dpm-finder

dpm-finder

Grafana PS tool ranking Prometheus metrics by DPM with per-series breakdown. Source: https://github.com/grafana-ps/dpm-finder

Prerequisites

• Python 3.9+
• A Grafana Cloud Prometheus endpoint URL + numeric stack ID + API key (glc_…, metrics:read scope)

Common Workflows

One-shot analysis (most common)

# 1. Clone + venv + install
git clone https://github.com/grafana-ps/dpm-finder.git
cd dpm-finder
python3

If step 5 is empty, lower -m or confirm the endpoint URL has no trailing path after .net.

Discover stack details with gcx

If gcx is installed it can derive the endpoint + username:

The Prometheus endpoint pattern is https://prometheus-{cluster_slug}.grafana.net. Username is the numeric stack ID.

Without gcx: look up in the Grafana Cloud portal, or query grafanacloud_instance_info{name=~"STACK_NAME.*"} on the usage datasource.

Multi-stack runs

Limit to max 3 concurrent runs to avoid GCloud rate limits. Batch the stacks and wait for each batch before the next.

Interpreting results

• DPM = max data points per minute across that metric's series
• series_count = active time-series count for that metric
• series_detail[] (JSON / text only) = per-label-combination DPM breakdown — use this to spot the offending label
• Sort by DPM descending → noisiest metrics; combine with --cost-per-1000-series to prioritize by spend

Troubleshooting

• 401 / 403 — API key invalid or missing metrics:read; confirm PROMETHEUS_USERNAME is the numeric stack ID
• Timeouts — bump --timeout to 120+ for stacks with thousands of metrics
• HTTP 422 — metric has aggregation rules; tool warns + skips automatically
• Empty results — lower -m; verify endpoint has no trailing path
• Connection errors — exponential backoff retries up to 10 times; persistent failure usually = network/firewall

References

• references/cli.md — full flag reference, output-format details, exporter mode, Docker invocation, auto-exclusion rules, retry behavior

Resources

• dpm-finder GitHub
• gcx