Services Work Pricing Blog About Book a free call →

CloudWatch EKS telemetry filters: cut ingestion cost

AWS added include and exclude filters to the CloudWatch EKS add-on. What changed, who should care, and how to trial it safely this week.

On November 20, 2026, AWS announced telemetry filtering for the Amazon CloudWatch Observability EKS add-on. You can now set cluster-wide include and exclude lists for namespaces, metric names, and pod and node labels, so the agent stops sending OpenTelemetry data you never look at. If your EKS observability bill grows with every new service, this is a direct lever on it. It needs add-on version 8.0.0 or later, and it is worth a careful trial rather than a quick switch, because a filter can also remove data your alarms depend on.

What was announced, and what is new?

AWS says the add-on's telemetry is billed per GB of ingestion, so collecting less lowers the bill. Telemetry sent through the add-on is charged at the Amazon EKS OpenTelemetry ingestion rate of $0.08 per GB, as of November 2026. AWS's CloudWatch pricing page says ingestion for Container Insights with OpenTelemetry on EKS has no separate charges for API calls, storage or metric series count. Query charges for PromQL apply separately, per the same page. AWS says the filters are available in all Regions where CloudWatch supports OpenTelemetry metrics.

Before this, you could already turn off features you did not use, such as the OTLP endpoint, span metrics or individual Container Insights solutions, and version 7.0.0 removed duplicate and generated data. What is new is a cluster-wide include and exclude layer. The telemetry filters documentation describes otelFilters settings (the filter scope is OpenTelemetry telemetry, not everything the add-on can send) that you change with one add-on configuration update. Each change restarts the affected agent pods, because the agent reads its configuration only at startup.

What can you filter?

The documentation lists three kinds of filter:

  • Namespaces. Drop the pod and container telemetry of some namespaces, for example kube-system or every namespace starting with test-.
  • Metric names. Drop metrics by name, for example container_network_*.
  • Pod and node labels. Choose which labels the agent adds as k8s.pod.label.* and k8s.node.label.* attributes.

Matching works like this, per the documentation: an entry is an exact name, a prefix ending in *, or * alone. Matching is case-sensitive. An empty include list means "everything". Exclude wins over include.

The label filter matters most for cost. AWS's cost controls page says that for OpenTelemetry metrics, each pod and node label the agent adds as an attribute increases the data you store and query. High-cardinality labels cost money. By default, the add-on also removes a built-in list of such labels (topology, hostname, instance type and similar), through a setting called recommendedExclusions.

What does a filter not touch?

This is where trials go wrong. Per the documentation:

  • The namespace filter does not apply to node-level metrics, host logs, or control-plane component metrics such as Karpenter and KEDA.
  • Namespace and metric-name filters do not apply to ServiceMonitor and PodMonitor scraping. Label filters do.
  • GPU, Neuron and EFA device metrics are never dropped by the namespace filter. The agent keeps them and removes their pod, namespace and container attributes.
  • A non-empty metric include list keeps only the listed metrics. It drops every other OpenTelemetry Container Insights metric and every other OTLP metric, span metrics included.

That last rule is the sharp edge. A short include list looks tidy, but it can silently empty a dashboard or leave an alarm with no data. An exclude list is the safer first step.

Who should care, and who can ignore it?

Care if:

  • You run EKS with the add-on's OpenTelemetry Container Insights or OTLP endpoint, and the ingestion line on your bill is growing.
  • You have many short-lived or test namespaces that nobody dashboards.
  • Your teams label pods freely and those labels reach metrics.

Probably ignore it if:

  • OpenTelemetry Container Insights is off by default, per AWS. If you only use Enhanced Container Insights (Classic), these filters have little to act on.
  • You do not use this add-on, for example you run your own collector or another vendor's agent. The filters live in the add-on, not in your pipeline.
  • Your cluster is small and the observability cost is minor. A cheaper first step is to measure it before investing time.

How would a team try it this week?

A cautious plan, as an approach rather than a prescription:

  1. Check the version. Run aws eks describe-addon --cluster-name <name> --addon-name amazon-cloudwatch-observability --query addon.addonVersion --output text (Helm users: helm list -n amazon-cloudwatch). Filters need 8.0.0 or later. AWS's documentation says EKS does not update the add-on on its own, and that the 7.0.0 release changed or removed some series, so read its upgrade notes and check queries and alarms before a jump from 6.x.
  2. Use a non-production cluster first. Filter changes restart agent pods, so avoid doing it in a busy production window.
  3. Start with one namespace exclude. For example, a test-* pattern. It is easy to explain and easy to revert.
  4. Then label filters. Remove labels nobody groups by. Keep the ones used in dashboards and alarms.
  5. Compare ingestion before and after over several days, using your own AWS billing data.

What would we check first?

Before turning on any filter in production, we start with the consumers of the data, not the producers:

  • Which alarms and dashboards use the metrics or labels you plan to drop? An alarm on a missing label gets no data, and its state then depends on how it treats missing data. AWS says the same about the 7.0.0 upgrade.
  • Which labels do your cost reports use? If you allocate cost by team or app label, do not exclude them. If you use an include list, it keeps only the labels named in it unless you add *, per the documentation.
  • Who owns the config? Put the otelFilters block in the same infrastructure-as-code repository as the add-on, with review, so a filter is a reviewed change and not a console edit.

A useful rule of thumb: for each excluded item, write down who loses what. If the answer is "nobody", exclude it. If it is "maybe someone", ask first.

What does it cost to get this wrong?

The risk of over-filtering is not on the bill: it shows up during an incident, when a metric you dropped is the one you need. That is why we prefer excludes with a named reason over broad includes. AWS does not quantify typical savings in the announcement, and neither do we: measure your own before and after.

If you want a second pair of eyes on an observability bill or an EKS setup, our services cover cloud cost reviews, and the fractional CTO option covers ownership of decisions like this one.

This article is general information based on AWS's public documentation as of November 2026. It is not financial or professional advice; check current AWS documentation and pricing for your account and Region.