'Standard' vapor website drops 1.5% of requests, even at concurrency of 100!

Should do - I just cloned the repo at HEAD.

It's probably worth noting this with some prominence in your post, as while it's not necessarily wrong to raise these until they don't impact the benchmark (especially the file descriptor limit), there ostensibly are reasons that somaxconn exists (to help defend against SYN flood attacks, for example). So arguably it is detrimental, in contrast to the other frameworks tested, for Vapor to be so sensitive to low configuration values.

(though it's also arguable that 128 genuinely is too low for a modern system, probably having been established decades ago when RAM capacities were an order of magnitude lower)

ps or similar running at a high-enough sample rate should work, in theory. But "high-enough" is fuzzy… I'd say at least 10 Hz, but that's a somewhat arbitrary guess.

I seem to vaguely recall that there are more precise (and easier) ways to get peak memory usage. e.g. maybe getrusage (ru_maxrss specifically)? There's probably pre-existing scripts which use it to wrap over a subcommand, like time does. Even if not, it's trival to write a little wrapper.

I've unenthusiastically considered it, but it seems like wrk is abandoned. I could push up my own fork and just leave it, but then it too would be abandoned. :man_shrugging: