Details
-
Improvement
-
Resolution: Fixed
-
Major
-
3.1.0, 4.6.4, 5.0.0, 5.0.1
Description
The histogram bucketing for the disk_commit and storage_age histograms are incorrectly defined so all samples were frequently placed in a single bucket:
storage_age (65700 total)
|
0 - 1s : ( 98.35%) 64615 ####################################
|
1s - 2s : (100.00%) 1085
|
Avg : ( 16ms)
|
|
|
disk_commit (6273 total)
|
0 - 1s : (100.00%) 6273 ######################################
|
Avg : ( 0)
|
This can also be seen in showfast / cbmonitor graphs - note how avg_disk_commit_time is constantly 0.5s:
Also visible in the UI:
This is caused by the histograms being constructed with a poor choice of bucket sizes:
dirtyAgeHisto(GrowingWidthGenerator<UnsignedMicroseconds,
|
cb::duration_limits>(
|
ONE_SECOND.zero(), ONE_SECOND, 1.4),
|
25),
|
diskCommitHisto(GrowingWidthGenerator<UnsignedMicroseconds,
|
cb::duration_limits>(
|
ONE_SECOND.zero(), ONE_SECOND, 1.4),
|
25),
|
The same is seen in the UI - you'll typically never see better number than 0.5s.
In other words, starting with a bucket at (0,1.0]s, then (1.0, 2.4], (2.4, 4.36], ... Given that disk_commit times are typically in the order of microseconds; these arn't particulary suitable buckets.
