You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
While analysing real-world production data, I had an insight:
If we hit the configured NativeHistogramMaxBucketNumberbefore the configured NativeHistogramMinResetDuration has passed since the last reset, we'll reduce the resolution instead. So far so good.
However, once we have reduced the resolution, it is much less likely to ever hit NativeHistogramMaxBucketNumber again. Therefore, we might very well stay at the reduced resolution forever (i.e. for the lifetime of the binary). It would be better to reset the histogram immediately after NativeHistogramMinResetDuration has passed, because this will reinstate the originally configured resolution.
More context: In the real-world production scenario I have analysed, it is indeed a very rare even that the resolution has to be reduced, mostly triggered by short peaks of more intense traffic. However, because of the reason explained above, it essentially means that a single occurrence will cause the affected binary to expose a reduced-resolution histogram for the rest of its lifetime. Since histograms are often aggregated together, this needlessly reduced the resolution of the aggregated results with quite some blast radius.
The text was updated successfully, but these errors were encountered:
While analysing real-world production data, I had an insight:
If we hit the configured
NativeHistogramMaxBucketNumber
before the configuredNativeHistogramMinResetDuration
has passed since the last reset, we'll reduce the resolution instead. So far so good.However, once we have reduced the resolution, it is much less likely to ever hit
NativeHistogramMaxBucketNumber
again. Therefore, we might very well stay at the reduced resolution forever (i.e. for the lifetime of the binary). It would be better to reset the histogram immediately afterNativeHistogramMinResetDuration
has passed, because this will reinstate the originally configured resolution.More context: In the real-world production scenario I have analysed, it is indeed a very rare even that the resolution has to be reduced, mostly triggered by short peaks of more intense traffic. However, because of the reason explained above, it essentially means that a single occurrence will cause the affected binary to expose a reduced-resolution histogram for the rest of its lifetime. Since histograms are often aggregated together, this needlessly reduced the resolution of the aggregated results with quite some blast radius.
The text was updated successfully, but these errors were encountered: