Stage 3 · Build

Observability & Networking

Writing Prometheus Exporters

Collector pattern, metrics types, and the promhttp package for exposing custom metrics.

7 min readGo for SREBuild

Exporter Anatomy

A Prometheus exporter exposes metrics about your system in the Prometheus format. It runs an HTTP server, collects metrics on demand, and returns them when Prometheus scrapes the /metrics endpoint.

Gominimal-exporter.go

22 linesLn 1, Col 1Go

prometheus.MustRegister registers metrics with the default registry. promhttp.Handler serves all registered metrics in the Prometheus exposition format. Prometheus scrapes /metrics to collect data.

Metrics Types

Prometheus has four metric types: counter, gauge, histogram, and summary. Each serves a different purpose and has different query patterns.

Type	Use Case	Example
Counter	Monotonically increasing value	requests_total, errors_total
Gauge	Value that can go up and down	temperature, goroutines, queue_size
Histogram	Distribution of values with buckets	request_duration_seconds
Summary	Distribution with precomputed quantiles	request_duration_seconds (client-side)

Godefining-all-metric-types.go

26 linesLn 1, Col 1Go

Counter: use rate() and increase(). Gauge: use directly or with delta(). Histogram: use histogram_quantile() and rate(). Summary: quantiles are precomputed but cannot be aggregated across instances.

Collector Pattern

The Collector interface lets you generate metrics on demand. Instead of pushing metrics, you compute them when Prometheus scrapes. This is ideal for metrics that are expensive to compute or depend on external state.

Gocustom-collector.go

35 linesLn 1, Col 1Go

Describe sends metric descriptors. Collect sends current metric values. Reset clears old label values before setting new ones. This pattern ensures metrics are fresh on every scrape and handles dynamic label values.

Labels and Dimensions

Labels add dimensions to metrics. They let you filter and aggregate by specific values. Use labels sparingly — high cardinality labels create performance problems.

Golabel-best-practices.go

21 linesLn 1, Col 1Go

Each unique label combination creates a separate time series. If you have 10 methods, 5 status codes, that is 50 time series. If you add user_id with 1M users, that is 50M time series. Keep label cardinality low.

promhttp Handlers

promhttp provides HTTP handlers for serving metrics. Use the default handler for basic setups, or configure it for multi-tenant or filtered metrics.

Gopromhttp-configurations.go

23 linesLn 1, Col 1Go

HandlerFor serves from a custom registry. Handler configures error handling and compression. InstrumentedHandler adds metrics about the metrics endpoint itself. Multi-tenant filtering lets you expose only relevant metrics per tenant.

Testing Exporters

Test exporters by gathering metrics programmatically and asserting on their values. This ensures your metrics are correct without scraping an HTTP endpoint.

Gotesting-metrics.go

29 linesLn 1, Col 1Go

testutil.GatherAndCompare collects metrics from the registry and compares them against the expected output. This catches issues like wrong metric names, missing labels, and incorrect values. It is the standard way to test Prometheus metrics.

Avoid creating metrics in the Collect method

Always register metrics during initialization. Never create new prometheus.Desc or prometheus.Metric objects in Collect. Collect should only update existing metric values and send them to the channel.

Mark this lesson complete to store local progress and unlock a cleaner resume path the next time you visit.