Botond Dénes
Botond Dénes
@ndragazis there are conflicts, please rebase. @avikivity ping, see the performance impact explanation above.
> > One thing to notice is that the CRC32 latency doesn't simply add up to the compaction latency. Is this due to read-aheads? > > Could be that CRC32...
@ndragazis conflicts, please rebase.
> > @ndragazis conflicts, please rebase. > > @ndragazis - don't. Please hold off working on this until we finish higher priority items, such as tablets issues. CC @denesb ,...
> 10 minutes + additional reviewer's work. > I don't mind the 10 minutes, I mind the rest. We don't re-review after rebases. If this PR gets a green CI...
> ❌ - [Docker Test](https://jenkins.scylladb.com/job/scylla-master/job/scylla-ci-docker/352/) There is a problem with this test, will re-kick CI after it is fixed.
> I think there's a chance the idea is to reuse the buffer. > > ```c++ > inline read_status read_bytes(Buffer& data, uint32_t len, FragmentedBuffer& where) { > if (data.size() >=...
> It could be fixed by switching basic_fragmented_buffer::vector_type to be a small_vector. I don't see how this helps. Please explain. Maybe I should have explained it in the cover letter,...
New in v2: fix failing test, which started failing because reads now consume almost half as much memory with large cells.
> > > See how it reuses where.fragments(). This saves one allocation per value. > > > > But it doesn't reuse it, notice the `fragments.clear()`. It reuses `data`, via...