Tune Ruby GC sampler histogram buckets
What does this MR do?
Our current Ruby GC time distribution is as follows:
Element | Value |
---|---|
{le="+Inf"} |
7358827 |
{le="0.5"} |
7355490 |
{le="0.1"} |
1935066 |
{le="0.07"} |
446797 |
{le="0.04"} |
40495 |
{le="0.02"} |
27901 |
{le="0.01"} |
26518 |
{le="0.005"} |
23732 |
It's easy to see that our buckets to the upside of 100ms are too coarse, since they account for ~90% of the data. There is also a flatter tail to the downside of 40ms, but I would argue that this is not very interesting; for instance the 5ms bucket isn't useful since those GC pauses are so fast, they do not need to concern us. It's more interesting to break up the slow buckets.
Some data points from Thanos supporting the above
Does this MR meet the acceptance criteria?
Conformity
Edited by Matthias Käppler