Sampling overrides - Azure Monitor Application Insights for Java

Note

The sampling overrides feature is in GA, starting from 3.5.0.

Sampling overrides allow you to override the default sampling percentage, for example:

  • Set the sampling percentage to 0 (or some small value) for noisy health checks.
  • Set the sampling percentage to 0 (or some small value) for noisy dependency calls.
  • Set the sampling percentage to 100 for an important request type (for example, /login) even though you have the default sampling configured to something lower.

Terminology

Before you learn about sampling overrides, you should understand the term span. A span is a general term for:

  • An incoming request.
  • An outgoing dependency (for example, a remote call to another service).
  • An in-process dependency (for example, work being done by subcomponents of the service).

For sampling overrides, these span components are important:

  • Attributes

The span attributes represent both standard and custom properties of a given request or dependency.

Getting started

To begin, create a configuration file named applicationinsights.json. Save it in the same directory as applicationinsights-agent-*.jar. Use the following template.

{
  "connectionString": "...",
  "sampling": {
    "percentage": 10,
    "overrides": [
      {
        "telemetryType": "request",
        "attributes": [
          ...
        ],
        "percentage": 0
      },
      {
        "telemetryType": "request",
        "attributes": [
          ...
        ],
        "percentage": 100
      }
    ]
  }
}

How it works

telemetryType (telemetryKind in Application Insights 3.4.0) must be one of request, dependency, trace (log), or exception.

When a span is started, the type of span and the attributes present on it at that time are used to check if any of the sampling overrides match.

Matches can be either strict or regexp. Regular expression matches are performed against the entire attribute value, so if you want to match a value that contains abc anywhere in it, then you need to use .*abc.*. A sampling override can specify multiple attribute criteria, in which case all of them must match for the sampling override to match.

If one of the sampling overrides matches, then its sampling percentage is used to decide whether to sample the span or not.

Only the first sampling override that matches is used.

If no sampling overrides match:

  • If it's the first span in the trace, then the top-level sampling configuration is used.
  • If it isn't the first span in the trace, then the parent sampling decision is used.

Example: Suppress collecting telemetry for health checks

This example suppresses collecting telemetry for all requests to /health-checks.

This example also suppresses collecting any downstream spans (dependencies) that would normally be collected under /health-checks.

{
  "connectionString": "...",
  "sampling": {
    "overrides": [
      {
        "telemetryType": "request",
        "attributes": [
          {
            "key": "url.path",
            "value": "/health-check",
            "matchType": "strict"
          }
        ],
        "percentage": 0
      }
    ]
  }
}

Example: Suppress collecting telemetry for a noisy dependency call

This example suppresses collecting telemetry for all GET my-noisy-key redis calls.

{
  "connectionString": "...",
  "sampling": {
    "overrides": [
      {
        "telemetryType": "dependency",
        "attributes": [
          {
            "key": "db.system",
            "value": "redis",
            "matchType": "strict"
          },
          {
            "key": "db.statement",
            "value": "GET my-noisy-key",
            "matchType": "strict"
          }
        ],
        "percentage": 0
      }
    ]
  }
}

Example: Collect 100% of telemetry for an important request type

This example collects 100% of telemetry for /login.

Since downstream spans (dependencies) respect the parent's sampling decision (absent any sampling override for that downstream span), they're also collected for all '/login' requests.

{
  "connectionString": "...",
  "sampling": {
    "percentage": 10
  },
  "sampling": {
    "overrides": [
      {
        "telemetryType": "request",
        "attributes": [
          {
            "key": "url.path",
            "value": "/login",
            "matchType": "strict"
          }
        ],
        "percentage": 100
      }
    ]
  }
}

Span attributes available for sampling

Span attribute names are based on the OpenTelemetry semantic conventions. (HTTP, Messaging, Database, RPC)

https://github.com/open-telemetry/semantic-conventions/blob/main/docs/README.md

Note

To see the exact set of attributes captured by Application Insights Java for your application, set the self-diagnostics level to debug, and look for debug messages starting with the text "exporting span".

Note

Only attributes set at the start of the span are available for sampling, so attributes such as http.response.status_code or request duration which are captured later on can be filtered through OpenTelemetry Java extensions. Here is a sample extension that filters spans based on request duration.

Troubleshooting

If you use regexp and the sampling override doesn't work, try with the .* regex. If the sampling now works, it means you have an issue with the first regex and read this regex documentation.

If it doesn't work with .*, you might have a syntax issue in your application-insights.json file. Look at the Application Insights logs and see if you notice warning messages.