prometheus / client_ruby / 519

84%

master: 100%

Pull #161

travis-ci

Fix the "subtle" bug introduced in the previous commit

So, again, I don't propose we actually do this, this is just illustrative
of what we need to do, and didn't want it to get lost.

The problem with the previous commit is that for each metric, we open a
new FileMappedDict. All of them sharing the same file is "fine", actually.
The problem comes from the fact that each instance of FileMappedDict has
their own version of `@used`. When one metric adds a new entry to the file,
it moves its own `@used` pointed, but the other instances don't notice this.

As soon as a second instance that had already been loaded adds a new labelset,
it corrupts the file.

This only shows up when running the benchmarks, with histograms. Counters
(in our benchmark) don't reproduce it, because each metric in our benchmark
only has one labelset. They all get added initially to the file, and an
already loded Dict doesn't add new ones. The reason Histograms reproduce it
is that new observations may report a bucket that wasn't seen before, which
is effectively a new labelset.

The solution is for all MetricStore instances to share one single FileMappedDict.
This is not particularly hard, but the current MetricStore is not written
in a way that allows that without doing this crap.

We should find a better way. I'm just leaving this here as documentation
of this problem, for the next brave soul attempting this

53 of 53 new or added lines in 2 files covered. (100.0%)

3855 of 4587 relevant lines covered (84.04%)

33.18 hits per line