Reduce Docker image size from 4.8GB to 2.2GB

[pquentin]

Closes #600

This was done by using the torchcpu PyPI package and not storing the PyPI cache in the Docker image.

Tested using docker run -it --rm --network host -e ES_USERNAME=$ES_USERNAME -e ES_PASSWORD=$ES_PASSWORD eland eland_import_hub_model --cloud-id $CLOUD_ID --hub-model-id elastic/distilbert-base-cased-finetuned-conll03-english --task-type ner

[dolaru]

@pquentin It looks like the torch+cpu variant is only available for Intel/AMD CPUs:

torch-1.13.1+cpu-cp310-cp310-linux_x86_64.whl
torch-1.13.1+cpu-cp310-cp310-win_amd64.whl

This means the image won't build on systems running Docker Engine on Apple Silicon or other system with an ARM CPU. Are we comfortable with that restriction? 🤔

Later edit: When building for x86_64/amd64, it makes sense to use the torch+cpu variant as it doesn't install the large NVIDIA deps we don't need. For aarch64 though, with the normal torch package, the image size is only 1.3GB so we don't need to rely on the torch+cpu variant.

[pquentin]

Thanks! This is great.

The part that is left is building those images in Buildkite, modifying https://github.com/elastic/eland/blob/main/.buildkite/release-docker/run.sh. I've started reading https://docs.docker.com/build/building/multi-platform/ but it seems quite complicated. Do you have any suggestions about the best approach here?

[pquentin]

Thanks to @dolaru, I was able to adapt run.sh to build and push an ARM image. This command worked:

ENVIRONMENT=staging RELEASE_VERSION=8.9.0 with-vault-ci bash .buildkite/release-docker/run.sh

I would appreciate if an Elastic employee with an Apple Silicon Macbook could test the resulting image. (Ping me offline to get instructions.)

[dolaru]

Good stuff @pquentin! I can confirm that the ARM container image works as intended. 👍🏽

I successfully imported the elastic/distilbert-base-cased-finetuned-conll03-english model using eland_import_hub_model in that container image.