Support building functions with Dockerfile about openfunction

[lizzzcai]

Building from dockerfile is a nice feature and it can provide more flexibility to users.

The current build from source is not so optimized. I tried to build the same hello world program, one is from source and another is from dockerfile. For image built from dockerfile, the image size is about 1.9MB. For image built from source, the size is about 260MB. Hopefully, it can be optimized in the future.

A side topic, I want to check if OpenFunction can support customized run image. The target here is to reduce the image size so that can reduce the cool start time.

Usecase:
for machine learning use cases, there are usually a lot of dependencies like tensorflow, pytorch, opencv etc. The size of these kinds of libraries are quite big (up to 1-2 GB usually) and the actual code is about 1-2 MB (not including the models). If the base libraries can be separated and cached in the host machine, I think that will help to reduce the cool start time.

A similar result is shown in this post for your reference.

An example function yaml:

apiVersion: core.openfunction.io/v1alpha2
kind: Function
metadata:
  name: iris-python
spec:
  version: "v1.0.0"
  image: "lizzzcai/iris-python-func:v0.4.0"
  imageCredentials:
    name: push-secret
  port: 8080
  runImage: tensorflow/tensorflow:2.2.3-py3 # dependencies to run the python code
  build:
    builder: openfunction/builder:v1
    env:
      FUNC_NAME: "predict"
      FUNC_TYPE: "http"
      FUNC_SRC: "main.py"
    srcRepo:
      url: "https://github.com/lizzzcai/openfunction-samples.git"
      sourceSubPath: "iris-python"
  serving:
    runtime: Knative

I can create a feature request for this if needed. Thanks.

[benjaminhuo]

@lizzzcai Thanks for the suggestion!
I think there're two problems here:

1. Optimize the image created by cloud-native buildpacks approach to reduce image size
2. A method to preload the run image to nodes to reduce the cold start time

You're welcome to create an issue for this!