Glossary

Alertmanager — The companion program that receives Prometheus's firing alerts and owns notification: routing, grouping, deduplication, silences, and delivery to receivers.

Annotation — Human-facing text attached to an alert rule (summary, description), templated with the alert's labels and value, rendered when the alert fires.

Block — An immutable two-hour chunk of stored samples in Prometheus's TSDB; blocks are compacted into larger ones over time.

Cardinality — The number of unique series, driven by unique label combinations. The main scaling limit of a Prometheus server.

Counter — A metric that only increases (and resets on restart). Use rate() before believing it means anything.

Datasource — Grafana's named connection to a system it queries — here, a Prometheus server.

Evaluation interval — How often Prometheus runs rule expressions (evaluation_interval).

Exporter — An adapter program that publishes some system's statistics as a Prometheus-format /metrics page.

Exposition format — The plain-text format of a /metrics page: HELP and TYPE comment lines, then name{labels} value samples.

Firing — An alert whose expression has stayed true past its for: duration, now handed to Alertmanager. Before that, it is pending; after the expression goes false, resolved.

Gauge — A metric that goes up and down: temperature, free space, current load.

Grafana — The dashboard layer. It stores no metrics; every panel is a live query to a datasource.

Head block — The TSDB's in-memory recent window, backed by the write-ahead log.

Histogram — A metric type that counts observations into le buckets (plus _sum and _count), enabling estimated percentiles via histogram_quantile().

Instance — The label Prometheus adds naming the scraped target's address.

Job — A named group of targets in scrape_configs, sharing scrape settings; added to every series as the job label.

Label — A key="value" pair on a series. Name + label set is a series' identity — and its cost.

node_exporter — The standard operating-system exporter for Linux/Unix machines.

Panel — One visualization on a Grafana dashboard, bound to one or more queries.

Pending — An alert whose expression is true but whose for: clock hasn't expired yet. The anti-flap state.

PromQL — Prometheus's query language: selectors, rate(), aggregations, functions.

promtool / amtool — The command-line checkers and debuggers shipped with Prometheus and Alertmanager (check config, check rules).

Pull model — Prometheus initiates every scrape; targets just serve a page. The opposite of agents pushing data in.

Range vector — A selector with a window — metric[5m] — meaning "samples from the last five minutes." What rate() consumes.

Receiver — An Alertmanager notification destination: webhook, email, chat, pager.

Recording rule — A rule that stores a query's result as a new series on a schedule, pre-paying an expensive expression.

Retention — How long samples are kept before whole blocks are deleted: 15 days by default, set by age and/or size flags.

Scrape — One fetch of one target's /metrics page; happens per target every scrape_interval.

Series — One metric name plus one unique label set, and all its timestamped samples. The atom of storage, memory, and cost.

Silence — A time-boxed Alertmanager mute matching labels: the planned "we know" for maintenance and known incidents.

TSDB — Prometheus's local time-series database: head block, WAL, and compacted blocks on disk.

up — The synthetic series Prometheus makes per target: 1 when the last scrape succeeded, 0 when it didn't. The cheapest health signal there is.