@mattpap - gentle ping. Have you had a chance to look into the TypeScript compile issue with this PR?

Notes from today's call with @mattpap

Todo in this PR:

• Add licenses for code copied and modified from node-csv (it has an MIT license)
• Fix eslint issues
• Check how the CSV looks on the color scatter example

Leaving unsolved (or for a future PR):

• How to represent 2d arrays (and higher dimensions) in CSV

My latest code changes are a bit sloppy. Basically I added a function ndget which gets called when the column data source gets transformed into rows. This is to fix what felt like a broken CSV for the color_scatter example.

The color scatter example has 4000 circles, each with a color that is a function of the (x, y) coordinates of the circle's center point. When this example gets serialized and deserialized in the browser, the underlying column data source has a column called "fill_color" which contains a 12,000 long Uint8Array. Every three items in that array represent the red, green, and blue values for the color. So the problem is that (before my latest code changes), when this was serialized to CSV, the fill_color column in the CSV only gets the first 4,000 values of those 12,000. So I created a function that gets values from a flat-representation of an ndarray taking into account the shape of the array.

I also used the existing pretty printer class to stringify object types rather than use JSON.stringify.

Together this results in the following first two lines of CSV for the color_scatter example:

radius,x,y,fill_color
1.4252641788706195,41.7022004702574,32.664490177209615,"[Uint8Array([133, 95, 150])]"

This looks reasonable to me.

So the code changes are a bit sloppy, but I want to know if this is at least in the right direction.

My sense is that there really should be some way to index the NDArray class so that I do not have to create a code fork, i.e.:

if (is_NDArray(column)) {
  row[c] = ndget(column, r)
} else {
  row[c] = column[r]
}

But I don't have a good idea of how to go about that.

I want to add a few more test cases for the to_csv() method - specifically where the underlying test data source object has columns that are NDArrays, but otherwise this is ready for review so I am going to take it out of draft mode.

I believe @mattpap said that the following failing test is irrelevant to this PR.

✗ Bug
 └┬─ in issue #12718
  └── doesn't render Legend correctly when LegendItem.index is filtered out by CDSView
diff --git a/bokehjs/test/baselines/linux/Bug__in_issue_#12718__doesn't_render_Legend_correctly_when_LegendItem.index_is_filtered_out_by_CDSView.blf b/bokehjs/test/baselines/linux/Bug__in_issue_#12718__doesn't_render_Legend_correctly_when_LegendItem.index_is_filtered_out_by_CDSView.blf
index dbd519a..de67416 100644
--- a/bokehjs/test/baselines/linux/Bug__in_issue_#12718__doesn't_render_Legend_correctly_when_LegendItem.index_is_filtered_out_by_CDSView.blf
+++ b/bokehjs/test/baselines/linux/Bug__in_issue_#12718__doesn't_render_Legend_correctly_when_LegendItem.index_is_filtered_out_by_CDSView.blf
@@ -5,4 +5,4 @@ Figure bbox=[0, 0, 200, 200]
     Scatter bbox=[37, 92, 150, 0]
   LinearAxis bbox=[29, 178, 166, 22]
   LinearAxis bbox=[0, 5, 29, 173]
-  Legend bbox=[31, 15, 163, 42]
+  Legend bbox=[55, 15, 115, 42]

But I don't know about the following failing test:

✗ Plot
 └┬─ should support windowed auto-ranging
  └── with window_axis='y' when data changes
images differ by 4px (0.00%)