Run check_estimator on our estimators about dask-ml HOT 3 OPEN

Hmm, we'll have to think a bit more about this. It can involve some checks that might be expensive to do, e.g. length checking on a dask.dataframe.

See #72 for some of the edge cases it caught though.

from dask-ml.

Is the expensive part that length checking a dask.dataframe requires .compute()? If there are some parts of the check_estimator logic we have to skip, let's at least keep the checks for get_params() and set_params() as methods.

from dask-ml.

Yeah things like len can be expensive for dask.DataFrames. Other examples would be checking for NaNs / infs: https://github.com/dask/dask-ml/pull/72/files#diff-1be11e310bf210381f8dd957dbc27074R173. Everything else we can do with just the metadata on the dask collection itself.

Though perhaps I'm being overly paranoid about the costs here. Either way, we will want to inform users about which algorithms make multiple passes over the data (like KMeans), so that they'll want to .persist the data before running .fit.

from dask-ml.