couchdb-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Paul Joseph Davis (JIRA)" <>
Subject [jira] Created: (COUCHDB-396) Fixing weirdness in couch_stats_aggregator.erl
Date Sun, 28 Jun 2009 01:32:47 GMT
Fixing weirdness in couch_stats_aggregator.erl

                 Key: COUCHDB-396
             Project: CouchDB
          Issue Type: Improvement
          Components: Database Core, HTTP Interface
    Affects Versions: 0.10
         Environment: trunk
            Reporter: Paul Joseph Davis
            Assignee: Paul Joseph Davis
             Fix For: 0.10
         Attachments: couchdb_stats_aggregator.patch

Looking at adding unit tests to the couchdb_stats_aggregator module the other day I realized
it was doing some odd calculations. This is a fairly non-trivial patch so I figured that I'd
put in JIRA and get feed back before applying. This patch does everything the old version
does afaict, but I'll be adding tests before I consider it complete.

List of major changes:

* The old behavior for stats was to integrate incoming values for a time period and then reset
the values and start integrating again. That seemed a bit odd so I rewrote things to keep
the average and standard deviation for the last N seconds with approximately 1 sample per
* Changed request timing calculations [note below]
* Sample periods are configurable in the .ini file. Sample periods of 0 are a special case
and integrate all values from couchdb boot up.
* Sample descriptions are in the configuration files now.
* You can request different time periods for the root stats end point.
* Added a sum to the list of statistics
* Simplified some of the external API

The biggest change is in how time for requests are calculated. AFAICT, the old way was accumulating
request timings in the stats collector and just adding new values as clock ticks went by as
everything else does which makes sense in the case of resetting counters every time period.
In the new way I'm keeping a list of the samples in the last time period and when I get a
clock tick part of the update is to remove the samples that have passed out of the time period.
For a variable like request_time this would lead to unbounded storage.

The new method is calculating the average time of all requests in a single clock tick (1s).
One thing this loses is when you start having lots of variability in a single clock tick.
Ie, your average request time is 100ms, but 10% of your requests are taking 500ms. I've read
of people doing the averaging trick but also storing quantile information as well [1]. There
are also algorithms for doing single pass quantile estimation and the like so its possible
to do those things in O(N) time. The issue with quantiles is that it'd start breaking the
logic of how the collector and aggregators are setup. As it is now, there's basically a one
event -> one stat constraint. For the time being I went without quartiles to minimize the
impact of the patch.

This code will also be on github [3] as I add patches.

[2] (See the QuantileEstimator

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message