Testing of pangolearn analysis mode

[wm75]

I'm trying to update our Galaxy wrapper of pangolin to v4, but our tests with --analysis-mode pangolearn are failing without a proper error message. The same tests are passing locally for me.
Could it be that the pangolearn model has just become too resource-hungry by now to be compatible with our CI?
If that's the case, is there a less challenging way to test the pangolearn mode?

Here's my currently stalled PR: galaxyproject/tools-iuc#4494
I've pulled the important parts of the CI output into the conversation to make it easier for you to help if you can.

Thanks a lot for any insight!

[aineniamh]

Have a read through issue #395, we've added a warning to the latest version as random forest does take more RAM, but actually shows greater accuracy whilst still being a very fast method!

[wm75]

Thanks @aineniamh. Had missed the issue with the actual numbers.
Is there no workaround for testing purposes then? If possible, I would like to have the pangolearn branch covered by our Galaxy wrapper tests and not just test usher exclusively.

[wm75]

btw, my local machine only has 8GB of memory, and if I close all other programs I can make our tests (which use just one single sequence) pass using one thread only. So at least under ideal conditions you can still get away with less than the 12GB mentioned in #395 :)

[aineniamh]

So I think the actual ram usage is over 7GB because the github actions tests fail on a small file- but for macosx they have 14G available so they pass. I'm not the developer of pangolearn (she's sadly for us all moved on to a new role) so I'm not going to be much help with playing with the RAM requirements. I know that if we decrease the number of components in the random forest the RAM requirements decrease too, but with a big decrease in assignment accuracy too though.

[wm75]

Thanks for the clarifications! At least it's good to know what's going on.