[09:25:02] o/ looks like one of the wdqs hosts is missing some ammonut of data? https://phabricator.wikimedia.org/T396577#10903346 [09:27:48] although, i guess it might even only be a single triple thats missing, even though its 900 results missing [13:50:05] \o [13:53:24] * ebernhardson wonders when my irc clock changed from UTC back to PDT [14:02:08] Trey314159: updated report if you want to take a look before the wed meeting: https://people.wikimedia.org/~ebernhardson/T262612-dym-ab-analysis-DRAFT.html [14:04:00] ebernhardson: I may have a meeting, but if not I will take a look [14:06:21] one thing it's missing is raw numbers...but i couldn't decide which to include. Still pondering [14:06:38] Essentially i feel like this gives relative information, but you can't tell the impact with only % [14:25:19] I didn't miss the raw numbers. The top-level sample is large and the credible interval bars implicilty incorporate the scale of the underlying numbers. That said, I wouldn't object to more numbers—but I'm currently okay without them. (If any are small enough for you to worry about or big enough to want to show off, include them!) [14:25:53] Trey314159: i suppose whats missing is it's not obvious that auto-rewrite has 10x (100x?) the impact of manual rewrites, there isn't really a scale [14:27:51] Ah, true. You could add a little table before those graphs showing the number of relevant sessions/queries. [14:32:12] Are we meeting in ~10 minutes, or should we meet after the P&T staff meeting? [14:56:03] oh umm, i hadn't noticed that [14:56:36] i imagine p&t can be watched on youtube later [14:58:05] I prefer to watch the recording later, too, but if people want to be there live I understand. I guess we can see who shows up and whether we can have the conversations we want to have. [15:01:00] We should show up for the Wednesday meeting since we have guests [15:02:37] pfischer: ^^ [15:04:32] just a heads-up that I deployed cirrus-streaming-updater consumer in staging to clear out references to relforge hosts that are going away (ref T390565). It looks like it started OK, but if there's anything else I need to check LMK [15:04:33] T390565: decommission relforge100[34] - https://phabricator.wikimedia.org/T390565 [16:02:00] break, back in ~40 [16:48:06] ebernhardson: you're probably appear but just a heads up we have the traffic team meeting in 13' to deploy the switching search between datacenters work [16:48:13] probably aware* [17:00:42] nvm traffic pushed back to 13:30 EDT [17:10:52] ebernhardson I had a brain fart this morning and accidentally merged your mediawiki-config patch instead of scheduling a window. I reverted immediately and releng says I'm good, but I made a new patch if you have time to review it: https://gerrit.wikimedia.org/r/c/operations/mediawiki-config/+/1155738 [17:27:45] no worries :) [19:48:14] Trey314159: curious thing i came across. It's looking more narrowly at autorewrites where glent provided the query. Curious in multiple ways :P https://phabricator.wikimedia.org/F62295622 [19:48:41] ebernhardson: looking... [19:51:27] 51% is huge, and the 1.3% decline is not great. But then i also wonder about how if you keep drawing enough graphs eventually one will, by chance, be interesting :) [19:52:25] so glent auto-rewritten queries are much more clickable (~50%) than all auto-rewritten queries (~25%), but method 1 drags down the average a bit. [19:52:39] yea basically. not a show stopper at all, but curious [19:54:37] Seems like a net increase in clickthroughs, since method 1 generates more candidates. Recall goes up a little, precision goes down a smidge—sounds about right [20:10:24] if we just take some raw counts, m1 got 1300 extra impressions, and 320 extra clickthroughs. So clickthrough on the new suggestions was 24.6% [20:11:22] which matches the wider trend of phrase suggester suggested queries, but somehow glent m0 has 50%. It's just got good precision i guess [20:39:28] Yeah, M0 always seemed like the best & most targeted approach.. it's just more limited in scale. [21:18:03] inflatador: i think you mentioned something earlier today about cluster restarts, there is also https://gerrit.wikimedia.org/r/c/operations/software/opensearch/plugins/+/1143156