When do you need the Push-Gateway, and how does it fit the pull model?

For sources Prometheus can't scrape directly, such as network gear reachable only via SNMP or short-lived batch jobs. They push their metrics to the gateway, and Prometheus scrapes the gateway like any other exporter.

* Pushers on one side, a normal scrape on the other. *

Pull only works if the target runs long enough and can host an HTTP endpoint. Two cases break that:

• Devices you can't install an exporter on. A typical switch or router only speaks SNMP. A script queries it and pushes the values into the Push-Gateway.
• Short-lived jobs. A backup script that runs for 20 seconds may be finished before the next scrape, so it pushes its result (e.g. duration, success) instead.

The gateway holds the latest pushed values and exposes them on its own endpoint. Prometheus keeps pulling; it just pulls from a middleman. The payoff is that the whole data centre can be represented as metrics in Prometheus.

Go deeper:

• Prometheus — When to use the Pushgateway — the official advice, including when NOT to use it.
• Simple Network Management Protocol — Wikipedia — the protocol most network gear speaks instead of HTTP metrics.