Actually, Drew's presentations about Netflix, FreeBSD, ZFS, saturating high-bandwidth network adapters, etc. are legendary and have been posted far and wide. But having him available to answer questions on HN just takes it to a whole 'nother level.
You're making me blush.. But, to set the record straight: I actually know very little about ZFS, beyond basic user/admin knowledge (from having run it for ~15 years). I've never spoken about it, and other members of the team I work for at Netflix are far more knowledgeable about ZFS, and are the ones who have managed the conversion of our fleet to ZFS for non-content partitions.
I’ve devoured your FreeBSD networking presentations but I guess I must have confused a post about tracking down a ZFS bug in production written by someone else with all the other content you’ve produced.
Back to the topic at hand, it’s actually scary how few software expose control over whether or not sendfile is used, assuming support is only a matter of OS and kernel version but not taking into account filesystem limitations. I ran into a terrible Samba on FreeBSD bug (shares remotely disconnected and connections reset with moderate levels of concurrent ro access from even a single client) that I ultimately tracked down to sendfile being enabled in the (default?) config - so it wasn’t just the expected “performance requirements not being met” with sendfile on ZFS but even other reliability issues (almost certainly exposing a different underlying bug, tbh). Imagine if Samba didn’t have a tubeable to set/override sendfile support, though.
