← Blog
A lavender abacus with cream beads and one orange bead on a pale background.
Published on
2026-11-16

Self-hosted analytics: plan the app, database, and upkeep

Self-hosted analytics means running the software that collects and reports your website’s usage on infrastructure you control. You choose where it runs and how it is configured, while your team takes responsibility for updates, access, capacity, and recovery.

That can suit a team with a clear need for control and someone able to operate the service. If nobody owns maintenance or a failed restore, a managed analytics service is usually the stronger starting point. Compare the work required to produce dependable reports, not just the subscription price against a VM bill.

Start with the website questions you need to answer

Write down the decisions the reports should support. A documentation site might need popular pages and referral sources. An agency might need separate access for each client. A small software company might want to know which pages lead to a completed signup. These needs imply different events, permissions, and reporting features.

Define what each event means before installing a tracker. A button click and a completed signup are different outcomes. Decide which one you will record, where the confirmation comes from, and how repeated actions will be handled. Collect only the fields needed for the question; free-form form content, account details, and access tokens do not belong in an analytics event.

Also define what your counts represent. Browsers, blockers, site configuration, and collection rules can affect what reaches the service. A recorded visitor is a tool-defined measurement, not a verified person. Check the chosen tool’s definitions before comparing its results with another service.

Follow a page view from browser to report

A website analytics deployment has several parts that can fail independently. The dashboard can remain available while new events stop arriving. Trace the whole path:

PartWhat it doesUseful check
Website trackerObserves configured page views or actions and sends requestsDoes it load on the intended pages, once per intended action?
Collection endpointReceives and processes the analytics requestsCan the browser reach the correct endpoint, and are requests accepted?
DatabaseStores the records needed for reportsAre new records available, with enough storage and suitable access controls?
DashboardQueries data and presents it to authorized usersDoes the expected event appear under the correct site, date range, and filters?

For example, Umami’s collection guide requires a tracker on the website and recommends checking browser network requests when data is missing. A successful dashboard login does not exercise that path. Test both initial page loads and navigation within a single-page application, if your site uses one.

Keep the reporting interface and database access restricted. The collector must accept the website traffic you intend to measure, but that does not require exposing database ports or making every report public. Plan the domain, HTTPS, routing, account access, and monitoring as part of the deployment.

Choose an edition and its operating requirements

Umami and Plausible Community Edition are useful examples because their deployment requirements differ. Treat them as candidates to assess against your needs, rather than assume that “self-hosted analytics” describes one interchangeable stack.

Umami’s installation documentation describes source and container options with PostgreSQL. Its supplied Compose setup includes the application and database. Follow the supported application and database versions for the release you select, and preserve the required configuration and secrets when moving or restoring it.

The official Plausible CE v3.2.1 Compose definition includes the application, PostgreSQL, and ClickHouse. That means a recovery plan must account for multiple data stores and their compatible versions. Do not assume that backing up only PostgreSQL captures the complete analytics service.

Read the Plausible Cloud and Community Edition comparison before choosing an edition. Features, release timing, filtering, and support differ. Check the exact feature your team needs, especially where a managed-service demonstration influenced the decision. Operating the software yourself does not automatically reproduce the managed offering.

Self-hosting also does not establish how private a deployment is. Review what the tracker sends, what the application retains, and what proxies and logs record. Umami’s metric definitions, for instance, distinguish analytics records from other infrastructure processing. Choose the collection settings, access rules, and retention policy deliberately; the location of the VM does not make those choices for you.

Size the service from traffic and retention evidence

Monthly page views are a useful starting point, but they do not describe the whole workload. Record custom-event volume, busy-period request rates, the number of websites, retention, and how many people query reports at once. A campaign spike and a long-range report can compete for resources even when the monthly average looks modest.

Begin with the selected edition’s documented requirements, then measure a representative pilot. Observe CPU use, memory, database growth, disk space, write latency, and dashboard response times. Include normal background work such as backups and retention jobs. Avoid turning a minimum-memory requirement into a promise about how much traffic a VM can serve.

