seastar::make_file_output_stream by-pass fdatasync?

[nvartolomei]

Hi, we have some cases where we use seastar::output_stream created by calling seastar::make_file_output_stream for writing temporary files (quite a few of them) and I think we could benefit from avoiding calling fdatasync on them. Currently, to make sure the (output stream buffered) data made it to the file we have to call seastar::output_stream::flush() (and close() too) but that flush also calls unconditionally seastar::file::flush() which results in a fdatasync call.

Any thoughts on adding an option for avoiding that? I.e. as an option inside ss::file_output_stream_options.

[dotnwat]

do i understand correctly that you want to be able to (1) flush buffered data from the output_stream to the underlying file so that it is visible through the file interface, and (2) do that without incurring a call to fdatasync, but can't avoid it because output_stream::close|flush both call flush on the underlying file?

that's what the behavior looks like to me and the alternative makes sense. from what i've observed, the writes incurred by flushing the output_stream buffers to the underlying file (which are always necessary) are expensive in comparison to the fsync which is usually very quite speedy on nvme.

[nvartolomei]

Yes, you described it perfectly.

from what i've observed, the writes incurred by flushing the output_stream buffers to the underlying file (which are always necessary) are expensive in comparison to the fsync which is usually very quite speedy on nvme.

I've made the opposite observation when writing small temporary files due to disk volatile caches being involved. Especially on consumer grade hardware. I guess the same would be the case with EBS like storage too.