-zero 'ce->index' and without its CE_UPDATE_IN_BASE flag set. Now, with the fix in patch 5/5 both of those code paths have the is_race_timestamp() check.
