Gleb Natapov 60a851d3a5 commitlog: always flush segments atomically with writing ...

db::commitlog::segment::batch_cycle() assumes that after a write
for a certain position completes (as reported by
_pending_ops.wait_for_pending()) it will also be flushed, but this is
true only if writing and flushing are atomic wrt _pending_ops lock.
It usually is unless flush_after is set to false when cycle() is
called. In this case only writing is done under the lock. This
is exactly what happens when a segment is closed. Flush is skipped
because zero header is added after the last entry and then flushed, but
this optimization breaks batch_cycle() assumption. Fix it by flushing
after the write atomically even if a segment is being closed.

Fixes #5496

Message-Id: <20191224115814.GA6398@scylladb.com>

2019-12-24 14:52:23 +02:00

86 KiB

Raw Blame History

View Raw