[ https://issues.apache.org/jira/browse/HIVE-21823?focusedWorklogId=254240&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-254240 ]
ASF GitHub Bot logged work on HIVE-21823: ----------------------------------------- Author: ASF GitHub Bot Created on: 05/Jun/19 07:59 Start Date: 05/Jun/19 07:59 Worklog Time Spent: 10m Work Description: pvary commented on pull request #660: HIVE-21823: New metrics to get the average queue length / free executor number for a given time window URL: https://github.com/apache/hive/pull/660#discussion_r290619644 ########## File path: llap-server/src/java/org/apache/hadoop/hive/llap/metrics/LlapDaemonExecutorMetrics.java ########## @@ -397,4 +425,106 @@ public JvmMetrics getJvmMetrics() { public String getName() { return name; } + + /** + * Generate time aware average for data points. + * For example if we have 3s when the queue size is 1, and 1s when the queue size is 2 then the + * calculated average should be (3*1+1*2)/4 = 1.25. + */ + @VisibleForTesting + static class TimedAverageMetrics { + private final int windowDataSize; + private final long windowTimeSize; + private final Data[] data; + private int nextPos = 0; + + /** + * Creates and initializes the metrics object. + * @param windowDataSize The maximum number of samples stored + * @param windowTimeSize The time window used to generate the average in nanoseconds + */ + TimedAverageMetrics(int windowDataSize, long windowTimeSize) { + this(windowDataSize, windowTimeSize, System.nanoTime()); + } + + @VisibleForTesting + TimedAverageMetrics(int windowDataSize, long windowTimeSize, + long defaultTime) { + assert(windowDataSize > 0); + this.windowDataSize = windowDataSize; + this.windowTimeSize = windowTimeSize; + this.data = new Data[windowDataSize]; + Arrays.setAll(data, i -> new Data(defaultTime, 0L)); + } + + /** + * Adds a new sample value to the metrics. + * @param value The new sample value + */ + public synchronized void add(long value) { + add(System.nanoTime(), value); + } + + /** + * Calculates the average for the last windowTimeSize window. + * @return The average + */ + public synchronized long value() { + return value(System.nanoTime()); + } + + @VisibleForTesting + void add(long time, long value) { + data[nextPos].nanoTime = time; + data[nextPos].value = value; + nextPos++; + if (nextPos == windowDataSize) { + nextPos = 0; + } + } + + @VisibleForTesting + long value(long time) { + // We expect that the data time positions are strictly increasing and the time is greater than + // any of the data position time. This is ensured by using System.nanoTime(). + long sum = 0L; + long lastTime = time; + long minTime = lastTime - windowTimeSize; + int pos = nextPos - 1; + boolean firstRound = true; + while (firstRound || pos != nextPos - 1) { + // Loop the window + if (pos < 0) { + pos = windowDataSize - 1; + } + // If we are at the end of the window + if (data[pos].nanoTime < minTime) { + sum += (lastTime - minTime) * data[pos].value; Review comment: If we do not have enough data points and the time window starts before the first data point then we have to estimate the first part of the interval with the last data: timeWindowStart, time1/value1, time2/value2, time3/value3, timeWindowEnd We would like to calculate the average so before time1 we would like to apply value1 too. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org Issue Time Tracking ------------------- Worklog Id: (was: 254240) Time Spent: 1.5h (was: 1h 20m) > New metrics to get the average queue length / free executor number for a > given time window > ------------------------------------------------------------------------------------------ > > Key: HIVE-21823 > URL: https://issues.apache.org/jira/browse/HIVE-21823 > Project: Hive > Issue Type: Sub-task > Components: llap > Reporter: Peter Vary > Assignee: Peter Vary > Priority: Major > Labels: pull-request-available > Attachments: HIVE-21823.patch > > Time Spent: 1.5h > Remaining Estimate: 0h > > We need to calculate the average queue size / free executor size for a window > to have good data for making routing decisions. > Interesting things to consider: > * The timeĀ gap between arriving request can be different, so simple average > is not enough to have correct data > * We need to have 2 parameters > ** Time window length > ** Maximum data point numbers - so we will not collect "infinite" amount of > data at high load -- This message was sent by Atlassian JIRA (v7.6.3#76005)