Support HTTPS for model downloading

[ellistarn]

Might be worth upstreaming this KFServing PR here: kserve/kserve#979

[tduffy000]

@ellistarn I'm happy to take a crack at it, unless you had other thoughts

[ellistarn]

Could sync with the engineer developing it if you're looking for contributors. This is the future of KFServing's python server.

[adriangonz]

Hey @ellistarn @tduffy000 ,

Thanks for opening up this issue!

I've been thinking lately about this particular point. Should the storage / downloading logic live within the inference server? Currently, on both Seldon Core and KFServing, we perform the download on the init container. Thus, the logic is contained completely outside of the server itself. This becomes extra relevant with multi-model serving where it seems there will be some sort of "puller" doing the download across models.

Recently, in Seldon Core, we also decided to remove the cloud storage logic out of the Python wrapper to reduce the number of dependencies of the seldon_core package (SeldonIO/seldon-core#2310). Instead, we just use the same image as KFServing for the model initialiser.

It would be great to get your thoughts on this! We should also bring this topic to the next WG meeting.

[ellistarn]

This is actually a great point. This logic should probably live in Bloomberg's model puller. Let's discuss at WG.

[adriangonz]

Closing this one as this is already supported in KFServing's agent.