Buildkite Agent Metrics

A command-line tool for collecting Buildkite agent metrics, focusing on enabling auto-scaling. Currently AWS Cloudwatch, StatsD, Prometheus, Stackdriver, New Relic, and OpenTelemetry are supported.

Installing

The latest binary is available from Github Releases.

Container images

We also publish Docker Images to Amazon Public ECR. These are based on Alpine Linux.

You can run the container with a Docker CLI command such as:

docker run --rm public.ecr.aws/buildkite/agent-metrics:latest \
  -token abc123 \
  -interval 30s \
  -queue my-queue

AWS Lambdas

We also publish an AWS Lambda to S3 in region us-east-1. Each version is published to s3://buildkite-lambdas/buildkite-agent-metrics/v${VERSION}/handler.zip and is also available from Github Releases as handler.zip.

Example

aws lambda create-function \
  --function-name buildkite-agent-metrics \
  --memory 128 \
  --role arn:aws:iam::account-id:role/execution_role \
  --runtime provided.al2 \
  --code S3Bucket=buildkite-lambdas,S3Key=buildkite-agent-metrics/v5.10.0/handler.zip \
  --handler handler

Installing from source

Go can be used to download and build the binary from source:

go install github.com/buildkite/buildkite-agent-metrics/v5@latest

This typically installs the buildkite-agent-metrics binary into ~/go/bin.

Running

Several running modes are supported. All of them require an Agent Registration Token, found on the Buildkite Agents page.

Running as a Daemon

The simplest deployment is to run as a long-running daemon that collects metrics across all queues in an organization.

buildkite-agent-metrics -token abc123 -interval 30s

Restrict it to a single queue with -queue:

buildkite-agent-metrics -token abc123 -interval 30s -queue my-queue

Restrict it to multiple queues by repeating -queue:

buildkite-agent-metrics -token abc123 -interval 30s -queue my-queue1 -queue my-queue2

When using clusters, you can pass a cluster registration token to gather metrics only for that cluster:

buildkite-agent-metrics -token clustertoken ...

You can repeat -token to gather metrics for multiple clusters:

buildkite-agent-metrics -token clusterAtoken -token clusterBtoken ...

Running as an AWS Lambda

An AWS Lambda bundle is created and published as part of the build process. The Lambda will require the cloudwatch:PutMetricData IAM permission.

It requires a provided.al2 environment and respects the following env vars:

• BUILDKITE_BACKEND : The name of the backend to use (e.g. cloudwatch, statsd, newrelic, opentelemetry. For the lambda, prometheus and stackdriver are not supported).
• BUILDKITE_QUEUE : A comma separated list of Buildkite queues to process (e.g. backend-deploy,ui-deploy).
• BUILDKITE_QUIET : A boolean specifying that only ERROR log lines must be printed. This accepts either 1 or true to enable.
• BUILDKITE_CLOUDWATCH_DIMENSIONS : A comma separated list in the form of Key=Value,Other=Value containing the Cloudwatch dimensions to index metrics under.
• BUILDKITE_CLOUDWATCH_HIGH_RESOLUTION : Whether to enable High-Resolution Metrics which incurs additional charges. This accepts either 1 or true to enable.

To override the endpoint use the following env var:

• BUILDKITE_AGENT_ENDPOINT: Endpoint URL of the Buildkite Agent API (default https://agent.buildkite.com/v3)

To adjust timeouts, and connection pooling in the HTTP client use the following env vars:

• BUILDKITE_AGENT_METRICS_TIMEOUT : Timeout, in seconds, TLS handshake and idle connections, for HTTP requests, to Buildkite API (default 15).
• BUILDKITE_AGENT_METRICS_MAX_IDLE_CONNS : Maximum number of idle (keep-alive) HTTP connections for Buildkite Agent API. Zero means no limit, -1 disables pooling (default 100).

To assist with debugging the following env vars are provided:

• BUILDKITE_AGENT_METRICS_DEBUG : A boolean which enables debug logging. This accepts either 1 or true to enable.
• BUILDKITE_AGENT_METRICS_DEBUG_HTTP : A boolean which enables printing of the HTTP responses. This accepts either 1 or true to enable.

