Skip to main content

Monitoring & Performance

A Feldera instance and its pipelines can be monitored using various tools. This tutorial will guide you through setting up monitoring for your Feldera instance.

Metrics with Grafana and Prometheus

Metrics are helpful to check the health of your Feldera instance and to identify resource bottlenecks. Feldera exposes a metrics endpoint that can be scraped by Prometheus. Grafana is then used to visualize these metrics.

Setup

  1. Prometheus: You must have Prometheus installed.
  2. Connect Prometheus to Feldera:
    • Add the following to your prometheus.yml configuration file, usually located in /etc/prometheus/prometheus.yml:
    - job_name: 'feldera'
    scrape_interval: 1s
    static_configs:
    - targets: ['localhost:8081']
    • Replace localhost:8081 with the address of your Feldera instance. Note that 8081 is the default port for metrics.
    • Restart Prometheus.
  3. Grafana: You must have Grafana installed.
  4. Add Prometheus To Grafana:

Setup with Docker

Alternatively, with docker, you can avoid installing Prometheus and Grafana to your local machine. To run Feldera with both Prometheus and Grafana:

docker compose -f deploy/docker-compose.yml up grafana --force-recreate

This spins up:

  1. Feldera
  2. Prometheus
  3. Grafana

If you want to run Prometheus without Grafana:

docker compose -f deploy/docker-compose.yml up prometheus --force-recreate

Set up the monitoring Dashboard

  1. **Copy the JSON of the Feldera template dashboard **
  2. Import the dashboard into Grafana
    • Under Dashboards, click the "New" icon, then click "Import Dashboard".
    • Paste the JSON copied from the template in the text-box and click "Load".

A Feldera Instance Monitoring dashboard will be created in Grafana. The dashboard is a template and may need to be adjusted to fit your specific needs. Look for the feldera_* metrics in Grafana to add more metrics to the dashboard.

Tracing with Jaeger

A Feldera pipeline can be traced using Jaeger. Tracing is useful if you need to analyze throughput or latency bottlenecks of your pipeline, as it instruments every step of the pipeline execution and provides detailed information about the execution time of each step.

Setup

  1. Jaeger: You must have Jaeger installed.
  2. Start Jaeger using the jaeger-all-in-one script:
    ./jaeger-all-in-one
  3. Tracing a Feldera pipeline
    • Enable/disable tracing and specify the Jaeger endpoint in the pipeline runtime_config:
    curl -i -X PATCH http://localhost:8080/v0/pipelines/tracing-pipeline \
    -H 'Content-Type: application/json' \
    -d '{
    "runtime_config": {"tracing": true, "tracing_endpoint_jaeger": "host.docker.internal:6831", <other config settings> }
    }'
    • Use host.docker.internal:6831 if you are running Feldera in docker or 127.0.0.1:6831 if you run Feldera directly.
    • Make sure to specify other settings you changed from non-default values in the runtime_config as well. runtime_config can be retrieved with:
    curl -X 'GET' \
    'http://127.0.0.1:8080/v0/pipelines/tracing-pipeline' \
    -H 'accept: application/json'

DBSP Profiles

A DBSP profile is a graph of the pipeline's circuit where each node represents an operator and each edge represents a data flow between operators. The profile includes information about how much CPU time or memory each operator consumes.

The API endpoint /v0/<pipeline_name>/dump_profile can be used to download the DBSP profile of a running pipeline. It returns a zip file containing multiple profiles (one for each worker) as .dot files, and a Makefile to transform the .dot files into .pdf files.