V5 — fully robust, no transpilation, native ESM

[mourner]

• [x] Switch the package to "type": "module" and native ESM, requiring Node v12+.
• [x] Use robust orientation check from robust-predicates to solve #61, add more robustness test cases.
• [x] Remove buble transpilation and drop firsthand support for IE11.

The robust-predicates change makes the bundle ~32% bigger — from 2286 to 3029 bytes minified/gzipped. Previously I thought I should expose a separate "robust" bundle in addition to the smaller one, but now it seems that at a scale of few kilobytes, it's not worth the added complexity and the increase isn't a big deal.

Adding robust incircle would increase it more significantly (to 5322 bytes), but it doesn't seem necessary since this check only affects which triangles flip, and doesn't affect the integrity of the triangulation unlike orient2d.

@mbostock @Fil what do you think?

[Fil]

On principle I agree that we still too often have these precision issues, in particular in generative art code where points will often be aligned (vs "actual" data where Delaunator never fails). So I'm all for the extra cost if it can solve these.

I wanted to build d3-delaunay with this branch, however I am so bad with javascript plumbing, and failed.

[redblobgames]

I was wondering how much the robust check would affect performance. I ran the bench test but my machine shows quite a bit of variance from one run to the next, and the slowdown below 1,000,000 points is within the margin of error (+/- 10ms). At 1,000,000 points you can see a slight slowdown:

v4, 1,000,000 points, 4 runs

test	run 1	run 2	run 3	run 4
uniform	658.504ms	653.551ms	649.729ms	654.352ms
gaussian	640.57ms	644.188ms	648.516ms	641.86ms
grid	531.747ms	538.196ms	538.529ms	536.82ms
degenerate	239.893ms	235.587ms	241.171ms	229.518ms

v5, 1,000,000 points, 4 runs

test	run 1	run 2	run 3	run 4
uniform	679.483ms	674.571ms	666.818ms	660.141ms
gaussian	661.089ms	665.268ms	662.053ms	664.995ms
grid	546.916ms	555.071ms	555.103ms	553.64ms
degenerate	248.789ms	245.65ms	241.847ms	243.951ms

I'm surprisingly bad at benchmarking though so I wouldn't be surprised if I did something wrong.

[mourner]

@Fil didn't expect any problems with integrating this into d3-delaunay, can look at it next week. Of if you have any more details, I'd be happy to help.

@redblobgames that looks similar to what I've been getting and I think the results are pretty reasonable, even good considering the change.

[Fil]

Nah no real problem, it's just that I don't know how to do it. What I tried was to replace the delaunator folder in npm_modules! probably not the right method :-)

[mourner]

@Fil you can just do npm install mapbox/delaunator#v5 and it will install the branch there.

[Fil]

Thanks! With this it builds*, but alas it doesn't seem to solve this half-baked experiment https://observablehq.com/d/62b5c48194cba6b2 ; although the errors are different!

* build with yarn rollup -c — as far as I understand, the build is correct, but the unit tests fail:

❯ yarn test
yarn run v1.22.5
$ tape -r esm 'test/**/*-test.js' && eslint src test
/Users/fil/Sites/d3/d3-delaunay/node_modules/delaunator/index.js:1
Error [ERR_REQUIRE_ESM]: Must use import to load ES Module: /Users/fil/Sites/d3/d3-delaunay/node_modules/delaunator/index.js
    at new NodeError (node:internal/errors:278:15)
    at Object.Module._extensions..js (node:internal/modules/cjs/loader:1125:13) {
  code: 'ERR_REQUIRE_ESM'
}

[mourner]

@Fil great test case! At least the Delaunay triangulation on it visually looks correct, so there might be an issue hidden in the Voronoi code. Although Delaunay might be broken in a non-visual way (bad connectivity) — will check this one next week.

[mourner]

but the unit tests fail

Yes, this will need adjustments — in short:

• remove the -r esm flag because that dep is incompatible with native ESM
• add "type": "module" so that .js files are treated as modules
• replace tape <glob> with individual node <test-file> because of https://github.com/substack/tape/issues/514 (or you could use uvu instead of tape)

[mourner]

@Fil just to make sure, I added a sample test case from your notebook to the test suite and it passes robustness tests (triangulation is valid, with no broken halfedges etc.), so the issue must be on the d3-delaunay side.

[Fil]

Sorry for the delay. (And thank you for the "adjustments" in tests!). I can now confirm that the issue lies with d3-delaunay, in the special case it does for circumcenters of degenerate triangles in https://github.com/d3/d3-delaunay/blob/master/src/voronoi.js#L43

I can fix it for this example, but then it breaks a few other tests which are themselves a bit dubious. More later, but please don't hesitate to merge this.

[jrus]

This is great!