# Tail Sampling — OpenTelemetry

Source: https://www.skillbyai.com/en/opentelemetry/s-tail

> Decide after the trace completes.

## Keep errors and slow traces

**Tail sampling** buffers spans until a trace is (likely) complete, then decides using its content: keep all traces with errors, all traces slower than a threshold, and a small percentage of the rest. In OpenTelemetry it is implemented by the Collector's **tail_sampling** processor (contrib). It needs memory to buffer spans for a decision window and all spans of a trace must reach the same Collector instance, so it runs in a gateway tier.

## A tail sampling policy

Errors and slow traces kept, 5% of the rest.

```yaml
processors:
  tail_sampling:
    decision_wait: 10s
    num_traces: 50000
    policies:
      - name: keep-errors
        type: status_code
        status_code:
          status_codes: [ERROR]
      - name: keep-slow
        type: latency
        latency:
          threshold_ms: 1000
      - name: sample-rest
        type: probabilistic
        probabilistic:
          sampling_percentage: 5
```

## Combine head and tail sampling carefully

Heavy head sampling before a tail sampler removes the errors the tail sampler would have kept.

**Quiz:** Why must tail sampling run where all spans of a trace arrive together?

- [ ] Spans are encrypted otherwise
- [ ] It only works on metrics
- [x] The decision needs the complete trace
- [ ] Because exporters require it

*Answer:* The decision needs the complete trace. Route by trace ID to one instance.