Cleanup _plot_args_replacer logic - #10872

may fail e.g. if a user would use a string-list for data. I assume that's ok-ish.

I'm worried this might mess categorical axes. I'll make a note to check on that.

This will 'work' if you do ax.plot('foo', 'bar', data=['foo', 'bar'])

Why do we not want to pull out the data as part of the test? (i.e. try data[name] and catch attributeerror) "Should" be cheap...

"Probably" you are right.

In the worst case, it could get a factor 2 slower because we'll have to fetch the data later anyway. But I think this is negligible compared to all the other stuff done during plotting. If someone has data with a slow getitem, they can still resort to plot(data['x'], data['y']) to only access them once.

So I'm +1 on try data[name] and catch AttributeError.

Is there more support for "try data[name] and catch AttributeError"? If so, I would create an alternative PR.

h5py objects can be slow.

I would lean towards continuing to use __contains__ and documenting that support that is part of the required API on the things passed in to data=.

If h5py is the couter-example that item access can be slow, I'd leave the implementation logic as is, just with the cleaned up code from this PR.

I've added a note that the data object must support membership test.

I don't mind the rebase, but note that #10928 will probably make all this obsolete.

Don't we need both?

#10928 kills _plot_args_replacer and uses the standard _replacer logic used by everyone else (try to get the item and if it's there, swap it in -- so access is only done one anyways)

(and yes this PR is the original reason why I did #10928)

Note: Adapted _has_item() to fix #11389