Ashley Scillitoe

Postponing this to `v0.12.0` so that we can combine it with deprecation of legacy saving and #723 .

Leaving this until after v0.11.0 based on offline discussion...

Heterogeneous workers would also be beneficial for the recently added support for online drift detectors (https://github.com/SeldonIO/MLServer/pull/1108), since these detectors must currently be run with `parallel_workers = 0`.

Also echoing this, in my experience the results from cuML's UMAP are often unusable, which is a shame as it's so fast! Funnily enough though I've seen the opposite behaviour...

Regarding the codecov report, the `4.32%` decrease is not accurate. This is based on an old `master` commit where we were still counting tests in coverage.

@arnaudvl @ojcobb (and @jklaise/@mauicv) could do with your thoughts on this. In the latest implementation, I have removed the new `reset_state` method. State can still be reset with the existing...

Additional side-note, this issue with setting random seeds not giving deterministic behaviour for a given operation (in this case the `torch.randperm` called by `reset` -> `_initialise` -> `_configure_ref_subset`) is something...

@arnaudvl @ojcobb a possible alternative strategy to make `reset` deterministic is to rework `_initialize` methods to ensure they are deterministic. For example, for `LSDDDriftOnline`, this would involve avoiding the `while`...

## Offline vs online state Based on offline, our current view is that we have a distinction between offline state (stateful attributes such as `self.thresholds` computed at instantiation) and online...

> I will take a closer look at the code next week but the overall approach you've described sounds sensible to me. I agree that there's no need for a...