[exporter/signalfx] Translate metrics together instead of individually

[crobert-1]

Description:

Translation rules were originally done on metrics individually. This means that only datapoints within the same metric could be operated on together. This is insufficient for the new host metrics format removing the direction dimension. To properly form metrics to match their previous format, translations must be done across metrics, especially aggregations. This only changes the fact that multiple MTS will be operated on at the same time, it doesn't change any translation rules. The solution is to do aggregations at the ScopeMetric metric level, so metrics within the same ScopeMetric can be aggregated.

This change also properly changes test metric data to be their proper format, moving metrics into corresponding ResourceMetrics and ScopeMetrics, instead of all being lumped into the same slices.

This is an internal refactoring and testing change that should not impact any input or output data. Because of this, I don't think there should be any CHANGELOG entry.

Link to tracking Issue:

This PR is the first step to solving issue #12641

Testing:

Ensured existing tests pass, data being emitted is the same as before change.

Documentation:

NA since this is a refactor without any external impact.

[dmitryax]

@crobert-1 PTAL at the linter failures

[crobert-1]

It looks like tests are setting up resource metrics in unsupported formats, so that aggregation isn't working as expected. The extra checks would take care of this, but assuming proper metric format causes these to fail.

I'll update the test format (resource metrics should all have the same time stamp, rather than scope metrics within resource metrics having different timestamps) unless anyone disagrees with my findings.

[dmitryax]

Where is the actual change now? :)

[crobert-1]

Where is the actual change now? :)

The change at this point is just updating test data format. It could go with another PR that will update translation rules, but I figure it would be good to have the format change attached to the discussion we've had on this PR.

[dmitryax]

but we still apply translateAndFilter on data points from one pdata.Metric, the idea is to apply it on set of data points from one pdata.ScopeMetrics object

[crobert-1]

but we still apply translateAndFilter on data points from one pdata.Metric, the idea is to apply it on set of data points from one pdata.ScopeMetrics object

I don't know why, but I got mixed around with this review. You're right, we want to apply it at the ScopeMetrics level. For some reason I got confused and thought the inner-most loop was the scope level, rather than individual metric level.

I'll update the description again and move translateAndFilter call.

[crobert-1]

I've added a benchmark test that takes hostmetrics and times how long it takes to run MetricsToSignalFxV2. The only change to this functionality was the scope of the translation loop.

Here's the performance comparison before and after this change on the same dataset:
Before:

BenchmarkMetricConversion-12       	    2332	    504435 ns/op
BenchmarkMetricConversion-12       	    2166	    558897 ns/op
BenchmarkMetricConversion-12       	    1760	    637230 ns/op

After:

BenchmarkMetricConversion-12       	    2235	    488777 ns/op
BenchmarkMetricConversion-12       	    2451	    488565 ns/op
BenchmarkMetricConversion-12       	    2424	    514079 ns/op

I believe this shows there is a slight performance improvement for the signalfx exporter with the new functionality.

[dmitryax]

@crobert-1 thanks for adding the benchmark 👍 Can you please post before and after results?

[dmitryax]

Nice! Seems like there is a clear improvement. I believe we can merge it regardless of duration attribute's changes