[HADOOP-13804] MutableStat mean loses accuracy if add(long, long) is used

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 2.6.5
• Fix Version/s: 2.8.0, 2.7.4, 3.0.0-alpha2
• Component/s: metrics
• Target Version/s: 2.7.4
• Hadoop Flags: Reviewed
• Assignee: Erik Krogen
• Reporter: Erik Krogen
• Created: 07/Nov/16 22:49
• Updated: 08/Nov/16 00:40
• Resolved: 08/Nov/16 00:23

Description

Currently if the MutableStat.add(long numSamples, long sum) method is used with a large sample count, the mean that is returned will be very inaccurate. This is a result of using the Welford method for variance calculation, which assumes that each sample is processed on its own, to calculate the mean as well. For variance this is fine, since variance numbers lose meaning if you add many samples at once, but the mean should still be accurate.

Issue Links

• blocks: HADOOP-13782 Make MutableRates metrics thread-local write, aggregate-on-read