Travis Downs

``` { "duplicate": "https://github.com/redpanda-data/redpanda/issues/18372" } ```

``` { "duplicate": "https://github.com/redpanda-data/redpanda/issues/18372" } ```

`seastar::internal::pending_io_request` is 56 bytes, so this means that the iosink queue of resizing to 458752 can hold 8192 of them, i.e., there are > 4096 ios in flight in the...

@StephanDollberg I didn't but nice to see our analyses line up. I'll take a shot at swapping in chunked_fifo.

This swap is not totally trivial because: - _pending_ios has `operator[]` used on it, which chunked_fifo can't support efficiently (but the use can be moved, it is actually being used...

Floating solutions upstream: https://github.com/scylladb/seastar/issues/2355

Yeah that's a good idea, let me do that.

I ended up going with the `chunked_fifo` approach as `operator[]` on boost::deque has terrible performance (as discussed elsewhere). The result is: https://github.com/scylladb/seastar/pull/2357

Recent x86 failure: https://buildkite.com/redpanda/vtools/builds/15600

Per the discussion on the seastar issue above, it is possible that io properties would mitigate this issue in some cases, since the scheduler may limit the per-tick movement of...