A metric type is an indication of what you are trying to represent with your metric and its emission source. If you refer to the metric types documentation, you can see that the COUNT and RATE metric types are quite similar to each other, as they represent the same concept: the variation of a metric value over time. However, they use different logic:
RATE: Normalized value variation over time (per second).COUNT: Absolute value variation over a given time interval.Depending on your use case and submission method, one metric type may be more suited than the other for submission. For instance:
| Metric type submitted | Use case |
|---|---|
RATE | You want to monitor the number of requests received over time across several hosts. |
RATE | You do not have control over the consistency of temporal count submissions across your sources, so you’re normalizing by each individual interval to be able to compare them upstream. |
COUNT | You want to count the number of times a function is called. |
COUNT | Counting the amount of revenue that have been made over a given amount of time. |
Since RATE and COUNT aren’t the same metric type, they don’t have the same behavior or shape within Datadog graphs and monitors. To change metrics on the fly between RATE and COUNT representations, use Datadog’s in-application modifier functions within your graphs and monitors.
The two main in-application modifiers are as_count() and as_rate().
| Modifiers | Description |
|---|---|
as_count() | Sets the operations necessary to display the given metric in COUNT form, giving you the absolute variation of a metric value over a rollup interval. Note that since it depends on the rollup interval, graphing a longer time interval changes your graph shape. |
as_rate() | Sets the operations necessary to display the given metric in RATE form, giving you the absolute variation of a metric value per second. |
Depending on the metric type you applied them to, the behavior differs:
as_count():SUM.as_rate():SUM.[1,1,1,1].as_rate() with a rollup interval of 20s would produce [0.05, 0.05, 0.05, 0.05].Note: There is no normalization on very small intervals (when no time aggregation occurs), thus the raw metric value counts are returned.
as_count():[0.05, 0.05, 0.05, 0.05].as_count() with a rollup interval of 20s would produce [1,1,1,1].as_rate():SUM.GAUGE metric types represent the absolute and final value of a metric; as_count() and as_rate() modifiers have no effect on them.
While it is not normally required, it is possible to change a metric’s type in the metric summary page:
Example use case:
You have a metric app.requests.served that counts requests served, but accidentally submitted it via StatsD as a GAUGE. The metric’s Datadog type is, therefore, GAUGE.
You wanted to submit app.requests.served as a StatsD COUNT metric for time aggregation. This would help answer questions like “How many total requests were served in the past day?" by querying sum:app.requests.served{*} (this would not make sense for a GAUGE metric type.)
You like the name app.requests.served, so rather than submitting a new metric name with a more appropriate COUNT type, you could change the type of app.requests.served by updating:
dogstatsd.increment('app.requests.served', N) after N requests are served, andRATE.This causes data submitted before the type change for app.requests.served to behave incorrectly because it was stored in a format to be interpreted as an in-app GAUGE (not a RATE). Data submitted after step 3 is interpreted properly.
If you are not willing to lose the historical data submitted as a GAUGE, create a new metric name with the new type, leaving the type of app.requests.served unchanged.
Note: For the AgentCheck, self.increment does not calculate the delta for a monotonically increasing counter; instead, it reports the value passed in at the check run. To send the delta value on a monotonically increasing counter, use self.monotonic_count.