Additionally, one of the following groups of environment variables must be set in order to define how the Lambda function should obtain the required Buildkite Agent API token:

Option 1 - Provide the token(s) as plain-text

• BUILDKITE_AGENT_TOKEN : The Buildkite Agent API token to use. You can supply multiple tokens comma-separated.

Option 2 - Retrieve token from AWS Systems Manager

• BUILDKITE_AGENT_TOKEN_SSM_KEY : The parameter name which contains the token value in AWS Systems Manager. You can supply multiple names comma-separated.

Note: Parameters stored as String and SecureString are currently supported.

Option 3 - Retrieve token from AWS Secrets Manager

• BUILDKITE_AGENT_SECRETS_MANAGER_SECRET_ID: The id of the secret which contains the token value in AWS Secrets Manager. You can supply multiple ids comma-separated.
• (Optional) BUILDKITE_AGENT_SECRETS_MANAGER_JSON_KEY: The JSON key containing the token value in the secret JSON blob.

Note 1: Both SecretBinary and SecretString are supported. In the case of SecretBinary, the secret payload will be automatically decoded and returned as a plain-text string.

Note 2: BUILDKITE_AGENT_SECRETS_MANAGER_JSON_KEY can be used on secrets of type SecretBinary only if their binary payload corresponds to a valid JSON object containing the provided key.

aws lambda create-function \
  --function-name buildkite-agent-metrics \
  --memory 128 \
  --role arn:aws:iam::account-id:role/execution_role \
  --runtime provided.al2 \
  --code S3Bucket=buildkite-lambdas,S3Key=buildkite-agent-metrics/v5.10.0/handler.zip \
  --handler handler

Building and running the container

You can build a Docker image for the buildkite-agent-metrics from a clone of the source repo with the following:

docker build -t buildkite-agent-metrics .

This will create a local docker image named as buildkite-agent-metrics that you can tag and push to your own registry.

You can use the command-line arguments in a docker execution in the same way as described before:

docker run --rm buildkite-agent-metrics \
  -token abc123 \
  -interval 30s \
  -queue my-queue

Supported command line flags

$ buildkite-agent-metrics --help
Usage of buildkite-agent-metrics:
  -backend string
        Specify the backend to use: cloudwatch, newrelic, prometheus, stackdriver, statsd, opentelemetry (default "cloudwatch")
  -cloudwatch-dimensions string
        Cloudwatch dimensions to index metrics under, in the form of Key=Value, Other=Value
  -cloudwatch-region string
        AWS Region to connect to, defaults to $AWS_REGION or us-east-1
  -debug
        Show debug output
  -debug-http
        Show full http traces
  -dry-run
        Whether to only print metrics
  -endpoint string
        A custom Buildkite Agent API endpoint (default "https://agent.buildkite.com/v3")
  -interval duration
    	  Update metrics every interval, rather than once
  -cloudwatch-high-resolution
        If `-interval` is less than 60 seconds send metrics to CloudWatch as [High-Resolution Metrics](https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/publishingMetrics.html#high-resolution-metrics) which incurs additional charges.
  -max-idle-conns int
        Maximum number of idle (keep-alive) HTTP connections for Buildkite Agent API. Zero means no limit, -1 disables connection reuse. (default 100)
  -newrelic-app-name string
        New Relic application name for metric events
  -newrelic-license-key string
        New Relic license key for publishing events
  -prometheus-addr string
        Prometheus metrics transport bind address (default ":8080")
  -prometheus-include-default-metrics
        Include Go runtime, process, and HTTP handler metrics in the Prometheus endpoint (default true)
  -prometheus-path string
        Prometheus metrics transport path (default "/metrics")
  -queue value
        Specific queues to process
  -quiet
        Only print errors
  -stackdriver-projectid string
        Specify Stackdriver Project ID
  -statsd-host string
        Specify the StatsD server (default "127.0.0.1:8125")
  -statsd-tags
        Whether your StatsD server supports tagging like Datadog
  -timeout int
        Timeout, in seconds, for HTTP requests to Buildkite API (default 15)
  -token value
        Buildkite Agent registration tokens. At least one is required. Multiple cluster tokens can be used to gather metrics for multiple clusters.
  -version
        Show the version

Backends

By default metrics will be submitted to CloudWatch but the backend can be switched to other systems using the -backend argument.

