I tried inlining batchStreamImpl and doing away with the func and intermediary interfaces, I tried commenting out the drpc-related code in the dialer, and I could not reliably produce an improvement. I looked at profiles but couldn't find any of the changed code being responsible for anything signficant. It is possible that it's just due to importing a new library and having more code (and perhaps realigning certain hot code blocks w.r.t cachelines - something PGO in go1.23 is supposed to help with). I will try again looking at exactly the merge commit and the parent one and perhaps importing the change piece by piece.

Unrelated, but I did notice that according to the profile, shouldUseBatchStreamPoolClient takes about 0.5% and we should improve that:

It's hard to get accurate results since the difference is less than the usual standard deviation. But I think reverting all the client-side code gets part of it back (maybe 0.2-0.3%) - probably just a handful of otherwise insignificant changes. Even after reverting most server-side code I still see some regression. I will stop spending time on this now :)