Optimize docker build

[kelvin-wong]

Split from #109

Problem 3: Python images are slow to build

Cause of slow build

From the logs on building Docker images using skaffold. There are 2 main time-consuming processes. They are PNPM package installation and Poetry package installation.

Proposed optimization

Make use of PNPM workspace

PNPM could define dependencies for different packages in the monorepo. The way it works is similar to our implementation on subset.config.js. The main difference is that the PNPM workspace shares a single node_modules for all packages. It will save a lot of time on downloading packages/building packages.
https://pnpm.io/workspaces
@solomondefi-dev could look into this.

Extract base build stage for Python apps

The current base build stage for each Python apps is duplicated. We could extract it into another image and save some build time.
Another potential optimization could be a similar workflow like PNPM workspace as Poetry cache packages in .cache folder by default, the cached packages could be shared across different Python apps.

[kelvin-wong]

https://github.com/jacksmith15/poetry-workspace-plugin is one of the PoC of the python workspaces for poetry. It's just quite an early stage of the development but it basically shares packages/libs used in different projects in monorepo.

[kelvin-wong]

Some notes on slow docker build for Python apps.
Currently, pre-built binary packages (wheel) are built from glibc. But Alpine is based on musl and it will be regarded as a different platform for pip. Building from source is unavoidable for some packages until PEP656 is employed or a general platform package is available.

[solomondefi-dev]

That makes sense. I wonder if it's worth trying bullseye-slim instead of alpine, it seems like with Python, most of the benefits of alpine are lost.

[solomondefi-dev]

Closing this since the task is completed as originally defined. Will add an icebox issue for exploring other base images.