← All posts Prometheus

Set a Prometheus cardinality budget before adding labels

The Tryssh team ·

For Prometheus cardinality budget, the safest approach is a bounded operational change, not a command pasted without context. This runbook starts with effective state, shows the smallest candidate action, and finishes by repeating the real user or system path.

TL;DR: Cardinality is the number of distinct active label sets; multiplying unbounded dimensions increases memory, storage, query, and remote-write cost. Capture a baseline with promtool query instant http://127.0.0.1:9090 'count({__name__!=\"\"})', make the reviewed change only when the evidence matches, then verify with promtool query instant http://127.0.0.1:9090 'prometheus_tsdb_head_series' and keep the rollback ready.

Audience: operators who can inspect Prometheus targets, labels, rules, and query results without treating a dashboard as ground truth. This guide assumes familiarity with Prometheus operations and a change window appropriate to the system.

Set a Prometheus cardinality budget before adding labels observe, interpret, change, and verify workflow

The direct answer

New instrumentation has an estimated series ceiling and observed head-series growth remains within it. That is the success condition for Prometheus cardinality budget; command completion by itself is not enough.

The important boundary is instrumentation, scrape discovery, relabeling, sample ingestion, query evaluation, rule state, and notification delivery. Cardinality is the number of distinct active label sets; multiplying unbounded dimensions increases memory, storage, query, and remote-write cost. If an observation does not identify which side of that boundary failed, collect a narrower observation before changing state.

How the mechanism works

For Prometheus cardinality budget, use this mental model: Prometheus discovers targets, transforms labels, scrapes timestamped samples, evaluates PromQL over those series, and turns rule results into recorded series or alert states. The model prevents a common mistake—treating configuration text, control-plane acceptance, process state, and end-user behavior as the same proof.

Follow four stages:

  1. Observe: identify the exact host, object, version, owner, and active configuration.
  2. Interpret: write the expected result, the abnormal result, and what would remain inconclusive.
  3. Change: apply one reviewed action at the narrowest layer that contradicts the baseline.
  4. Verify: repeat the original path and compare the same evidence, including adjacent safety controls.

Preflight and safety boundary

Run new queries over bounded ranges, estimate series growth, test rule files, and keep existing alerts until replacement coverage is proven.

Before Prometheus cardinality budget, record UTC time, the current version or digest, the exact target, recent changes, and who owns the workload. The rollback for this runbook is: revert the instrumentation or metric relabeling change, restart only if the client requires it, and allow obsolete series to age out under retention.

Do not continue if the target identity is ambiguous, the current state cannot be saved, the only recovery session would be at risk, or the proposed command affects more objects than the brief names.

Capture the read-only baseline

Run these commands one at a time. Replace example names and addresses deliberately; do not paste production secrets into a transcript.

promtool query instant http://127.0.0.1:9090 'count({__name__!=\"\"})'
promtool query instant http://127.0.0.1:9090 'topk(20, count by (__name__) ({__name__!=\"\"}))'
promtool query instant http://127.0.0.1:9090 'prometheus_tsdb_head_series'

Interpret the baseline before moving on:

  • Expected: new instrumentation has an estimated series ceiling and observed head-series growth remains within it.
  • Abnormal: user IDs, paths, errors, hashes, or pod churn create values with no bounded vocabulary.
  • Inconclusive: missing output can also mean the wrong context, permissions, namespace, log window, binary, or target. Prove those assumptions before treating absence as health.

Save the decisive output, exit status, and timestamp. Redact credentials, customer data, private topology, tokens, and complete environment dumps.

Apply the smallest candidate change

The following is state-changing example syntax, not an instruction to run it unchanged:

sudo promtool check rules /etc/prometheus/rules/cardinality.rules.yml
sudo kill -HUP \"$(pidof prometheus)\"

For Prometheus cardinality budget, the proposed change is acceptable only when the read-only baseline predicts its effect and the rollback is available. The key risk is: dropping labels at query time does not remove already ingested series cost.

Prefer an immutable artifact, validated configuration, dry-run, transaction, candidate object, or staged target when the tool supports one. Record the exact command and UTC time so later telemetry can be correlated to the change.

Verify the result from the outside in

promtool query instant http://127.0.0.1:9090 'prometheus_tsdb_head_series'
promtool query instant http://127.0.0.1:9090 'topk(20, count by (__name__) ({__name__!=\"\"}))'
promtool query instant http://127.0.0.1:9090 'scrape_samples_post_metric_relabeling'

Verification for Prometheus cardinality budget has three layers:

  1. The control plane or command reports the intended effective state.
  2. The process, resource, or data path reflects that state without a new pressure signal.
  3. The original user-visible or dependent-system path succeeds from an independent vantage point.

If promtool query instant http://127.0.0.1:9090 'prometheus_tsdb_head_series' succeeds but the original path still fails, stop. The change may have repaired a local symptom while DNS, policy, routing, caching, dependency, or client state remains broken.

Failure branches

The baseline does not match this runbook

When user IDs, paths, errors, hashes, or pod churn create values with no bounded vocabulary, do not force the candidate command. Return to identity and scope, compare a healthy peer only through effective settings, and name a new falsifiable mechanism.

The change succeeds but behavior does not

A successful sudo promtool check rules /etc/prometheus/rules/cardinality.rules.yml proves that one interface accepted a request. It does not prove convergence, readiness, data compatibility, external routing, or client recovery. Re-run the same evidence at each downstream boundary.

The change makes the system worse

Execute the written rollback: revert the instrumentation or metric relabeling change, restart only if the client requires it, and allow obsolete series to age out under retention. Preserve the failed candidate and relevant logs long enough to explain the outcome; do not destroy the evidence with broad cleanup or repeated restarts.

Operator checklist

  • Confirm the exact target, context, identity, version, and active owner.
  • Capture the read-only baseline and one disconfirming observation.
  • Label sudo promtool check rules /etc/prometheus/rules/cardinality.rules.yml as state-changing during review.
  • Keep recovery access and rollback independent of the path being edited.
  • Change one layer, record UTC time, and wait for its real convergence boundary.
  • Verify the original path, adjacent controls, resource pressure, and persistence.
  • Update the runbook when observed behavior differs from the source-reviewed model.

Investigate it in Tryssh

Tryssh keeps the command, approval boundary, output, and verification beside the host conversation.

Tryssh can preserve this evidence loop and show a state-changing command for human approval. It does not make the operator's identity, recovery access, rollback, or platform authority decisions.

Evidence and review status

This Prometheus cardinality budget runbook was source-reviewed on 2026-07-29 against current first-party documentation. The commands are illustrative and use example targets. The page does not claim that the change was reproduced across every distribution, managed service, version, network, or workload.

Limitations and trade-offs

Dropping labels at query time does not remove already ingested series cost. Managed platforms may generate configuration, restrict privileges, replace local state, or expose a different control plane than the upstream project. Confirm the installed version and provider contract before applying a repair.

Search visibility is not proof of operational correctness. Treat this page as a decision aid, preserve independent recovery, and stop when the evidence contradicts its assumptions.

Related operator runbooks

Continue with SLO burn rate alerts, Prometheus relabel_configs, the Prometheus operations foundation guide, and the SSH hardening checklist.

Sources and further reading