Prometheus is pull-based rather than push-based. So, whatever the metrics source is, it must expose the metrics in Prometheus format, and Prometheus will periodically query them with HTTP request.

If directly exposing the metrics is not possible, the metrics source can push the metrics to some intermediate component which exposes the metrics in Prometheus format so that Prometheus can query them.

It seems this is the approach taken by the document you're referring to. The metrics are submitted from the source via PubSub to a Metrics Telemetry Converter pod running in the Kubernetes cluster, which exposes them in Prometheus format.

You then have to configure Prometheus to scrape the metrics from this pod, as you would configure it for any other job.