CloudWatch

The CloudWatch backend supports the following arguments:

• -cloudwatch-dimensions: A optional custom dimension in the form of Key=Value, Key=Value

StatsD (Datadog)

The StatsD backend supports the following arguments:

• -statsd-host HOST: The StatsD host and port (defaults to 127.0.0.1:8125).
• -statsd-tags: Some StatsD servers like the agent provided by Datadog support tags. If specified, metrics will be tagged by queue otherwise metrics will include the queue name in the metric. Only enable this option if you know your StatsD server supports tags.

Prometheus

The Prometheus backend supports the following arguments:

• -prometheus-addr: The local address to listen on (defaults to :8080).
• -prometheus-include-default-metrics: Whether to include Go runtime, process, and HTTP handler metrics (defaults to true). Use -prometheus-include-default-metrics=false to expose only Buildkite metrics.
• -prometheus-path: The path under prometheus-addr to expose metrics on (defaults to /metrics).

Stackdriver

The Stackdriver backend supports the following arguments:

• -stackdriver-projectid: The Google Cloud Platform project to report metrics for.

New Relic

The New Relic backend supports the following arguments:

• -newrelic-app-name: String for the New Relic app name
• -newrelic-license-key: The New Relic license key. Must be of type INGEST

OpenTelemetry

The OpenTelemetry backend allows you to send metrics and traces to any OpenTelemetry-compatible using OTLP.

Configuration

Command Line Flags:

• --backend opentelemetry: Select OpenTelemetry as the metrics backend

Environment Variables:

• OTEL_SERVICE_NAME: OpenTelemetry service name (default: buildkite-agent-metrics)
• OTEL_EXPORTER_OTLP_PROTOCOL: Protocol to use: http/protobuf or grpc (default: http/protobuf)
• OTEL_EXPORTER_OTLP_ENDPOINT: Endpoint to send metrics to (default: http://localhost:4318 for http/protobuf and http://localhost:4317 for grpc)
• OTEL_EXPORTER_OTLP_HEADERS: Headers to send with each request (default: none)

See OpenTelemetry SDK documentation for more complete list of supported environment variables.

• https://opentelemetry.io/docs/specs/otel/configuration/sdk-environment-variables/
• https://opentelemetry.io/docs/specs/otel/protocol/exporter/#endpoint-urls-for-otlphttp

Usage Example

export OTEL_SERVICE_NAME="buildkite-metrics"
export OTEL_EXPORTER_OTLP_ENDPOINT="https://your-otlp-endpoint.com:4317"
export OTEL_EXPORTER_OTLP_HEADERS="authorization=your-api-key"
export OTEL_EXPORTER_OTLP_PROTOCOL="grpc"

buildkite-agent-metrics \
  --backend opentelemetry \
  --token $YOUR_BUILDKITE_TOKEN \
  --interval 30s

What Gets Sent

Metrics: The following metrics are exported to OpenTelemetry:

• buildkite.jobs.scheduled: Number of scheduled jobs
• buildkite.jobs.running: Number of running jobs
• buildkite.jobs.unfinished: Number of unfinished jobs
• buildkite.jobs.waiting: Number of waiting jobs
• buildkite.agents.idle: Number of idle agents
• buildkite.agents.busy: Number of busy agents
• buildkite.agents.total: Total number of agents
• buildkite.agents.busy_percentage: Percentage of busy agents
• buildkite.collection.duration: Time taken to collect metrics

All metrics include attributes for:

• org: Buildkite organization
• cluster: Buildkite cluster name
• queue: Queue name (for per-queue metrics)

Traces: Distributed tracing is provided for:

• Metrics collection operations
• HTTP requests to the Buildkite API
• Backend metric publishing

Each trace includes relevant attributes such as organization, cluster, and queue information.

