muratbuker/exporter-litellm

This project is a Prometheus exporter for LiteLLM, a tool that routes requests to different AI language models. In plain terms, it reads usage and cost data out of the database that LiteLLM keeps and turns that data into metrics that monitoring tools like Prometheus and Grafana can display on dashboards. It is a fork of another exporter called ncecere/exporter-litellm, with the main change being how it organizes time periods for the numbers it reports. Instead of letting each person configure their own rolling time window for spend and usage stats, this exporter reports most metrics in three fixed windows: the last 30 days, the last 1 day, and the last 1 hour. Each version of a metric gets its own name, so a Grafana dashboard can pull exactly the time range it needs without extra calculation. The three windows also refresh on different schedules, so the fast moving hourly numbers update often while the slower 30 day totals are not recalculated as frequently, which keeps the load on the underlying Postgres database manageable. The metrics themselves cover things like total spend, spend broken down by user, team, key, or organization, token counts, request counts, error rates, budget usage, and information about active or expired API keys. Some data, like request duration, is reported as a single ongoing measurement since it is meant to be queried across any time range directly in Prometheus. To use this project, you connect it to your LiteLLM Postgres database by setting environment variables for the host, port, database name, and credentials, and the documentation recommends creating a dedicated read only database user for security. You can also configure how often each of the three metric tiers refreshes and which port the exporter listens on. This is a backend infrastructure tool aimed at people already running LiteLLM who want visibility into cost and performance through existing monitoring dashboards.