Interpolation is not about filling arbitrary large gaps in a metric series, but *aligning* several series together, to make it possible to perform aggregation across sources.

Most of the time graphing in Datadog is all about mix together data from separate sources into a single line for your graph. However separate sources might not submit data at the same time and with the same frequency.

```
error.count | 3:00:00 3:00:10 3:00:20 3:00:30 3:00:40 ...
------------------+-------------------------------------------------
1: host:A,env:prod| 15 25
2: host:B,env:test| 10 40
------------------+-------------------------------------------------
sum (1+2) | 15? 10? 25? 40?
The above example shows that merging sources directly produces absurd results just because sources are not naturally aligned.
```

Interpolation solves this problem by providing relevant values *just for calculations*.

```
error.count | 3:00:00 3:00:10 3:00:20 3:00:30 3:00:40 ...
------------------+-------------------------------------------------
1: host:A,env:prod| 15 18.3 25 X
2: host:B,env:test| Y 10 30 40
------------------+-------------------------------------------------
sum (1+2) | 15 + Y 28.3 55 40 + X
where X and Y are interpolated using data after and before the interval displayed.
```

**In which cases does interpolation occur?**

Interpolation occurs when more than 1 source corresponds to your graph query i.e.:

- for space-aggregation: avg:system.cpu.user{env:prod}. If you have 2 or more hosts with the tag “env:prod”, our system will compute the average over time and needs interpolation to do so.
- for group queries:
*error.count{*} by {host}*.

Interpolation is not needed:

- when you graph 1 metric submitted from 1 source: avg:error.count{host:a} (we assume this host submits this metric always with the same tag list).

Interpolation is not performed for multi part queries (e.g. "avg:system.cpu.user{env:prod},avg:system.cpu.user{env:dev}").

The type of interpolation described in this article is also not performed for arithmetic. When evaluating queries, our backend rolls data up into intervals (one for each point in a time series graph, see this article for more details: https://help.datadoghq.com/hc/en-us/articles/204820019-Graphing-with-Datadog-from-the-query-to-the-graph). If a query involves arithmetic, and one of these intervals is missing data for part of a query, the query system will substitute 0 for that interval. This behavior cannot currently be controlled with the fill modifier.

**How to control interpolation?**

The default interpolation is *linear* and is performed *up to 5 min* after real samples.

The `fill`

modifier controls interpolation parameters:

- fill(linear,X) gives you a linear interpolation up to X seconds after real samples.
- fill(last, X) just replicates the last sample value up to X secs.
- fill(zero, X) inserts 0 where the interpolation is needed up to X secs.
- fill(null, X) disables interpolation, the value of X doesn’t matter.

**FAQ**

There’s a gap in my metric, fill(zero) doesn’t do anything, I still have a long straight line on my graph.

Graphs in Datadog are just a series of datapoints joined by lines. If you have a long period without any data, this will translate into a long straight line as interpolation is *not meant to add values to* your metric series.

Rather, interpolation is about *aligning* series to make aggregation and multi-line graphs possible.

I have disabled interpolation but I see my metrics dropping to 0 which is not expected.

These artificial dips are caused by front-end visualization enhancement. See this article for more information.

How to choose the interpolation method?

*linear* interpolation is a great fit for metrics reported on a steady basis from the same sources. For sparse metrics/ metrics reported from varying sources over time, it's often more interesting to disable interpolation.

*last* makes sense for instance if you send datapoints only when the value of the thing you measure changes.*zero* is good if you want to get accurate totals for `count`

metrics (statsd “increment”) like sum:error.count{*}.as_count()*null* will prevent graphs from displaying interpolated values 5 min after the last real value etc.

## 0 Comments