Logs: Structured logging is integrated with OpenTelemetry:

• Error events are recorded as span events with error details
• Info events provide operational context
• All logs include relevant attributes and are correlated with traces

Troubleshooting

Check if OpenTelemetry is enabled: Look for this log message on startup:

OpenTelemetry backend initialized successfully

Compatibility:

• OpenTelemetry backend can be used as a complete replacement for other backends
• Supports both HTTP and gRPC OTLP protocols
• Works with any OpenTelemetry-compatible system (Jaeger, Honeycomb, HyperDX, Grafana Cloud, etc.)
• No impact on other backend functionality when not selected

Upgrading from v2 to v3

1. The -org argument is no longer needed
2. The -token argument is now an Agent Registration Token — the same used in the Buildkite Agent configuration file, and found on the Buildkite Agents page.
3. Build and pipeline metrics have been removed, focusing on agents and jobs by queue for auto–scaling. If you have a compelling reason to gather build or pipeline metrics please continue to use the previous version or open an issue with details.

Development

This tool is built with Go 1.20+ and assumes Go Modules by default.

You can build and run the binary tool locally with Go installed:

go run *.go -token [buildkite agent registration token]

Currently this will publish metrics to Cloudwatch under the custom metric prefix of Buildkite, using AWS credentials from your environment. The machine will require the cloudwatch:PutMetricData IAM permission.

The token package

It is an abstraction layer enabling the retrieval of a Buildkite Agent API token from different sources.

The current supported sources are:

• AWS Systems Manager (a.k.a parameter store).
• AWS Secrets Manager.
• OS environment variable.

Tests

All the tests for AWS dependant resources require their corresponding auto-generated mocks. Thus, before running them, you need to generate such mocks by executing:

go generate token/secretsmanager_test.go
go generate token/ssm_test.go

Metrics

The following metrics are gathered when no specific queue is supplied:

Buildkite > (Org) > RunningJobsCount
Buildkite > (Org) > ScheduledJobsCount
Buildkite > (Org) > UnfinishedJobsCount
Buildkite > (Org) > WaitingJobsCount
Buildkite > (Org) > IdleAgentsCount
Buildkite > (Org) > BusyAgentsCount
Buildkite > (Org) > BusyAgentPercentage
Buildkite > (Org) > TotalAgentsCount

Buildkite > (Org, Queue) > RunningJobsCount
Buildkite > (Org, Queue) > ScheduledJobsCount
Buildkite > (Org, Queue) > UnfinishedJobsCount
Buildkite > (Org, Queue) > WaitingJobsCount
Buildkite > (Org, Queue) > IdleAgentsCount
Buildkite > (Org, Queue) > BusyAgentsCount
Buildkite > (Org, Queue) > BusyAgentPercentage
Buildkite > (Org, Queue) > TotalAgentsCount

When a queue is specified, only that queue's metrics are published.

We send metrics for Jobs in the following states:

• Scheduled: the job hasn't been assigned to an agent yet. If you have agent capacity, this value should be close to 0.
• Waiting: the job is known to exist but isn't schedulable yet due to dependencies, wait statements, etc. This information is mostly useful to an autoscaler, since it represents work that will start soon.
• Running: the jobs that are currently in the "running" state. These jobs have been picked up by agents and are actively being executed.
• Unfinished: the jobs that have been scheduled but have not yet finished. This count includes jobs that are in the states "running" and "scheduled".

Detailed explanations about some of the metrics:

RunningJobsCount: This metric counts the number of jobs that are currently in the "running" state. These jobs have been picked up by agents and are actively being executed.

UnfinishedJobsCount: This metric includes all jobs that have been scheduled but have not yet finished. This includes jobs in the "running" and "scheduled" states.

ScheduledJobsCount: This metric counts jobs that have been scheduled but are not yet started by any agent. These jobs are in the queue, waiting for an available agent to start executing them.

WaitingJobsCount: This metric counts jobs that are in a "waiting" state, which could mean they are waiting on dependencies to resolve, on other jobs to finish, or on any other condition that needs to be met before they can be scheduled.

License

See LICENSE.md (MIT)