remove cached trace._interpz from heatmaps/contour maps with gaps

[alexcjohnson]

Fixes #2652 - turns out this isn't a react-specific issue, it's just about changing z data for contours with gaps in the data (and heatmaps with connectgaps: true). We were caching the interpolation results, even across recalc calls, because this can be a pretty slow (iterative) process. I removed that cache, after first trying to figure out whether we could salvage it, based on comparing the incoming known (non-interpolated) z values to the old ones. It turned out that was difficult (and expensive) to get right, especially for edge cases like removing a previously filled-in value without changing anything else.

We're doing a better job than we used to do (3+ years ago when this cache was added) at avoiding unnecessary recalcs, so I don't think this should be too big a cost, but this is more motivation to push in that direction, especially things like recalc'ing only the traces that need it.

cc @etpinard

[etpinard]

but this is more motivation to push in that direction, especially things like recalc'ing only the traces that need it.

Good call, just to list a few things high on my priority list:

• make autorange edits not go through editType: 'calc' (more in Faster axis autorange relayout #2546)
• make trace visible edits not go through editType: 'calc' either

[etpinard]

Nice fix and tests 💃

Would you mind sharing some benchmarks for interp2d? Mostly to document how much of a slow down users can expect in the next release in case I don't have time to resolve #2546 by then.

[etpinard]

Amazingly ✨

[etpinard]

thanks for this comment 📚

[alexcjohnson]

Benchmarks - compare a full heatmap, which plots in around 70ms for me at n=100 (100x100 heatmap):

function hmFull(n) {
  var z = new Array(n); for(var i = 0; i < n; i++) { z[i] = new Array(n); for(var j = 0; j < n; j++) z[i][j] = Math.random(); }
  var t0 = performance.now();
  Plotly.newPlot(gd, [{z: z, type: 'heatmap', showscale: false}], {width: 500, height: 500})
  return performance.now() - t0;
}

to a gapped one:

function hmPartial(n) {
  var z = new Array(n); for(var i = 0; i < n; i++) {
    z[i] = new Array(n);
    for(var j = 0; j < n; j++) z[i][j] = null; z[i][Math.floor(Math.random() * n)] = Math.random();
  }
  var t0 = performance.now();
  Plotly.newPlot(gd, [{z: z, type: 'heatmap', showscale: false, connectgaps: true}], {width: 500, height: 500})
  return performance.now() - t0;
}

A fairly full heatmap (n=100, m=50, so about 50% gaps) spends about 40ms on interp2d.
A fairly empty one (n=100, m=5, so about 95% gaps) spends about 450ms on interp2d. Only twice as many gap points to fill in, but it iterates much more slowly.