Skip to content

AGENTS.md

Latest commit

 

History

History
29 lines (15 loc) · 2.36 KB

AGENTS.md

File metadata and controls

29 lines (15 loc) · 2.36 KB

xk6-output-influxdb

k6 output extension that ships test metrics to InfluxDB v2. Exists as a separate extension because the v2 API has breaking changes that cannot coexist with the v1 output bundled in k6 core.

Architecture

The extension registers itself via k6's output extension system at init time. A registration shim at the repo root delegates to the implementation package.

Configuration consolidates from three sources in priority order: JSON config from k6 options, environment variables (all prefixed with K6_INFLUXDB_), and the URL passed via k6's output argument. Each layer overrides the previous. The bucket name is required; the address defaults to localhost:8086.

The output implementation buffers metric samples and flushes them to InfluxDB on a periodic interval. Flush runs in a goroutine pool bounded by a semaphore channel (default 4 concurrent writers). If a flush takes longer than the push interval, a warning is logged but no samples are dropped -- they queue for the next cycle.

Tags can be promoted to InfluxDB fields (making them non-indexable but queryable with more types) via a configuration option. Each promoted tag gets a type annotation (string/int/float/bool) parsed from a colon-delimited format. By default, VU number, iteration count, and URL are promoted to fields.

The tag-to-field extraction mutates the tags map in place -- it deletes promoted keys from the tags map and adds them to the values map. This is cached per unique tag set pointer for the duration of a single flush batch to avoid repeated conversion.

The InfluxDB client's internal logger is explicitly disabled at init time to prevent noisy output.

A docker-compose setup provides InfluxDB, Grafana, and k6 for local testing, with pre-provisioned Grafana dashboards for visualizing test results.

Gotchas

  • Flush errors are logged but silently swallowed. If InfluxDB is unreachable or rejects writes, metrics are lost with only a log line. There is no retry mechanism.

  • The semaphore channel blocks the periodic flusher goroutine when all concurrent write slots are occupied. Under sustained backpressure, this causes sample buffer growth in memory without bound.

  • The linter config is not checked in. It is downloaded from k6 core's master branch on first lint run. Do not commit it.

  • Tests mock the InfluxDB server with Go's httptest. No real InfluxDB instance is needed for unit tests.