[09:49:13] errand [14:15:23] o/ [14:49:16] o/ [14:50:31] \o [14:56:50] Taking a look at T414697 and it looks like we can just run `/usr/share/opensearch/bin/opensearch-plugin install analysis-icu` during the image build. Is that acceptable, or do we need to get it from maven, build it ourselves, etc? [14:56:50] T414697: Build the required plugins for opensearch 3 - https://phabricator.wikimedia.org/T414697 [14:58:25] inflatador: well, depends. I suppose the historical reason was that it locks down the security side of updates. When installing a plugin by the shell script what we get varies, when it's a hash in a git repo we know that it (probably) never changes. But for how important that is...i dunno [15:07:34] yeah, me neither. Looks like maven has checksums https://repo1.maven.org/maven2/org/opensearch/plugin/analysis-icu-client/3.3.2/ . Maybe that's an option. I might look at our flink images, I know we're using checksums there too [15:08:40] https://gerrit.wikimedia.org/r/plugins/gitiles/operations/docker-images/production-images/+/ec3785527a1dada3e022a57a50c43da269afe82d/images/flink/flink/maven-download.sh [15:24:33] meh, forbiddenapi's doesn't like me using the "jdk internal/non-portable" HttpServer in unit tests [15:27:04] o/ [15:36:40] rebooting, back in a few [16:06:18] * ebernhardson never enjoys guessing throught a build.gradle to figure out how it publishes [16:08:57] oh, for the opensearch-innerhits plugin? yes no clue how this is working :( [16:09:15] atm i'm just trying to copy what we do in the stconvert plugin, but not there yet : [16:09:17] :) [16:15:43] huh, never noticed before but gitlab project access tokens have a "self_rotate - grants permission for token to rotate itself" scope. [16:25:06] interesting, does it copy the token to the expected CI vars? [16:26:17] indeed i probably have to look more into how it works, because it doesn't auto-save the variable i expect it also wouldn't auto-update it. But by default you can't even get the token value later on [17:55:39] * ebernhardson is somehow never happy with gradle :P [18:31:41] hm not sure what to do with the mapping, mainly adding fields we extract by defaults (from _source) but I see we display the token_count for I'd have to index the whole content and configure much more analysis settings... [18:32:20] dcausse: can we skip it? I suppose might be annoying in cirrus / result classes, but really thats for a Special:Search UI affordance [18:33:40] true, but I'm afraid it'll fail if not declared? would need some adaptation to the cirrus result class but I think this is something you had to implement anyways? [18:34:20] yes I see https://gerrit.wikimedia.org/r/c/mediawiki/extensions/CirrusSearch/+/1229645/7/includes/Search/SemanticResultsType.php [18:34:49] so we could possibly drop this line 42 [18:35:30] dcausse: yea i think we can drop it, i'm not sure the downstream implications but we can probably put a null somewhere [18:35:50] ack, not adding the whole content just yet [18:37:53] meh, managed to publish innerhits to gitlab...but no zip :( [18:38:01] also it published to innerhits and inner-hits :S [18:42:41] :/ [18:43:51] it's the CI+gradle process, maybe by the 4th or 5th publication it will work [19:33:11] seems it was a good guess. -wmf5 published just the zip and pom without any scary warnings [19:33:41] javadoc complains, but i don't agree that, for example, you have to have a blank default constructor so it can have a doc block [20:25:07] err, sigh..i added the plugin to relforge1008, but missed analysisicu and now we have red and green indices...not sure how fixable yet [20:38:22] oh good, it didn't delete anything on failure. Added the plugin and ran _cluster/reroute?retry_failed=true [20:53:35] {◕ ◡ ◕} [20:54:32] the fun part now is configuring everything :P I'm going to start a patch in https://gitlab.wikimedia.org/repos/search-platform/cirrus-toolbox to start holding the json files that define everything [20:55:08] record what we've configured so far, and then start adding the bits for inner_hits llama reranking [21:02:35] separately interesting, claude.ai now has a Q/A format where it presents multiple-choices and you select. [21:10:01] hmm, maybe cirrus-toolbox is the wrong repo...but trying to not have a million repos