The bike-rentals-auto-ml image has 1.6 GiB

[adelton]

Unlike the tensorflow-housing image that has 120.6 MiB (at https://quay.io/repository/rhoai-edge/tensorflow-housing?tab=tags&tag=latest, and also reproduced in my setup), the bike-rentals-auto-ml reported size is 1.6 GiB, at https://quay.io/repository/rhoai-edge/bike-rentals-auto-ml?tab=tags and again in my setup -- see #93 (comment) and #93 (comment).

Assuming both just ship a runtime without bundling the AI model, there seems to be some serious bloat in bike-rentals-auto-ml.

The pipelines/README.md should likely explain why those container images differ by an order of magnitude in size, what is so special about the second one that it warrants the extra bits.

[piotrpdev]

Assuming both just ship a runtime without bundling the AI model, there seems to be some serious bloat in bike-rentals-auto-ml.

Can you elaborate? Both include the serving runtime and model.

The serving runtime image alone for bike-rentals is almost 900MB compressed: seldonio/mlserver:1.3.5-slim. It is large, but supports a wide range of models: https://github.com/SeldonIO/MLServer#inference-runtimes

ai-edge/pipelines/containerfiles/Containerfile.seldonio.mlserver.mlflow

Lines 28 to 30 in d2972ab

	# Create the MLServer container. Use the slim image, since we are providing an environment tarball.
	#
	FROM docker.io/seldonio/mlserver:1.3.5-slim

The bike-rentals image also contains all of the dependencies for the model:

ai-edge/pipelines/containerfiles/Containerfile.seldonio.mlserver.mlflow

Lines 16 to 30 in d2972ab

	# Download model dependencies and create a portable tarball
	# The tarball is placed inside the model directory.
	RUN . ~/miniconda3/bin/activate && \
	conda env create -n mlflow-env -f model/conda.yaml && \
	conda activate mlflow-env && \
	pip install mlserver-mlflow && \
	conda list && \
	conda deactivate && \
	conda activate && \
	conda install conda-pack && \
	conda-pack -n mlflow-env -o model/environment.tar.gz
	# Create the MLServer container. Use the slim image, since we are providing an environment tarball.
	#
	FROM docker.io/seldonio/mlserver:1.3.5-slim

See #27 for seldonio/mlserver vs openvino/mlserver.

[adelton]

Our discussions typically revolved around the model being huge. In this case though, the model is small (54 MB) and the runtime is huge.

I wonder if we are able to demonstrate the behaviour with a runtime that might not be so generic but whose size would be more palatable.

[piotrpdev]

seldonio/mlserver and openvino/mlserver are popular and cover most of the model frameworks and types/flavors. If we're looking for the best compatibility using these along with MLflow is probably the best option.

We can take the KServe approach and write our own runtimes if we want many smaller runtimes instead of a couple of large ones ¯\(ツ)/¯.