spcl/serverless-benchmarks

Our PyTorch-based benchmark has been recently ported to support Python 3.12. We should verify that it runs on 3.13 and 3.14 on all platforms. Furthermore, we should port dependencies to also run on ARM Lambda instances. Since PyPI packages for torch should have CPU-only wheels for ARM, we can try to use that. On x64, we use custom wheels because we don't want to ship the heavy CUDA bundle.

This benchmark relies on a static build of ffmpeg, which we source from an external website. Instead of pulling the ffmpeg binary at every single build, which adds a lot of pressure on that website in CI builds and can lead to errors when pulling data (e.g., rate limits), we could put a pinned and reproducible variant of ffmpeg in our benchmarks-data instead. We should a new resources category in benchmarks-data, and change init.sh script to copy that binary during the initial build. EDIT: we should also revise the command in the function to make sure that we explicitly select codecs and are reproducible.

SeBS keeps a local cache of all deployed cloud resources and locally created artifacts, e.g., function builds. To support distributed deployments, we should support a mode where we merge the cache from N existing caches - remove duplicates without losing allocated resources, as long as they share the same SeBS resource ID. This should also drop local resources, as such merged cache should be usable on any machine.

Currently, the benchmark 411.image-recognition has only a single input class - one image. Since each image taskes roughly the same amount of time, we should generate different classes of input size like small and large by creating a batch of requests to be processed by the function.

Azure will retire the Consumption Plan in 2028, and they recommend a switch to the Flex Consumption plan. We should add support for this, and make it configurable (like GCP gen1 and gen2). We should also try Windows functions. https://learn.microsoft.com/en-us/azure/azure-functions/functions-scale