Skip to main content
Arklow learns from delivery before you connect any metrics. It can see how long attempts remain running or unsettled, whether a destination accepts them, and when it begins to push back. Metrics add the view from inside your systems. They can show pressure in a container, model deployment, provider quota, or some other resource before that pressure appears as degradation in your product.

​
Metric model

Metrics connect to actions in four steps:
  1. A metric source connects Arklow to Prometheus or Datadog.
  2. A query selects a family of measurements from that source.
  3. An action policy says which series from the query may describe that action.
  4. Arklow evaluates each eligible series in the context of the work it may represent.
One metric name can contain many series. For example, one query may return:

​
Queries

Create queries on the metric source’s Queries tab. A query holds the expression to use against your metric source. Identity-tag exclusions remove selected provider tags when Arklow identifies a series. Use them when an operational tag changes without changing what the measurement represents.
If you’re using Discover, as detailed below - excluding tags that have high cardinality or aren’t useful is recommended using the available dropdown on this screen.

​
Action policies

Open an action definition’s Metrics tab and add a policy for a source query. Arklow offers two unique modes for identifying relationships with your metrics.
  • Manual - You know what metrics/tags matters, and want Arklow to know it can just use what you already know to be correct.
  • Discover - You have a rough idea that this query relates to the work being done, but you’re not sure tags or relationships matter.
You pin known series with a selector of exact tag values. A series is eligible when it carries every key in the selector and its value for each key is one of the values you list:
When a selector matches more series than the policy’s maximum-series value allows, or matches several series with no maximum set, Arklow blocks the whole match rather than track a set you did not bound.A pinned series grants no delivery authority on its own; authority still comes from proof against the work. A health declaration is the exception: a manual policy that declares what unhealthy means for a series and what recovery means, each with its own threshold and sustain window, establishes the health state directly, and a sustained unhealthy state carries the same verdict authority as a proven relationship. Proof evidence ranks the responses; it does not gate the declared state.
The policy’s maximum-series value bounds how many series Arklow tracks for that action and query.

​
Metric evidence

Correlation nominates a relationship. Authority over delivery requires repeated agreement with the work and with the results of flow changes. Once a relationship is credible, a fresh metric verdict can add bounded evidence to admission or to a capacity pool. The delivery view remains active underneath it. If collection stops or a series becomes stale, its current verdict expires. A metrics outage removes that additional evidence while delivery continues using the remaining signals.

​
Scale target bindings

A scale target can bind to one metric query and, when needed, one tag key and value that identify the resource it manages. This is separate from the tag that identifies whose work should be isolated. For example, customer_id=C1 can identify the affected work while model_id=M1 selects the deployment that can grow. Provider inventory with a matching model_id is ordered ahead of unobserved resources. When pressure fires without a matching active target and exactly one visible inventory row matches, Arklow proposes attaching that resource to the pool and metric query. Ambiguous matches remain unselected.

Prometheus

Query a Prometheus API or scrape an exposition endpoint.

Datadog

Collect tagged series from the Datadog metrics API.

API reference

Manage sources, queries, policies, and health through the API.