Self-hosted analytics in 2026 is a three-horse race: Umami if you want the lightest thing that works, Plausible if you want speed at scale with a clean UI, Matomo if you want every feature Google Analytics ever had — on your own server. All three end the same problems: data leaving your infrastructure, consent-banner complexity, and GA4's interface. They differ in stack, appetite and depth, and those differences decide which server you need.
Stack and appetite
| Umami | Plausible CE | Matomo | |
|---|---|---|---|
| Stack | Node.js + PostgreSQL/MySQL | Elixir + ClickHouse + PostgreSQL | PHP + MySQL/MariaDB |
| Realistic idle RAM | ~512 MB total | 2–4 GB (ClickHouse is hungry) | 1–2 GB + MySQL sized to data |
| Cookieless by default | yes | yes | optional mode |
| Data depth | essentials + custom events | essentials + goals, funnels, props | everything: visitor logs, ecommerce, heatmaps*, A/B* |
| License | MIT | AGPL (CE) | GPL (core), * = paid plugins |
Umami: analytics as a sidecar
Umami's whole point is that analytics should not need its own infrastructure. One Node process, one Postgres database, ~512 MB — it happily shares a 2 GB VPS with the site it measures. Setup is a docker-compose file and a tracking script. You get pageviews, referrers, countries, devices, UTM breakdowns and custom events; you do not get funnels-of-funnels or per-user session replays. For a blog, a SaaS landing, a portfolio of small sites — this is the honest default, and MIT licensing means zero strings.
Plausible: the scale-with-grace option
Plausible Community Edition runs on Elixir with ClickHouse as the events store — the same class of column database that powers commercial analytics vendors. The practical consequence: dashboards stay sub-second at tens of millions of events where MySQL-based tools start crawling. The cost is baseline appetite: ClickHouse alone wants 2 GB+ to be happy, so plan a 4 GB machine minimum just for the analytics stack. If your properties do six-figure daily pageviews or you keep years of history hot, this architecture is why you pick Plausible.
Need a dedicated server?
Compare prices from top providers. Configure and order in minutes.
Matomo: when you actually need everything
Matomo (ex-Piwik, since 2007) is the only one of the three that genuinely replaces old-school Google Analytics feature-for-feature: individual visitor profiles, ecommerce tracking with product-level revenue, form analytics, tag manager, and (via paid plugins) heatmaps and session recordings. It is also the only one where GDPR needs configuration rather than defaults — full visitor logs are personal data. PHP+MySQL means it runs anywhere but scales the least gracefully: past a few million pageviews a month you will be tuning MySQL and archiving jobs. Pick it for feature need, not for fun.
Server sizing that matches reality
- Up to ~100k pageviews/mo: Umami on the VPS you already have. Done.
- Up to ~1M pageviews/mo: any of the three on a 4 vCPU / 8 GB VPS ($15/mo class). Matomo will want its MySQL on NVMe.
- Multi-million and history-heavy: Plausible/ClickHouse on a dedicated box — the $59 Xeon-E 2136 (6c/32 GB ECC/NVMe) in our catalog holds years of events for a network of sites; ClickHouse loves the RAM and honest disks.
- Agency running analytics for clients: one dedicated server + one Plausible/Matomo instance with multiple sites beats N micro-VPSes on cost and upkeep.
Whichever you choose, put the tracking endpoint behind your own domain (first-party) — adblock lists hit the vendors' default domains, and self-hosting's accuracy advantage comes precisely from serving the script yourself. For the reverse-proxy layer, our self-hosted stack notes apply unchanged.
Verdict
- Umami — default for individuals and small teams: lightest, MIT, zero drama.
- Plausible — high-traffic properties and long history: the ClickHouse architecture earns its RAM.
- Matomo — feature checklist requirements (ecommerce, visitor logs, tag manager) that the other two simply do not have.
Facts checked July 8, 2026. All three ship active releases; resource figures are steady-state observations, not vendor minimums.
