The correct option is By scraping application endpoints that serve metrics over HTTP.

Prometheus operates by periodically polling HTTP endpoints that expose metrics in the Prometheus exposition format or in OpenMetrics. Applications and exporters expose metrics on an HTTP path and Prometheus is configured with scrape targets and service discovery so it can collect those metrics at the configured intervals.

The pull model makes it easy to discover services and to control scrape frequency. For host and system metrics there are exporters such as node_exporter that read system information and then expose those metrics over HTTP for Prometheus to scrape.

By running a local agent on each monitored host that pushes metrics to the server is incorrect because Prometheus uses a pull based scrape workflow by default. There is a Pushgateway for short lived jobs that cannot be scraped but that is an optional exception rather than the normal collection method.

By relying on a managed monitoring agent such as Cloud Monitoring agent is incorrect because those agents are used to send metrics to specific cloud monitoring services and are not how Prometheus itself obtains metrics by default. You can integrate with other agents or exporters but Prometheus still scrapes endpoints it can reach over HTTP.

By querying host hardware interfaces directly to extract system metrics is incorrect because Prometheus does not directly query hardware interfaces. Instead exporters or local services collect hardware and OS metrics and expose them over HTTP for Prometheus to scrape.

Full AWS Practitioner Certification Question