Storage planning should include the live database, indexes, logs, temporary working space, and separately protected backups. Measure growth over a useful interval and test what happens near your alert thresholds. Do not assume that every event uses a fixed number of bytes across tools, versions, or configurations.

A single VM can simplify an initial deployment, but it also puts the application and local data in the same failure boundary. A separate or managed database changes the responsibilities and connection design. Choose based on acceptable downtime, data loss, operating capacity, and the selected application’s support for that arrangement.

Assign maintenance and rehearse recovery

Make ownership explicit before relying on the reports. Each task needs a named person or team, a way to detect failure, and a record of the last successful check.

ResponsibilityWhat the owner needs to maintain
CollectionA repeatable test after website releases, tracker changes, and routing changes
AccessAccount permissions, protected credentials, and removal of access when people leave
CapacityAlerts for storage pressure, collection errors, resource exhaustion, and slow reports
UpdatesA supported version record, release review, test environment, and recovery plan
BackupsProtected copies, completion monitoring, retention, and access independent of a failed VM
RecoveryA tested procedure that restores collection and useful historical reports within the agreed limits

Updates can change configuration and database schemas. Umami’s update instructions include release-specific requirements and migrations. Read the relevant release notes before upgrading, preserve a recoverable copy, and test the collection path afterward. Returning to an older application image may be insufficient after a schema change.

Set two recovery requirements: how much recent data you can lose and how long the service can be unavailable. Then test whether your actual backup and restore procedure meets them. A persistent disk or container volume retains data under some conditions; it is not an independently protected backup.

For the PostgreSQL part, our PostgreSQL backup and restore guide explains a logical backup exercise and its limits. A complete analytics recovery also needs the other data stores, application configuration, secrets, compatible software, and access rules. Test in an isolated environment so the exercise does not send reports or collect production traffic by mistake.

A successful restore should recover known historical data and accept a new controlled event. Measure the full process, including obtaining backups and preparing the destination. Reports cannot reconstruct events that were never collected unless your application has a separately designed, valid replay source.

Run an acceptance check before committing

Use a dedicated test website or a clearly separated test stream. Prepare known actions and expected results rather than infer success from an upward-moving chart. The following checks give the pilot a concrete finish line:

  • Collection: load a known page and trigger one configured event; follow each through the browser request and into the correct report.
  • Counting: repeat navigation and reloads, check for duplicate instrumentation, and confirm that the behavior matches your event definitions.
  • Separation: verify site boundaries, user permissions, report-sharing settings, and the handling of internal test traffic.
  • Data fields: inspect the actual payload and retained fields for unexpected URLs, query parameters, identifiers, or form content.
  • Failure detection: interrupt collection in the test environment and confirm that monitoring distinguishes missing data from a real traffic drop.
  • Recovery: restore a protected copy into an isolated instance, verify known history, and send a new test event before accepting the procedure.

Record failures as well as passes. If the pilot misses its recovery requirement or nobody can respond to an alert, resolve that before moving your reporting dependency. A managed service can be the better choice even when the self-hosted software has all the features you want.

Bring the workload to a CPU hosting decision

Website analytics is a practical CPU hosting use case. The relevant constraints are the application and database requirements, storage behavior, collection traffic, and report queries. A GPU is not the starting assumption for these deployments.

Compute with Hivenet provides CPU infrastructure for applications your team operates. Evaluate an instance against the selected software and confirm storage persistence, network access, region, capacity, and recovery arrangements for your intended use. Running analytics on a VM does not include a managed analytics service.

Bring your expected page-view and event volume, peak load, retention, database requirements, recovery targets, and operating owner when you discuss the workload with Hivenet. A measured pilot and a maintenance plan make that conversation more useful than a page-view estimate alone.

Your next workload belongs on Hivenet.

Pick one AI, compute, or storage workload and see the difference for yourself. Spin it up in minutes, or let our team map your fastest path to production.