Open cgwalters opened 6 years ago
Without doing a lot of measurement, I suspect we want to have potentially different behavior between pulling to a releng (mirror/dev/test) repo vs a system (or separate flatpak etc.) repo.
A releng repo might be on a large data volume and the cost of a syncfs()
is probably much greater than individual fsyncs.
Updated proposal:
Add
We've had lots of issues with this over time...anyways I was debugging a segfault from what I thought was some outstanding work and saw the stack below.
Basically we're doing
fsync()
on object commit even if we're in a txn by default still. I think this is pretty broken. Actually embarassing since we've done all this work on usingsyncfs()
and the per-boot-id staging directory etc. that is much less useful if we're doing individual syncs.