OpenTelemetry's metrics SDK enforces a default cardinality limit of 2000 unique attribute combinations per metric stream to prevent unbounded memory growth. When this limit is exceeded, overflow measurements are folded into a single data point marked with `otel.metric.overflow=true`, stripping original attributes. This means totals remain correct but filtered or grouped queries can undercount — affecting dashboards, SLOs, and alerts. The guide covers how to size limits appropriately (including worked examples for delta vs. cumulative temporality), how to detect overflow using PromQL queries, how to set up continuous monitoring and alerting, and a decision flow for responding when overflow occurs. Key advice: treat overflow as a signal to inspect instrumentation, not an automatic reason to raise the limit.
