dash-prometheus 2026.9.1

dash-prometheus

_{PROMETHEUS FOR PLOTLY DASH}
How often each Dash callback runs, and how long it takes.

dash-prometheus wraps every dash.callback in a counter and a duration histogram and serves them on a Prometheus endpoint mounted on the app's own server. No sidecar, no second port: one import and one call.

pip install dash-prometheus

import dash_prometheus  # must come before `import dash`
import dash

app = dash.Dash(__name__)
dash_prometheus.add_middleware(app)  # serves /metrics

add_middleware(app, slug="stats") serves /stats instead.

Metrics

file is the source file that defines the callback and callback is its function name. Histogram buckets are 1 ms, 10 ms, 100 ms, 1 s, 5 s, 10 s, 30 s, 1 min, 2 min and 5 min.

Multiple workers

Metrics use prometheus_client's multiprocess mode, so every gunicorn worker reports into one set of numbers. Set PROMETHEUS_MULTIPROC_DIR to choose the directory; without it, the package creates a fresh one under /tmp. Either way the directory is emptied on import, so point it at a path used for nothing else.

Known limits

• Import order matters: dash_prometheus patches dash.callback at import time and raises if dash is already loaded, rather than silently missing callbacks.
• Only callbacks registered through dash.callback are measured; app.callback and clientside callbacks are not.
• A callback that raises is not counted: the counter and histogram record completed calls.
• prometheus-client is pinned to 0.18.0.

Development

mise run setup     # cold start: toolchain, frozen deps, git hooks, verify
mise run check     # lint + format + tests: the definition of done

See AGENTS.md for the working agreement.

Releasing

Publishing runs from CI, not from a laptop: bump version in pyproject.toml, merge, then push a tag v<version>. .github/workflows/release.yml re-runs the full check, refuses a tag that disagrees with pyproject.toml, and publishes to PyPI with trusted publishing, so no API token is stored anywhere.

Licence

MIT