Avi Kivity

We can auto-detect latency sensitive classes. If the queue is empty (or there are no running tasks) then that queue is latency sensitive. Once a scheduling group becomes overloaded, it...

Since it's experimental, we should wait until it's generally available. Otherwise a change in the protocol could create incompatibility.

It's not enough to state it's experimental. 1. deploy a Seastar application with this code 2. Prometheus changes the protocol and graduates the feature to GA 3. User deploys new...

> > It's not enough to state it's experimental. > > If something is experimental, it needs to be default-off on this side, not somewhere that's not controlled by the...

https://github.com/scylladb/seastar/pull/1786

> @avikivity, if you will merge this PR, I will send the follow-up patches with the experimental flag support (either explicitly like you wrote in your comment, or following Prometheus'...

pipes aren't suitable as files, because our files are non-streaming. You can't implement dma_read (which is pread(2)) on a pipe. Can't S3 files handle append-only in dma_write? Gather buffers from...

> > Can't S3 files handle append-only in dma_write? > > That's the biggest problem. Typically output_stream + file_data_sink_impl would push 128-k buffers into S3 file in a form of...

Linux does this multiple layers of customization points. Maybe it can make sense for zoned devices too.

There's the physical block size in /sys/block/nvme0n1/queue/physical_block_size. Note: we still want to write the commitlog with logical block size, there's hope it avoids RMW since it's a stream.