Q: Your Prometheus instance is crashing with OOM every few hours. You identify the culprit: a Kubernetes deployment metric with a `pod_name` label containing every pod UUID ever created in the cluster (including deleted pods). Why is cardinality so deadly, and how do you prevent this without restarting Prometheus?
Each unique combination of label values creates a separate time-series. Prometheus stores each series' metadata, recent data points, and indices in memory.
Want to master this scenario in a live sandbox? The Linux Foundation's Prometheus Certified Associate (PCA) & Monitoring Labs covers this exact problem with hands-on terminal drills.
🛠️ Production Runbook & Step-by-Step Resolution
Direct Answer: "Our SRE team tackled this monitoring and metrics bottleneck to eliminate false-positive alert fatigue. The interviewer is testing: Understanding metric cardinality, label design, active remediation.. I structure my answer around systematic triage first, root cause analysis second, and permanent remediation third."
Initial Diagnostics & Root Cause Analysis
Each unique combination of label values creates a separate time-series. Prometheus stores each series' metadata, recent data points, and indices in memory.
- 10 jobs × 100 instances × 50,000 pod UUIDs × 10 containers = 500 million series
- Each series occupies ~1KB of memory (metadata, indices) = 500GB needed (will OOM instantly on a 64GB server).
- Label Design (Prevent): Never use unbounded identifiers like pod_name or user_id as labels. Use only bounded dimensions:
Remediation & Permanent Safeguards
The Math: If you have a metric with labels {job, instance, pod_name, container}: Why deletes matter: When a pod is deleted, Prometheus doesn't immediately purge its cardinality. The metric remains "recorded" until the TSDB compaction cycle runs (days later), so cardinality grows unbounded. Prevention & Remediation: Reload Prometheus: kill -HUP (no restart, configs reloaded live). New scraped metrics won't have pod_name, and Prometheus will garbage-collect the old series over time. If a metric's cardinality exceeds 10,000, auto-alert before it crashes.
# Bad: pod_name has infinite cardinality
http_requests_total{method, status, pod_name}
# Good: namespace and deployment are bounded (~100s)
http_requests_total{method, status, namespace, deployment}
- Metric Relabeling (Active Remediation): Without restarting, add
metric_relabel_configsto your scrape config to drop the problematic label: - Cardinality Budgets: Proactively monitor cardinality:
- 10 jobs × 100 instances × 50,000 pod UUIDs × 10 containers = 500 million series
- Each series occupies ~1KB of memory (metadata, indices) = 500GB needed (will OOM instantly on a 6...
- Label Design (Prevent): Never use unbounded identifiers like pod_name or user_id as labels. Use o...
Senior hiring managers and bar raisers evaluate how you prioritize during outages. Avoid these common anti-patterns:
- Injecting high-cardinality label dimensions (such as user IDs or request UUIDs) into Prometheus metrics when configuring Observability Prometheus instance High-cardinality label values cause exponential time-series growth, consuming gigabytes of TSDB memory and crashing Prometheus servers.
- Evaluating application performance solely through arithmetic mean or p50 latency while ignoring tail latency (p99) when configuring Observability Prometheus instance Averages hide catastrophic tail latency outages where 1% of high-value user transactions suffer severe timeouts.
- Failing to inject and extract W3C distributed trace context across message brokers and RPC calls when configuring Observability Prometheus instance Breaking distributed trace context propagation fragments end-to-end transaction waterfalls into disconnected, un-correlatable trace segments.
- Configuring alert rule evaluation intervals without sufficient duration damping (for: 5m), inducing alert flap when configuring Observability Prometheus instance Instant alerts triggered by transient sub-second CPU or latency blips page on-call engineers unnecessarily and dilute response urgency.