Move the SDK to be just a set of cohesive dependencies (and maybe some initialization code)

[llucax]

Currently the SDK is including a lot of code from different areas of the very complex Frequenz ecosystem. This is starting to be an issue, as any user of the SDK basically needs to pull the dependencies for ALL the ecosystem, even when they are not needed or used.

A classical example is cloud apps, which only use cloud APIs and run in the cloud, needing to pull all the microgrid API dependencies.

I think in the future the SDK should just be a declaration of dependencies that are tested together and are known to work well. Many dependencies can be even optional, so if you are writing a cloud app, you could do pip install frequenz-sdk[cloud] and that won't pull any microgrid related dependencies to your code, just cloud APIs. We can even have more granular optional dependencies, like pip install frequenz-sdk[electricity-trading], which will only pull electricity trading code, or pip install frequenz-sdk[microgrid] if you only want to work with the microgrid.

There will still be stuff that the SDK might need to do, common setup that we need even if you only want to work only with one API, like setting up the Configuration system (ConfigManager), or handling logging properly, etc.

[thomas-nicolai-frequenz]

A classical example is cloud apps, which only use cloud APIs and run in the cloud, needing to pull all the microgrid API dependencies.

Does this hold true when using the microgrid sandbox API or edge apps using the cloud APIs directly?

[llucax]

I'm not sure I understand the question/scenario, but there will always be examples of people wanting to write an app with the SDK that only use a subset of APIs, and it will be overkill to force them to pull the whole Frequenz ecosystem to just use one or two APIs.

That said, I am aware this mostly a very abstract idea, we'll have to see how this works out in practice when we have time to address it. The main issue IMHO is to see a way that a simple and small project doesn't need to pull hundreds of unused dependencies and end up with a huge dependency bloat.