ridiculytics

Privacy-first web analytics that stores nothing and speaks Prometheus.

One ~15 MB Go binary. No database, no disk writes, no cookies, no persistent identifiers. A 1 KB script posts events, they land in bounded in-memory counters, and they leave as Prometheus metrics. Your TSDB is the storage layer and Grafana is the dashboard — neither is something extra to run if you already have monitoring.

Free forever, and structurally so: no hosted tier, no paid build, nothing behind a licence check.

flowchart LR
    js["counter.js<br/>~1 KB, in the page"]

    subgraph collector["ridiculytics"]
        ingest[":8080 ingest<br/>public"]
        mem["in-memory<br/>counters"]
        metrics[":9090 /metrics<br/>private"]
        ingest --> mem --> metrics
    end

    prom[("Prometheus")]
    graf["Grafana"]

    js -->|"POST /api/event"| ingest
    prom -->|scrape| metrics
    prom --> graf

Quick start

docker compose -f docker-compose.e2e.yml up

Ridiculytics, Prometheus and Grafana with a provisioned dashboard, on :8080, :9091 and :3000. Local only — no TLS, no auth.

Add to your site:

<script defer
        src="https://cdn.jsdelivr.net/npm/ridiculytics@1/web/counter.min.js"
        data-host="https://stats.example.com"
        data-site="example.com"></script>

data-host is mandatory and there is no fallback endpoint in the source, so a fork that forgets it is a silent no-op rather than a script quietly reporting to somebody else's server.

For a real deployment — ridiculytics + nginx + optional certbot, configured entirely through environment variables — see docs/deployment.md.

Two things to know before you adopt it

It stores marginals, not a cube. Each dimension gets its own metric family carrying at most two labels: its own, plus path. You can slice any dimension by page — the correlation people actually ask for. You cannot slice by anything else, and there are no three-way slices. That trade is what keeps series counts linear instead of multiplicative, and it is permanent.

Unique visitors come from sketches, and sketches do not merge. Uniques exist only for the windows the server computes (1h, 24h, 7d, 30d). Prometheus cannot merge HyperLogLog sketches, so uniques over an arbitrary range are not derivable. A rotating in-memory salt means no IP is ever stored or exported.

Both are inherent to in-memory plus Prometheus, not gaps to close later. If you need free n-way slicing or funnels, you want an event store and a columnar database — a different project with a very different operational cost.

What you get in exchange is the reason this exists: analytics in the same TSDB as your infrastructure metrics. Pageviews and p99 latency on one Grafana row. An Alertmanager rule that fires when a deploy takes traffic to zero.

Documentation

| | | |---|---| | Deployment | Production stack, TLS, ports, GeoIP, abuse resistance | | Configuration | Every environment variable, per-site overrides, YAML | | Metrics | Metric families, label sentinels, PromQL recipes | | Cardinality | The one knob that decides your series count | | Contributing | Tests, releases, project layout |

Privacy

No cookies, no localStorage, no fingerprinting, no cross-site identifier. The visitor hash is salted with a rotating in-memory secret and is not linkable across sites or across rotations. Referrers are reduced to a bare host. IP addresses live for the duration of one request handler and are never stored, logged, or exported.

This is the usual basis for arguing no cookie banner is required. It is not legal advice, and your regulator is not us.

Licence

web/counter.js is MIT — it gets embedded in other people's pages, where anything copyleft would be a non-starter. Everything else is AGPL-3.0, which is the licence that stops someone wrapping this in a closed hosted product.