[jira] [Created] (KAFKA-13994) Incorrect quota calculation due to a bug

["Divij Vaidya (Jira)" <ji...@apache.org>]

Divij Vaidya created KAFKA-13994:
------------------------------------

             Summary: Incorrect quota calculation due to a bug
                 Key: KAFKA-13994
                 URL: https://issues.apache.org/jira/browse/KAFKA-13994
             Project: Kafka
          Issue Type: Bug
          Components: core
            Reporter: Divij Vaidya


*Problem*
This was noted by [~tombentley] at [https://github.com/apache/kafka/pull/12045#discussion_r895592286] 

The completion of a sample window in `SampledStat.java` is based on comparison of `

recordingTimeMs` with startTimeOfPreviousWindow [1]. `recordingTimeMs` is calculated as a System.currentTimeMillis which:
1. is not guaranteed to be monotonically increasing due to clock drifts. 
2. is not necessarily the current time when it arrives at [1] because the thread may be blocked at `synchronized` at {{{}Sensor.recordInternal [2]{}}}, because synchronized provides no guarantee about fairness for blocked threads.

Hence, it is possible that when isComplete comparison is made at [1], recordingTimeMs < endTimeOfCurrentWindow whereas the wallClockTimeAtTheMoment >  startTimeOfCurrentWindow + window length.

The implication of this would be:
1. The current sample window will not be considered completed even if it has completed as per wall clock time.
2. The value will be recorded in a sample window which has elapsed instead of a new window where it belongs.

Due to the above two implications, the metrics captured by the sensor may not be correct which could lead to incorrect quota calculations.

 [1] [https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/metrics/stats/SampledStat.java#L138] 

 [2] https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/common/metrics/Sensor.java#L232



--
This message was sent by Atlassian Jira
(v8.20.7#820007)