Is the endgoal of this PR (and further work) to be as close as possible to the ones in numpy? I ask because I'm a long time user of numpy, as is everyone in the company I work for, and nobody used this numy.ma, ever. I mean, we do use masked arrays all the times in medical imaging, mostly for ignoring invalid voxel (outside of brain, etc.), but the tools in numy.ma never had any appeal I guess.

In brief, I see 3 problems with this (not your PR spedifically, but the concept)

1. AFAIK, it's mostly useless. This is my opinion and the result of a survey in my small company. I don't really believe it, but keep in mind that it's possible that we're simply ignorant of the right usage, or have simply never needed it.
2. It forces the library to duplicate code. Like, mean for Array and MaskedArray, std-dev, etc. Maybe this is less of a hassle than I think.
3. Most all(?) usages can be replaced with a select() or a Zip::from(arr).and(mask). That's what we're currently doing. in fact, I think that's the problem with numy.ma: it's super easy to avoid. You simply use arr[mask > 0] and some other indexing tools, et voila.

What I would gladly use is a Zip::masked function, like

Zip::from(&brain).and(&wathever).and(&mut out).masked(mask).par_apply(|&b, &w, o| {
 // No mask == false here, yay! But it's simply avoiding a if...
})

but this is somewhat irrelevant to the current discussion :)

I appreciate reading your sketch andrei, you're more productive than me, just having a go at a draft instead of trying to make something perfect.

I think it's been mentioned before yeah, the question whether to have masked arrays or masked operations on arrays. I dread the complexity of either. Thanks nilgoyette for the candid thoughts too.

I think we should start with masked operations. I think that's what a masked array type (if it were to exist) would need as basis anyway. And it allows having a separate mask too - which should hopefully be more efficent (packed or sparse bitmap?)

I realise this is an older PR but I would vote in favour of masks. I'm by no means experienced in numpy so the following may have a different solution.

My current use case is a 2D jagged array of ids. Each row represents the following:

[parent_id, child_id, facet1_id, facet2_id, ...facetN_id]

Since numpy docs state 2D arrays must be rectangular, not jagged, I use np.resize to set the length of every row to the row with the highest number of facets and np.ma.array to mask the padded arrays. Another solution is padding arrays with sentinel values, but masking is more correct IMO, and required for the next step.

Following that, I mask the parent_id and child_id and perform masking operations to determine matching rows with AND/OR facet ids. e.g. I want to find all parent_id/child_id rows that contain facets (A OR B) AND (C). I do not want to discard the values, and I do not want them part of the comparison. I can do this via the following:

facet_groups = [[A, B], [C]]
filtered = arr[np.logical_and.reduce([np.isin(arr, facet_ids).any(axis=1) for facet_ids in facet_groups])]

Now I can remove the mask and extract all the parent_id/child_id values with a slice:

filtered.mask = np.ma.nomask
parent_ids = filtered[:,0]
child_ids = filtered[:,1]

In addition, I use np.unique on the masked array to calculate the total counts for each remaining facet.

All in all, it's a very concise bit of code that performs very well for the small dataset of a few hundred thousand values.

Yes, I highly agree with @stuarta0, there needs to be a masked_fill feature in ndarray.

Currently, you would do so on a nth dimensional array using loops, where if a specific value is mask, then replacing it with z.

Yes, I highly agree with @stuarta0, there needs to be a masked_fill feature in ndarray.

Currently, you would do so on a nth dimensional array using loops, where if a specific value is mask, then replacing it with z.

At least this is what I do, there most definitely is a simpler version out somewhere.