Auto mark tracking issue

[mbostock]

Fast follows from #1238:

Bugs

• Auto mark: no display with non-zero reducer and bar mark #1340
• The auto mark should not default to bars when the reduce is not sum or count (e.g., don’t stack means) #1272
  • only default to bar for x or y is zeroed #1273

Robustness — aspiring to produce good plots for any combination of options

• Assigning ordinal dimension to size results in an empty Plot. We should handle this as a "very filtered data" warning, plus propagating warnings more clearly to the frontend.
  • Communicate information about filtered data points #493
  • Expose warnings #1192
  • Sketch PR: Auto: error on reducer without input; warn on ordinal size; expose warnings #1424
• Handle high-cardinality ordinal domains better, e.g. with limit and sort plus an "Other" category so we don't silently truncate.
  • A transform to consolidate ordinal values outside the top n into an “other” category, perhaps in conjunction with the group transform. #144
  • https://observablehq.com/d/68e4c3583be08872
  • Also think about high-cardinality ordinal color swatches
• Handle non-ordinal facets better
• Detect 1:1 domains suitable for bars https://observablehq.com/@tophtucker/plot-auto-bijection-test
• Warn or error if non-count reducer is set without a field (very common in user testing)

Options — making more configuration possible

• Pass through other channel options, e.g. thresholds for binning
• Pass through other mark options, e.g. strokeWidth Auto mark passes through other options to mark #1261
• How to specify z for multi-series line or area marks?

Other enhancements

• Shorthand for arrays of primitives (e.g., Plot.auto(numbers).plot() gives a histogram)
• Better default (non-error) state when no options are given (e.g., Plot.auto(aapl).plot())
• Eject to explicit marks Sketch: compile auto mark to explicit marks #1373
• Auto mark: add more override-only mark types? #1422

Docs

• Document the zero option only default to bar for x or y is zeroed #1273 (review)

[tophtucker]

Maybe lower-priority, but not to be forgotten:

• "Shorthand" (make it work for passing arrays of primitives)
• Better default unconfigured state for Plot.auto(aapl).plot() than an error (notes)

[tophtucker]

It’s a little hard to say why, but this should be lines, not rects:

Is it because date_of_birth is continuous, albeit not monotonic, and reduce is something other than count? Implicitly stacking means feels wrong.

[tophtucker]

An explicit size reducer should stop us from setting an implicit y reducer:

[mbostock]

reduce is something other than count

Yes, we should only implicitly stack when the reduce is count, not when it’s something else. You can always add counts but adding means/medians/etc. often doesn’t make sense.

I’ve added your TODO’s to the list above, along with this one.

[Fil]

sum is also additive :-) I'm adding it to the TODO.

[tophtucker]

Setting size to an ordinal dimension shows nothing. That makes sense, since Plot.dot(data, {x: "culmen_depth_mm", r: "sex"}).plot() shows nothing either, and I'm not sure what it'd mean… maybe size by the value's index in the domain? But it feels broken when I'm trying every combination and every other one works. Could also throw an error.

[mbostock]

Setting size to an ordinal dimension shows nothing.

Related #493; the data here is being filtered because the size channel implicitly coerces to numbers producing NaN.

[tophtucker]

Should we throw an error when a column isn't found in the data?

Two notable cases. You might've made a typo (as described in Robert's Levenshtein distance comment #493 (comment), though imo that's overkill):

Plot.auto(aapl, {x: "close"}).plot()

Or you might've changed the data source, invalidating your column selections:

Plot.auto(olympians, {x: "Close"}).plot()

Other marks don't do this, so I imagine you've considered it and decided against it, but it might fit with Plot.auto's more aggressive and opinionated mandate.

[mbostock]

Should we throw an error when a column isn't found in the data? … Other marks don't do this

I think we plan on doing this eventually for all marks as part of #493 (though maybe as a warning rather than an error). I’d like to minimize the “unnecessary” differences between Plot.auto and concrete mark types, where it makes sense; detecting whether the column you referenced exists feels like something that could apply to all marks. I don’t see it as part of the “opinionated defaults” that is the raison d’être for Plot.auto.

[tophtucker]

"Error: implicit ordinal position domain has more than 10,000 values" is the most common (only?) error I've been seeing, which makes me itch for limit/sort. Maybe the channel options object could have a limit property; undefined would default to ~20; if null, it wouldn't limit; if it limits, it always sorts? {x: {value: "nationality", limit: 100}}