fix(stdlib): sort statuses also by _time

[alespour]

This PR fixes issue in alerting where, depending on check and rule scheduling and timing configuration, state changes are sometimes not detected and therefore notifications not sent.

The problem is that status of data points may be evaluated multiple times with certain scheduling, for example as reported in #3807, deadman check evaluates status based on data point time and not a metric value. That results in existence of multiple statuses with the same source timestamp but different time (and level) in the monitoring bucket. Then, when an alert rule attempts to detect state change, it gets incorrect sequence of statuses, because sorting by only (the same) source timestamp is insufficient.

Affected functions are stateChangesOnly and _stateChanges functions in monitor package, where before sort,
_level is dropped from the group key which results in a new series with possibly unpredictable order (by _time).

The issue is a side effect of the changes in #2725. Mea culpa.

Done checklist

• Test cases written

[wolffcm]

The fix looks good and the changes make sense. Thanks for fixing this!

Can I trouble you to adapt your test to our newer format which uses a Flux keyword testcase? We want to be using it moving forward. Here is an example:

flux/stdlib/universe/window_test.flux

Line 7 in 955bc03

testcase window_period_gaps {

[alespour]

@wolffcm Done. Nice feature!

[wolffcm]

Look great. Thanks again!

[wolffcm]

@alespour Unfortunately we had to revert this commit because the new test doesn't pass in the OSS version of influxdb---these tests (like the older ones) will use testing.load() to insert data into InfluxDB and then select it back.

The culprit seems to be that the data being passed to testing.load() is pivoted. There are a couple ways to solve this:

• Remove testing.load() and just run the test against in-memory data. This is fine because you're not testing push-down logic or planner rules or anything like that.
• Stream unpivoted data into testing.load().

The first option is the easiest and that's what I recommend. It should be as simple as just deleting the |> testing.load() line from your test.

Unfortunately, I'm not sure of the right steps to check if a new Flux test like this will pass in influxdb before actually merging to master. We need to look into that.

[jsternberg]

Fixed using @wolffcm's first suggestion in #4012.