Update CI conda deps to match training/scoring SDK #263
Conversation
- Tied SDK version to 1.2.x as with conda_dependencies.yml - Allow deps to fall out based on SDK - Kept the rest of the deps manually specified to keep image size small and minimize regressions
|
@tcare explicitly specifying the versions for the dependencies will be safer. |
|
Explicitly setting for the top level SDK and letting it resolve its dependency tree should be sufficient for most. However we should probably manually specify pip as well. Edit: we can also freeze these at the new versions if you like, however at least letting point versions update would be preferable. I feel we're going to get breaks either way, but specifying the larger top level dependencies should reduce most of the pain. |
|
For the CI image in particular, this would also let us update the image easily if there was a needed update from one of these deps. Otherwise the image will stay static since we deploy it manually and version tag it. |
|
I'm fine to lock at bug fix (i.e. 1.4.x) but no more loose than this. We've had a bunch of problems with deps breaking things in the past leaving no versions on (as you currently have in the PR). |
|
Do we know which ones in particular were causing issues? I didn't see too many in the yml file histories (apart from azure-cli which needed updating due to an upstream change) |
|
Versions have been locked down and a successful run is at https://aidemos.visualstudio.com/MLOps/_build/results?buildId=4391&view=results |
Passing build at https://aidemos.visualstudio.com/MLOps/_build/results?buildId=4388&view=results
Test image deployed at mcr.microsoft.com/mlops/python:tcare-test