1438953 - Timer Precision Tests need a more generous float fuzziness value

Status: RESOLVED FIXED

(Core :: DOM: Events, enhancement, P2)

Description

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

Currently in the resist fingerprinting/reduce timer precision tests we have an allowance of ~.0000001 (milliseconds) to account for JAVASCRIPT (not the C++ calculations) fuzziness in float calculations.

When we are working with milliseconds since the epoch, a double's precision is around .0002; so we need to up our epsilon, or face the possibility of intermittents.

Comment 1

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

Attachment: Bug 1438953 Increase the epsilon on timer reduction tests

In Javascript, we re-clamp timers to ensure they stay the same. Because of double imprecision
sometimes they don't stay the same, and are clamped downwards. If that happens we detect it,
and if we were originally off by an epsilon from a clamped value, we accept the value in the
name of double imprecision.

However, the epsilons were originally chosen somewhat arbitrarily. They worked for small
numbers, where doubles can represent very precise decimals. They don't work for large numbers
where doubles cannot represent such small decimals.

Review commit: https://reviewboard.mozilla.org/r/220996/diff/#index_header
See other reviews: https://reviewboard.mozilla.org/r/220996/

Comment 2

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

See https://bugzilla.mozilla.org/show_bug.cgi?id=1430841#c42 for something that should probably be addressed here.

Comment 3

[Nathan Froyd [:froydnj]]

Comment on attachment 8951713 [details]
Bug 1438953 Increase the epsilon on timer reduction tests

https://reviewboard.mozilla.org/r/220996/#review227238

This seems reasonable.  Only shame is that we can't (or haven't yet) pull this out into common code.

::: commit-message-f191f:9
(Diff revision 1)
> +numbers, where doubles can represent very precise decimals. They don't work for large numbers
> +where doubles cannot represent such small decimals.

This seems like an awkward way to express the idea that large doubles can't have precise fractional parts.  WDYT about rewording this?

Comment 4

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

Comment on attachment 8951713 [details]
Bug 1438953 Increase the epsilon on timer reduction tests

Review request updated; see interdiff: https://reviewboard.mozilla.org/r/220996/diff/1-2/

Comment 5

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5db4200c90a8
Increase the epsilon on timer reduction tests r=froydnj

Comment 6

[Andreea Pavel [:apavel]]

Backed out for failing mochitest at browser/components/resistfingerprinting/test/mochitest/test_animation_api.html

Push that contains the failure: https://treeherder.mozilla.org/#/jobs?repo=autoland&revision=d32d276d2917d39471d8aa45192436f8819c5df0

failure log: https://treeherder.mozilla.org/#/jobs?repo=autoland&revision=d32d276d2917d39471d8aa45192436f8819c5df0

Backout: https://hg.mozilla.org/integration/autoland/rev/ad3678a2943b860a240de123dc57520cf622572c

Comment 7

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

Comment on attachment 8951713 [details]
Bug 1438953 Increase the epsilon on timer reduction tests

Review request updated; see interdiff: https://reviewboard.mozilla.org/r/220996/diff/2-3/

Comment 8

[Tom Ritter [:tjr] (OOTO until 4/30 at least)]

Okay, fixed the false positive case!

Comment 9

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/5eeff2b96802
Increase the epsilon on timer reduction tests r=froydnj

Comment 10

[Raul Gurzau (:RaulG)]

https://hg.mozilla.org/mozilla-central/rev/5eeff2b96802