553 blending weights calculation overwrites weights of other ensemble members

[larsbonnefoy]

Previously NWP weights would be overwritten by each worker has they wrote to the same variable storing NWP weights.

The weights are now computed per ensemble member j inside the worker(j) in __blended_nowcast_main_loop, because the NWP skill can differ per member. They are kept on a per-member store (weights_per_member[j], so that members do not overwrite one another.

These changes passes the previous tests but I don't know what should be expected in this newer case and how to write a test for it. I could rely on AI to write one but would have very low confidence in it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.18%. Comparing base (55653a9) to head (7fc6398).

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #554   +/-   ##
=======================================
  Coverage   84.17%   84.18%           
=======================================
  Files         170      170           
  Lines       14985    14989    +4     
=======================================
+ Hits        12614    12618    +4     
  Misses       2371     2371           

Flag	Coverage Δ
unit_tests	84.18% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[RubenImhoff]

Hi @larsbonnefoy, great addition! I will directly have a look at it. If tests are required, I can always add them (although I am on vacation coming week, so it may take a week longer).

[RubenImhoff]

This looks great, @larsbonnefoy, well done!

[larsbonnefoy]

Hi @larsbonnefoy, great addition! I will directly have a look at it. If tests are required, I can always add them (although I am on vacation coming week, so it may take a week longer).

If you have an idea of how we could test this I could also write it ! Otherwise do you have concrete example where this led previously to an issue in a pysteps run ? I could always try a rerun of a past case as our domains overlap. Might be able to spot already some differences.

[RubenImhoff]

It seems challenging to test, because the models will run successfully without the fix. However, I notice a 'jumpy' behavior in the temporal consistency of the ensemble members. So not sure yet if defining a test is feasible here.