[10:53:42] <elukey>	 hello folks!
[10:54:58] <elukey>	 I am checking model.py for articlequality, I am a little confused about how it works
[10:55:03] <elukey>	 does it use the mediawiki api?
[11:38:15] <elukey>	 I am asking because I don't see in model.py the same mwapi request workflow, and I am confused about how it retrieves the data
[11:38:18] * elukey lunch
[14:17:18] <elukey>	 ok self answering - we currently pass "{"article_text": "I am the text of a page.  I have a <ref>word</ref>"}" as feature from the client, so there is no api call to fetch the wikitext or similar from the mw api
[14:17:22] <elukey>	 this bit was not clear to me :)
[14:37:29] <elukey>	 so it is weird that for enwiki-damaging I don't see traffic coming through the egress gateway
[14:38:11] <elukey>	 not only istio metrics, even kserve ones
[14:38:49] <elukey>	 it doesn't make sense
[14:40:54] <elukey>	 yeah I was testing codfw, good job Luca
[14:40:57] * elukey cries in a corner
[14:42:09] <elukey>	 everything works, updating codfw too :)
[14:49:37] <elukey>	 ml-serve-codfw should be in sync!
[15:07:02] <chrisalbon>	 Morning all!
[15:07:11] <chrisalbon>	 Today I have a LOOONG survey to fill out
[15:14:11] <elukey>	 morning!
[15:38:39] <accraze>	 o/
[15:41:06] <accraze>	 elukey: that is correct, articlequality predictor takes `article_text`, the transformer will be the one that hits mw api (or feature store)
[15:45:20] <elukey>	 accraze: perfect, it should be instructed (as we do now for editquality) with the egress gw's endpoint + host header, and then it should work :)
[15:45:39] <elukey>	 I am currently checking https://www.envoyproxy.io/docs/envoy/latest/configuration/upstream/cluster_manager/cluster_circuit_breakers#config-cluster-manager-cluster-circuit-breakers
[15:46:00] <elukey>	 seems really nice, I don't think we use anything like that in the other k8s clusters
[15:46:30] <elukey>	 the idea is to apply back pressure on our side if the traffic is too big towards the mw api
[15:46:37] <chrisalbon>	 whoa
[15:47:27] <elukey>	 istio ingress/egress gateways are basically envoy proxies, so we have a lot of nice things to use :)
[15:49:14] <chrisalbon>	 TIL our new team member will get an Apple M1 laptop. Anyone see any major issues with that?
[15:49:51] <accraze>	 chrisalbon: running revscoring was tough on osx, but is do-able with a vm or docker etc.
[15:50:11] <accraze>	 err macOs
[15:50:35] <accraze>	 but most of our work is either on ml-sandbox or other boxes so it should be fine
[15:50:36] <chrisalbon>	 I could never get it to run without vm or Docker on macOS
[15:50:42] <chrisalbon>	 Yeah, that was my thought
[15:50:49] <chrisalbon>	 okay cool
[15:50:51] <chrisalbon>	 10 cores!
[15:50:56] <accraze>	 dang!
[15:51:42] <accraze>	 the kserve stack will probably be able to run locally on that
[15:52:26] <chrisalbon>	     Apple M1 Max chip with 10-core CPU, 32-core GPU, and 16-core Neural Engine
[15:52:26] <chrisalbon>	     64GB unified memory
[15:52:26] <chrisalbon>	     1TB SSD storage
[15:52:26] <chrisalbon>	     16-inch Liquid Retina XDR display
[15:52:26] <chrisalbon>	     Three Thunderbolt 4 ports, HDMI port, SDXC card slot, MagSafe 3 port
[15:52:26] <chrisalbon>	     140W USB-C Power Adapter
[15:52:26] <chrisalbon>	     Backlit Magic Keyboard with Touch ID - US English
[15:52:27] <chrisalbon>	     Accessory Kit
[15:52:47] <accraze>	 wait what is the neural engine?
[15:53:11] <chrisalbon>	 Apple have been including a specific neural chip in its latest machines
[15:53:26] <chrisalbon>	 to do more edge ML
[15:53:42] <chrisalbon>	 I've never used it though
[15:54:35] <accraze>	 "enables up to 15x faster machine learning performance"
[15:54:37] <accraze>	 waaat
[15:54:59] <accraze>	 ok im kinda interested
[15:58:39] <chrisalbon>	 yeah, you can't use it for training (yet) BUT I think the idea was that it means that Apple and developers can put complex neural networks into their desktop applications and websites which Apple's neural chip will run locally
[15:58:47] <chrisalbon>	 rather than hogging the CPU inefficiently
[16:03:05] <accraze>	 wow interesting, i do think we'll see some neat stuff as edge AI/ML matures
[16:05:44] <accraze>	 it also opens up some interesting questions about privacy etc.. too
[16:08:09] <elukey>	 64G of ram :D
[16:13:49] <accraze>	 ^ might be more than all of our laptops combined lol
[16:14:17] <accraze>	 the kserve stack will def be able to run locally
[18:57:18] <accraze>	 i'll file a ticket for that pipeline issue we talked about during the tech meeting
[18:59:55] <elukey>	 <3
[19:00:00] * elukey afk!
[19:33:55] <wikibugs>	 10Lift-Wing, 10Machine-Learning-Team (Active Tasks): Fix pipeline image publishing workflow - https://phabricator.wikimedia.org/T297823 (10ACraze)