To avoid having too much labels combination and useless metrics. For instance a rejected request will never have a loading task created so the load_task_status label value will always be "not created" here. So I think it is better to have a separate metric for each save request status.

ack

Will Prometheus create a time-series for these extra combinations if we never report a value for them, though?

The problem with using different metric names is that, afaik, they can't be aggregated. And we probably want to aggregate accepted and rejected requests, to get the total number of handled requests. (and maybe pending requests too, to have the total number of requests received)

I think it's missing a trailing slash now

How about dropping the REJECTED_SAVE_REQUESTS_METRIC and PENDING_SAVE_REQUESTS_METRIC and merge them in the following gauge ?

Gauge(
    name=SAVE_REQUESTS_METRIC,
    documentation='Number of origin save requests',
    labelnames=['status', 'visit_type'],
    registry=SWH_WEB_METRICS_REGISTRY)

I think keeping the metric about accepted save requests is interesting as for instance it could be used to calculate the ratio of successful requests.

right

Can you calculate such a ratio in prometheus between different metrics?

Sure, you can write this query in PromQL:

sum(swh_web_accepted_save_requests{load_task_status="succeed"}) / sum(swh_web_submitted_save_requests)