[
https://issues.apache.org/jira/browse/HDFS-16949?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17701438#comment-17701438
]
ASF GitHub Bot commented on HDFS-16949:
---------------------------------------
rdingankar commented on code in PR #5486:
URL: https://github.com/apache/hadoop/pull/5486#discussion_r1139490394
##########
hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/metrics2/util/SampleQuantiles.java:
##########
@@ -243,7 +245,12 @@ synchronized public Map<Quantile, Long> snapshot() {
Map<Quantile, Long> values = new TreeMap<Quantile, Long>();
for (int i = 0; i < quantiles.length; i++) {
- values.put(quantiles[i], query(quantiles[i].quantile));
+ /* eg : effectiveQuantile for 0.99 with inverseQuantiles will be 0.01.
+ For inverse quantiles higher numeric value is better and hence we want
+ to query from the opposite end of the sorted sample
+ */
+ double effectiveQuantile = inverseQuantiles ? 1 - quantiles[i].quantile
: quantiles[i].quantile;
+ values.put(quantiles[i], query(effectiveQuantile));
Review Comment:
Inversing the traversal does not give us the same thing. The iterator is a
variable for calculating the allowableError, which becomes high if the
traversal is inverted. It will require more math in the allowableError() method
based on [CKMS
paper.](http://dimacs.rutgers.edu/~graham/pubs/papers/bquant-icde.pdf)
Looking for the inverse quantile (1 - quantile) however works as expected as
shown in the UT runs from the PR description.
> Update ReadTransferRate to ReadLatencyPerGB for effective percentile metrics
> ----------------------------------------------------------------------------
>
> Key: HDFS-16949
> URL: https://issues.apache.org/jira/browse/HDFS-16949
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: datanode
> Reporter: Ravindra Dingankar
> Assignee: Ravindra Dingankar
> Priority: Minor
> Labels: pull-request-available
> Fix For: 3.3.0, 3.4.0
>
>
> HDFS-16917 added ReadTransferRate quantiles to calculate the rate which data
> is read per unit of time.
> With percentiles the values are sorted in ascending order and hence for the
> transfer rate p90 gives us the value where 90 percent rates are lower
> (worse), p99 gives us the value where 99 percent values are lower (worse).
> Note that value(p90) < p(99) thus p99 is a better transfer rate as compared
> to p90.
> However as the percentile increases the value should become worse in order to
> know how good our system is.
> Hence instead of calculating the data read transfer rate, we should calculate
> it's inverse. We will instead calculate the time taken for a GB of data to be
> read. ( seconds / GB )
> After this the p90 value will give us 90 percentage of total values where the
> time taken is less than value(p90), similarly for p99 and others.
> Also p(90) < p(99) and here p(99) will become a worse value (taking more time
> each byte) as compared to p(90)
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
