Benchmarks reproductability

[killianLM]

Hey there,

I am opening this issue concerning the Benchmarks part of River for a few reasons:

• First of all, it currently seems hard to reproduce. I've had multiple issues with it, especially with dependencies, as I had a few dependencies not installing properly (or at all) such as vowpalwabbit. I also had issues with the functionalities of tqdm, where I was forced to remove its use in the run.py file.
• Second, it does not look fully up to date to the latest version of River. For example, the current version of benchmarks calls ensemble.AdaptiveRandomForestClassifier, which does not exist in the current version of River.

In our opinion, it is important that this benchmark be easily reproductible as soon as it is installed, both from a purely scientific point of view, and to have a very useful global test of non regression.
All in all, it looks like it has a few issues, including ease of use, which would benefit from your attention.

Have a good day,
Killian.

[MaxHalford]

Hey there @killianLM. Yes, I totally agree with you on both points. However, the attitude to have in the open source world is not just to point and call out what's wrong. We have limited spare time to work on River, because we have jobs on the side. In this case, why don't you share what exactly is going wrong? What Python exceptions are you facing? More than that, I encourage you to fix stuff yourself if you notice it.

Have a good day!