dpm-finder

name	dpm-finder
license	Apache-2.0
description	Grafana Professional Services tool for identifying which Prometheus metrics drive high Data Points per Minute (DPM). Analyzes metric-level DPM with per-label breakdown to help optimize Grafana Cloud costs. Use when the user asks about DPM analysis, high-cardinality metrics, metric cost optimization, finding noisy metrics, or running dpm-finder against a Grafana Cloud Prometheus endpoint.

dpm-finder

A Grafana Professional Services tool for identifying which Prometheus metrics drive high Data Points per Minute (DPM). Analyzes metric-level DPM with per-label breakdown to help optimize Grafana Cloud costs.

Source: https://github.com/grafana-ps/dpm-finder

Quick Start

Prerequisites

• Python 3.9+
• Access to a Grafana Cloud Prometheus endpoint (or any Prometheus-compatible API)

Setup

1. Clone the repo and create a virtual environment:

git clone https://github.com/grafana-ps/dpm-finder.git
cd dpm-finder
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Configure credentials by copying .env_example to .env and filling in values:
   • PROMETHEUS_ENDPOINT -- The Prometheus endpoint URL (must end in .net, nothing after)
   • PROMETHEUS_USERNAME -- Tenant ID / stack ID (numeric)
   • PROMETHEUS_API_KEY -- Grafana Cloud API key (glc_... format)

Stack Discovery with gcx

If gcx is available, use it to find stack details:

gcx config check              # Show active stack context
gcx config list-contexts      # List all configured stacks
gcx config view               # Full config with endpoints

The Prometheus endpoint follows the pattern:

https://prometheus-{cluster_slug}.grafana.net

The username is the numeric stack ID. gcx auto-discovers service URLs from the stack slug via GCOM.

Stack Discovery without gcx

Look up the stack in the Grafana Cloud portal, or query the usage datasource:

grafanacloud_instance_info{name=~"STACK_NAME.*"}

Extract cluster_slug for the endpoint URL and id for the username.

Running the Tool

One-Shot Analysis (primary use case)

./dpm-finder.py -f json -m 2.0 -t 8 --timeout 120 -l 10

CLI Flags Reference

Flag	Default	Description
-f, --format	csv	Output format: csv, text, txt, json, prom
-m, --min-dpm	1.0	Minimum DPM threshold to include a metric
-t, --threads	10	Concurrent processing threads
-l, --lookback	10	Lookback window in minutes for DPM calculation
--timeout	60	API request timeout in seconds
--cost-per-1000-series	(none)	Dollar cost per 1000 series; adds estimated_cost column
-q, --quiet	false	Suppress progress output
-v, --verbose	false	Enable debug logging
-e, --exporter	false	Run as Prometheus exporter instead of one-shot
-p, --port	9966	Exporter server port
-u, --update-interval	86400	Exporter metric refresh interval in seconds

Output Formats

Output files are written to the current working directory.

JSON (-f json) -> metric_rates.json

Best for programmatic analysis. Includes per-series DPM breakdown:

• metrics[].metric_name -- the metric name
• metrics[].dpm -- data points per minute (maximum across this metric's individual series)
• metrics[].series_count -- number of active time series
• metrics[].series_detail[] -- per-label-set DPM breakdown (sorted by DPM descending)
• total_metrics_above_threshold -- count of metrics above threshold
• performance_metrics.total_runtime_seconds -- total processing time
• performance_metrics.average_metric_processing_seconds -- avg time per metric
• performance_metrics.total_metrics_processed -- total metrics analyzed
• performance_metrics.metrics_per_second -- processing throughput

CSV (-f csv) -> metric_rates.csv

Columns: metric_name, dpm, series_count (plus estimated_cost if --cost-per-1000-series is set).

Text (-f text) -> metric_rates.txt

Human-readable format with per-series breakdown and performance statistics.

Prometheus (-f prom) -> metric_rates.prom

Prometheus exposition format suitable for Alloy's prometheus.exporter.unix textfile collector.

Interpreting Results

• DPM = data points per minute (maximum across this metric's individual series)
• series_count = number of active time series for that metric
• series_detail (JSON/text only) = per-label-combination DPM breakdown
• Sort by DPM descending to find the noisiest metrics
• For top metrics, examine series_detail to identify which label combinations drive the highest DPM
• If --cost-per-1000-series is set, use estimated_cost to prioritize by spend

Rate Limiting

When running dpm-finder against multiple stacks, limit to max 3 concurrent runs. Batch the stacks and wait for each batch to complete before starting the next.

Metric Filtering

The tool automatically excludes:

• Histogram/summary components: *_count, *_bucket, *_sum suffixes
• Grafana internal metrics: grafana_* prefix
• Metrics with aggregation rules defined in the cluster (fetched from /aggregations/rules)

Exporter Mode

Run as a long-lived Prometheus exporter instead of one-shot analysis:

./dpm-finder.py -e -p 9966 -u 86400

Serves metrics at http://localhost:PORT/metrics. Recalculates at the configured interval (default: daily). See README.md for full exporter and Docker documentation.

Docker

Alternative to local Python setup:

docker build -t dpm-finder:latest .
docker run --rm --env-file .env -v $(pwd)/output:/app/output \
  dpm-finder:latest --format json --min-dpm 2.0

See README.md for full Docker Compose, production deployment, and monitoring integration docs.

Troubleshooting

Common Errors

• Authentication failures (401/403): Verify the API key is valid and has metrics:read scope. Confirm PROMETHEUS_USERNAME matches the numeric stack ID.
• Timeouts: Increase --timeout for large metric sets. The default is 60s; use 120s or higher for stacks with thousands of metrics.
• HTTP 422 errors: Usually means the metric has aggregation rules. The tool logs a warning and skips these automatically.
• Empty results: Lower the --min-dpm threshold. Check that PROMETHEUS_ENDPOINT does not have a trailing path after .net.
• Connection errors: Verify network connectivity to the Prometheus endpoint. The tool retries with exponential backoff (up to 10 retries).

Retry Behavior

The tool retries failed API requests with exponential backoff (up to 10 retries). Rate-limited responses (HTTP 429) are backed off automatically. HTTP 4xx errors other than 429 are not retried.

Project Structure

dpm-finder.py          # Main CLI tool (one-shot + exporter modes)
requirements.txt       # Python dependencies
.env_example           # Template for credential configuration
Dockerfile             # Multi-stage Docker build
docker-compose.yml     # Docker Compose orchestration
README.md              # Full project documentation