Prometheus

Overview

Prometheus is a metrics database and scraper, and nothing else. It does not handle logs, traces or profiles, and being clear about that boundary is the most useful thing to know about it.

It pulls rather than receives: the server scrapes HTTP endpoints on a schedule. That single architectural decision shapes everything else, including why exporters exist and why short-lived jobs need a push gateway.

Its data model is multi-dimensional time series identified by a metric name plus label pairs, and PromQL, its query language, has become a standard others implement, including Azure Monitor and Grafana Cloud.

Scaling is deliberately simple rather than clever. Server nodes are autonomous with no distributed storage dependency; you scale by federating or remote-writing into another system rather than clustering.

Its documentation is refreshingly candid about the trade-off that follows. If you need complete accuracy, such as for per-request billing, the project says plainly that Prometheus is not a good choice.

Storage is a local time-series database organised into two-hour blocks with a write-ahead log, and the default retention is fifteen days, which catches people out when they expect a year of history.

Long-term retention means remote write into something else, with PromQL still evaluated locally after fetching the raw data back.

Alerting splits in two: the server evaluates rules, and Alertmanager, a separate component, handles deduplication, grouping, silencing and routing. On-call and incident response are explicitly other people's job.

Release discipline is published and predictable, with a new minor cycle beginning every six weeks, and a long-term support line for people who cannot move that fast.

It is genuinely foundation-governed, with a seven-member steering committee on staggered two-year terms, annual elections by a ranking method, and a cap of two seats per company.

Key features and capabilities

The same headings are used for every pipeline & infrastructure monitoring entry, so two tools can be read side by side.

What it collects

• Metrics only, with four core metric types; no logs, traces, profiles or synthetics
• Alerts are derived from metrics rather than ingested as events
• Exemplars link metric samples to trace identifiers, but traces themselves are not stored
• Its own docs say it is unsuited to cases needing complete accuracy, such as per-request billing

How data gets in

• Pull-based scraping of HTTP endpoints on a configured interval, the defining mechanism
• Exporters translate third-party systems into scrapeable endpoints, with a large ecosystem
• A push gateway exists solely for short-lived batch jobs that cannot be scraped
• Service discovery across Kubernetes and the major clouds, with Oracle Cloud added in 3.14
• Client libraries for direct instrumentation in the common languages

Storage and retention

• Its own local time-series database, grouping samples into two-hour blocks
• A write-ahead log allows the in-memory head block to be recovered after a restart
• Retention defaults to fifteen days if neither time nor size retention is configured
• Long-term storage means remote write to another system, with PromQL still evaluated locally
• No cardinality ceiling is published; it is bounded by host resources

Querying and dashboards

• PromQL, a functional query language built around the label dimensions
• A built-in expression browser rather than rich dashboards, which are deliberately Grafana's job
• Recording rules precompute expensive expressions into new series
• Federation lets one server scrape aggregated series from another

Alerting and incident response

• Alerting rules evaluated by the server against PromQL expressions
• Alertmanager, a separate Apache 2.0 component, handles deduplication, grouping, silencing and inhibition
• Routing to receivers is Alertmanager's job, which is where external tools plug in
• On-call and incident response are not part of the project at all
• Service level objective features are not published as built-in

Ecosystem and standards

• A large exporter ecosystem plus official client libraries across languages
• PromQL and the exposition and remote-write formats have become interoperability standards
• OpenTelemetry interoperability is in active development jointly with that community
• Managed Prometheus is sold by others, notably Azure Monitor and Grafana Cloud
• PromCon is the project's own community event

Where it runs and what it costs

• Self-hosted only; the project ships no managed service
• A single autonomous binary per node, with no distributed storage dependency
• Kubernetes operators exist in the community but none is published by the project itself
• Current stable is on the 3.x line, with a long-term support release for slower-moving estates
• A new minor release cycle begins every six weeks

Pricing

Open sourceFree, Apache 2.0

Free under Apache 2.0, with no paid tier, free-tier limits or trial, because none of those concepts apply. The foundation sells nothing. Your cost is the infrastructure you run it on, plus the operational work of sizing, sharding and retaining, which is the real trade-off against a managed service. If you want Prometheus without that work, managed Prometheus-compatible services are sold by Grafana Labs and Microsoft and are priced under those entries rather than this one.

Vendor pricing page →

Demos and videos

• Prometheus 3.0 overview from PromCon Webinar or talk · YouTube
• Prometheus Monitoring channel Channel or playlist · YouTube

About Cloud Native Computing Foundation

Prometheus was originally built at SoundCloud and joined the Cloud Native Computing Foundation in 2016 as its second hosted project after Kubernetes, graduating in August 2018. It is Apache 2.0 licensed and governed by a steering committee of seven serving staggered two-year terms, with an explicit contributor ladder from contributors through voting members to maintainers. Decisions default to consensus, with voting only when that fails, and a single company may hold at most two committee seats. Foundation statistics put it at over nineteen thousand contributors from more than three thousand organisations.

Founded 2016 · prometheus.io