log scale with zero data should give a better warning

[jowens]

In Altair, I had a column of data that I plotted on a log scale:

sh = Chart(df).mark_point().encode(
    x='Launch:T',
    y=Y('Pixel/unified shader count:Q',
        scale=Scale(type='log'),
        ),

and a zero snuck into the column that I didn't know about and the debugging of that took a really really long time. Here's what Altair output looked like:

I guess I would have expected that the zeroes would be silently dropped OR the console would have thrown a warning ... anyway, the Altair output was perplexing. One zero out of several hundred data points made everything plot on a single line.

@jakevdp suggested this was the right place to raise the issue (vega/altair#953).

[domoritz]

Vega-Lite actually never sees the data so the warning would need to be in Vega.

[jowens]

@domoritz should I go file this in vega's issues?

[domoritz]

One thing we could do in Vega-Lite is automatically filter out values less or equal to zero (in https://github.com/vega/vega-lite/blob/master/src/compile/data/filterinvalid.ts) but that seems like it could have all kinds of bad effects.

Yes, I think this needs to be in Vega itself.

[kanitw]

One thing we could do in Vega-Lite is automatically filter out values less or equal to zero (in /src/compile/data/filterinvalid.ts@master) but that seems like it could have all kinds of bad effects.

We used to, but not any more. (Look at the code you link to.)

[domoritz]

What do you mean? I know that we used to and we don't right now but I linked to the code where we would add this functionality.

[g3o2]

wouldn't it be better to plot the invalid values as missing values?

[domoritz]

@g3o2 yes. We haven't figured out a good way to show missing values in all chart types.

[kanitw]

What do you mean? I know that we used to and we don't right now but I linked to the code where we would add this functionality.

Oh I see. We intentionally remove that filter because it's magical so I don't think we should add it back. Plus, D3 log scale works with either strictly positive or strictly negative values, so it's quite tricky to write the right filter.

wouldn't it be better to plot the invalid values as missing values?

Yes, but invalid values is also a part of scale domain and thus it will break the scale like shown in the screenshot at the top of this issue. Filtering it correctly can be tricky.