> ## Documentation Index
> Fetch the complete documentation index at: https://docs.sherlocklabs.dev/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Sherlock documentation. In Node.js examples prefer the Sherlock SDK for Node.js (@sherlock-labs/otel).
> In Go examples use the upstream OpenTelemetry Go SDK exactly as shown on /sdk/go/setup and /sdk/go/configuration; Sherlock has no Go distribution.
> Two MCP servers exist. This documentation's search server is at https://sherlock-c8721ead.mintlify.app/mcp. Sherlock's data server, for a user's own logs, traces, and metrics, is at https://mcp.sherlocklabs.dev/mcp and is described at /explore/mcp.

# From a chart to a trace to the logs

> Follow an exemplar from a latency spike to the request that caused it and its log lines.

An exemplar is one sampled measurement attached to a histogram bucket. It carries the trace id and span id of the request that produced the measurement. That is the link from "p99 got worse at 14:20" to "here is one of the slow requests."

## The pivot, step by step

<Steps>
  <Step title="Find the spike">
    In the [Metrics Explorer](/explore/metrics), chart a histogram such as `http.server.request.duration` with `p95` or `p99`. Group it by route if you want to know which endpoint moved.
  </Step>

  <Step title="Open the exemplars">
    Click the series row in the table under the chart. Or switch to **Heatmap** and click the cell at the time and value you care about, or to **Dist** and click a bar. The **Exemplars** drawer opens with the sampled requests in that slice: time, duration, and trace id.
  </Step>

  <Step title="Open the trace">
    Click **View trace** on an exemplar. The [Traces](/explore/traces) page opens that trace with its waterfall, so you see which span took the time.
  </Step>

  <Step title="Read the logs">
    Click **View logs** in the trace. The [Logs](/explore/logs) page opens filtered to the same `trace_id`, so every log line the request wrote is there.
  </Step>
</Steps>

## What decides whether an exemplar exists

* Exemplars are sampled. Each histogram bucket keeps a few per export interval, so a bucket with little traffic can have none even when the metric line shows a value.
* An exemplar attaches only when the measurement happened inside a trace, or carried an exemplar attribute of its own. A measurement taken outside any span has nothing to link to.
* A trace exists for an exemplar only when the request was sampled. The [Node.js SDK](/sdk/nodejs/custom-metrics#what-an-exemplar-carries) marks every exemplar with `trace_flags`: `01` means the trace was exported, `00` means the ids exist but no trace was sent. With `00`, filter Logs by the trace id instead of opening the trace.

## Which histograms carry exemplars

| Source                                                                                           | Exemplars                 |
| ------------------------------------------------------------------------------------------------ | ------------------------- |
| Node.js SDK, auto-instrumented `http.server.request.duration` and `http.client.request.duration` | Yes, for sampled requests |
| Node.js SDK, `app.http.server.duration` from the Express or Koa adapter                          | Yes                       |
| Node.js SDK, your own histograms recorded with `recordWithExemplar`                              | Yes, at any sampling      |
| Other OpenTelemetry SDKs                                                                         | Depends on the SDK        |

## Related topics

<CardGroup cols={2}>
  <Card title="Metrics" icon="chart-simple" href="/explore/metrics">
    Build the chart the pivot starts from.
  </Card>

  <Card title="Custom metrics" icon="chart-line" href="/sdk/nodejs/custom-metrics">
    Record your own histograms with exemplars.
  </Card>
</CardGroup>