langflow/deploy/observability/grafana-loki

Langflow on Grafana + Loki

Reference stack that ingests Langflow's structured JSON logs into Loki and visualizes them with a pre-provisioned Grafana dashboard.

Use this as a starting point. The compose file, Promtail config, and dashboard JSON are independent of the rest of deploy/ and can be lifted into any environment.

What you get

• Loki 3.2 on :3100
• Promtail 3.2 scraping a directory of Langflow log files
• Grafana 11.3 on :3000 with the Loki datasource and the Langflow Logs dashboard already provisioned

Prerequisites on the Langflow side

The dashboard expects Langflow to be running in JSON mode with service metadata set. At minimum:

LANGFLOW_LOG_ENV=container
LANGFLOW_LOG_FILE=/absolute/path/to/langflow/logs/langflow.log
LANGFLOW_SERVICE_NAME=langflow
LANGFLOW_VERSION=1.10.0
LANGFLOW_ENVIRONMENT=production

Promtail scrapes a directory of *.log files, so LANGFLOW_LOG_FILE must point at a file inside the directory you expose to Promtail as LANGFLOW_LOG_DIR (see Run). Set both to the same directory, otherwise Promtail watches an empty folder and the dashboard stays blank. Use an absolute path: LANGFLOW_LOG_FILE is resolved against Langflow's working directory, not this one.

In JSON mode the file is a single JSON stream: application logs and third-party stdlib loggers (uvicorn, sqlalchemy, httpx, langchain) are all rendered as JSON and run through PII redaction, so the json parse stage and the Stdlib intercept routing panel work against it directly. This stack scrapes a file, so LANGFLOW_LOG_FILE is required. If you instead run Langflow as a container, you can drop the file and scrape its stdout by swapping Promtail's static_configs file target for docker_sd_configs (same JSON, same labels).

See Logs and observability for the full list of environment variables (per-logger overrides, extra PII redaction keys, trace correlation, etc.).

Run

From this directory:

# Point Promtail at the directory that holds the file you set in
# LANGFLOW_LOG_FILE above. Must be the same directory. Defaults to the
# bundled ./logs (used by the quick smoke test below).
export LANGFLOW_LOG_DIR=/absolute/path/to/langflow/logs

docker compose up -d

Then open http://localhost:3000/d/langflow-prod-logs. Default credentials are admin / admin (override with GF_ADMIN_USER and GF_ADMIN_PASSWORD).

To stop:

docker compose down -v

Quick smoke test (no Langflow required)

To verify the stack end to end without running Langflow, write a sample record into the bundled ./logs directory and query Loki directly:

mkdir -p logs
echo '{"event":"smoke test","level":"info","logger":"langflow.api.run","timestamp":"2026-06-01T00:00:00Z","service":"langflow","environment":"production","version":"1.10.0"}' >> logs/langflow.log

docker compose up -d

# Give Promtail a few seconds to tail the file, then confirm the line reached Loki:
sleep 5
curl -sG 'http://localhost:3100/loki/api/v1/query_range' --data-urlencode 'query={job="langflow"}' | grep -q "smoke test" && echo "OK: log reached Loki"

What each dashboard panel proves

Panel	LogQL it runs
PII leak count (must be 0)	sum(count_over_time({job="langflow"} |~ "sk-do-not-leak|hunter2|Bearer xyz" [$__range]))
Errors with structured tracebacks	{job="langflow", level=~"error|critical"} |= "exception" | json
Redaction proof	{job="langflow"} |~ "\\*\\*\\*"
Stdlib intercept routing	{job="langflow", logger=~"uvicorn.*|sqlalchemy.*|httpx.*|langchain.*"}
Service / environment / version coverage	sum by (service, environment, version) (count_over_time({job="langflow"}[$__range]))
Log rate by level	sum by (level) (rate({job="langflow"}[1m]))
Log rate by logger	topk(10, sum by (logger) (rate({job="langflow"}[1m])))

Notes

• Promtail only promotes level, service, environment, version, logger to labels. High-cardinality fields (user_id, flow_id, trace_id) stay in the log body — query them with | json in LogQL.
• Replace Promtail with Grafana Alloy if you already standardize on it; the JSON parse stage maps 1:1.
• If your runtime ships logs through a different transport (Fluent Bit, Vector, OTLP), only the scrape side changes — the dashboard and label schema stay the same.