Reduce memory consumption

We observed increased memory usage in our microservices, after applying micrometer. Heap dumps showed that a lot of memory is consumed by histograms stored in our timers. Following is an extract from such dump:

io.micrometer.core.instrument.step.StepTimer@0xf97e62a0
- histogram: TimeWindowLatencyHistogram@0xf97e6398
  - accumulatedHistogram: Histogram@0xf982e7a0
    - counts: [long]@0xf982e9f0 (10256 bytes) 
  - ringBuffer: [LatencyStats]
    0: LatencyStats@0xf9821a50
      - activePauseCorrectionsHistogram: Histogram@0xf98299f0
        - counts: [long]@0xf9829c40 (10256 bytes) 
      - activeRecordingHistogram: AtomicHistogram@0xf9826f70
        - counts: AtomicHistogram@0xf98271d0
          - array: [long]@0xf98271e0 (10256 bytes)
      - inactivePauseCorrectionsHistogram: Histogram@0xf9824510
        - counts: [long]@0xf9824760 (10256 bytes)
      - inactiveRawDataHistogram: AtomicHistogram@0xf9821a90
        - counts: AtomicHistogram@0xf9821cf0
          - array: [long]@0xf9821d00 (10256 bytes)
      - intervalEstimator: TimeCappedMovingAverageIntervalEstimator@0xf982c4c8
        - intervalEndTimes: [long]@0xf982c518 (8208 bytes) 

And there is 4 more LatencyStats objects in that ringBuffer array.

Summing only long arrays: 10256 + (10256 * 4) * 5 = 215376 bytes Total for all 80 timers that we have: 215376 * 80 = 16.4 mb

We don’t use histograms and percentiles, so we’ve overcame the problem with custom implementation of the Timer interface. But it would be great to have a better solution straight in the Micrometer.

[jkschneider]

After decreasing buffer length to 3, adding a new configuration option for digits of precision used in computing percentile approximations, and adding new histogram implementations that are more memory efficient when you don’t need client-side percentiles, we have this result.

1 digit of precision seems to be satisfactory for percentile approximations ±5% of true value. 3 digits of precision yields ±0.1%, but at a much higher memory cost.

[pooyas]

@jkschneider thanks!