Avi Kivity

> @avikivity you only changed the config where is the change in the code that actually reads this config value ? https://github.com/scylladb/scylladb/blob/8ee5b69f800481e31ecefbeae74d224d3d082152/gms/feature_service.cc#L32

> There's also the flag `enable_sstables_mc_format`, which is already marked "deprecated" but appears to have been worse than deprecated (it seems to have been ignored?) and should probably be removed....

It's mostly the CPU usage.

Our access patterns don't match that. We generate files incrementally, so we can't really upload in parallel (unless we stage them to local disk). Reads are also sequential (for compaction),...

Our compactions run at a rate of tens of megabytes per second, so I don't see pacing being a problem soon. Of course it may be in some workloads.

I agree about reusing parts of the SDK when possible. Could work with alternator too, though that implements the server side and not the client side. @nyh

The question is how much is it important for correctness. I remember looking at the signature stuff and thinking how complicated it is. If we adopt their code, we reduce...

Looks good.

I gave it a very shallow review, but I'm impressed.

These metrics are problematic. They don't count the number of large cells, they count the number of times an sstable writer encountered a large cell since the last restart. Neither...