How to graph percentiles in Datadog?

DogStatsD implementation

It’s possible to get percentiles in Datadog by submitting data as a histogram metric through DogStatsD. The Agent embeds a DogStatsD server that receives DogStatsD packets, perform data aggregation, and send final percentile metrics to Datadog.

Since this aggregation is taken care of on the collection side, this isn’t available as a graphing function in the GUI.

Out of the histogram data, you get: 95th percentile, 50th percentile, avg, max, count.

Additional percentiles

Through the “histogram_percentiles” line of the configuration file of the Agent, get extra percentiles, such as:

  • histogram_percentiles: 0.95, 0.75

More about histogram

Local aggregations

Histograms are computed every 10 seconds on a host per host basis by the Datadog Agent. This collection model comes with its advantages and its limitations.

Advantages

  • Raw datapoints used to compute histogram metrics are not exposed nor relayed to the Datadog site.
  • StatsD handles the relevant aggregations and submits the calculated data package to the Datadog server directly.

Disadvantages

  • If you have two reporting streams of aggregated data, it is not possible today to aggregate across the raw datapoints from both streams, only aggregate across the aggregates.

    • EX: Averaging across <METRIC_NAME>.avg for all regions, takes the average stream values for each region and produces an average of averages.
  • Making a change to increase tag complexity (adding additional tags to be more specific) leads to changes in the behavior of a rolled up metric visualization

    • EX: Whereas before the change <METRIC_NAME>.avg (without any tags) would be aggregating across all raw points (StatsD takes all the raw datapoints, aggregates it and then ships over a single metric stream), adding a tag like region (US, EU) tag causes StatsD to bin raw datapoints into two region bins, aggregate them, and ship over two streams. This means when graphing <METRIC_NAME>.avg AVG by * means an aggregate across the two streams rather than a single one.

Read more about the Datadog histograms characteristics