Exemplars for Exponential LongHistogram always have value 0.

[fstab]

I realized this while working on unit tests for #5940:

In order to switch to exponential histograms, you register a View with aggregation

Aggregation.base2ExponentialBucketHistogram()

However, under the hood there's only one implementation: DoubleBase2ExponentialHistogramAggregator. This is used for both DoubleHistogram and LongHistogram.

As a result, if you have a LongHistogram the Exemplars always have value 0, because the aggregation always uses ReservoirCell.doubleValue and never ReservoirCell.longValue.

It looks like a LongBase2ExponentialHistogramAggregator is missing.

[jack-berg]

Good catch! I just looked at the source code and it looks like the explicit bucket histogram aggregation doesn't suffer from this problem because it takes advantage of the exemplar reservoir being HistogramExemplarReservoir, a purpose built reservoir which knows that only double histograms are exported and casts the long measurement value to a double before recording in the reservoir.

Something similar is needed for exponential histograms, but even the explicit bucket histogram solution is brittle because it would be subject to the same failure mode once we make exemplar reservoirs configurable via views. We really need both explicit and exponential bucket histograms to always cast measurements to doubles before trying to record in the exemplar reservoir.