[09:53:43] sigh... needs to fix the metrics I calculated in T405869, recall baseline is incorrect, it's the recall of one the model after epoch 1, so perhaps there's some improvement after all, rerunning something [09:53:44] T405869: Tune the perfield_builder_relaxed query builder profile - https://phabricator.wikimedia.org/T405869 [09:53:56] lunch [10:21:59] yes there's a improvement now, baseline recall@10 is actually at 0.899 vs 0.927 for the trained version, I might upload the new weights, could not hurt [10:24:20] but a bit doubtful it'll change much the relforge report that Erik already compiled and the new MLR model [14:11:49] \o [14:21:51] o/ [14:28:06] looking into __DELETE_GROUPING__...apparently at some point i thought `testCase` was a good name for a test :P [16:20:30] :) [16:35:36] meh, the images cindy uses still uses debian buster, which is long EOL [16:46:25] * ebernhardson is having no luck reproducing... [16:59:48] :/ [17:00:03] heading out, have a nice week-end! [17:08:20] the only thing i've come up with so far is the upsert...but the counts seem way too high for those to all have been upserted [17:08:39] (but i'm pretty sure this does incorrectly upsert the __DELETE_GROUPING__, so something to fix) [17:09:13] or it might completely miss them...testing needed [17:59:47] random thought...we maintain the noopHints in the Update class that tracks what to do, but in the request converter we have to specifically handle those anyways...adding some bits to verify those are always set as expected at least [18:03:12] ebernhardson CC'd you on https://gerrit.wikimedia.org/r/c/operations/puppet/+/1196022 ...no action needed, just a heads-up that we're gonna start mirroring upstream opensearch repos and we're getting more specific w/our version numbers [18:20:21] coo [18:20:22] l