Michael McCandless

> Like with JMH we should also add enough warmup "per JVM" before we actually measure and then let the JVM run for much longer time. You can control the...

These gains are really quite incredible -- I tweeted about it: https://twitter.com/mikemccand/status/1664285218285694979?s=20 I'll try to cutover nightlies to the larger vectors...

I wonder if we also have some slowness on reading these `float[]` (and `byte[]`) vectors through `IndexInput`? Maybe the Panama native memory APIs could help with this limitation in our...

I'm having trouble evaluating the two options here :) I do agree, simplifying the onboarding of new devs using IDEs that are NOT emacs, heh, is important. Can someone compare...

I love the idea of a unified build system, but in my experience it remains an unachievable ideal -- I wind up hating nearly every build system I use ;)...

> Which do you prefer, loading the reader states eagerly (this PR) or lazily (existing functionality)? +1 to do it eagerly (this PR) -- thank you! The change looks good...

OK, I pushed a rough first cut at this! It produces [this file](https://home.apache.org/~mikemccand/lucenebench/for_all_time.json) each night. Maybe we can build on this to make alarms/anomaly detection/pre-release checks of some kind. There...

> 1. could we add some header column explaining the additional value(s)? Are they variances? Yeah +1 I'll try to add that (they are variances, I think!). > 2\. Can...

It's OK for the KNN results to change -- that's just needs a nightly regolding as long as we know/accept the source/reason for the change.

> > 1. fixing exception handling in DW.updateDocuments to handle this exception gracefully (non-aborting). > > the problem in this method was that we need to flush the DWPT to...