was thinking https://prometheus.io/docs/concepts/metric_types/#gauge makes more sense

With a gauge, you only get history as detailed as your metrics polling frequency. I think that a typical metrics polling frequency is a minute or more. That's fine if you want long-term data; it's not useful if you want detailed history of how the input batch size varied over, say, the last minute. So I think that the right kind of metric here depends on the goal (for debugging, I suspect that it is a histogram, possibly a sliding histogram like we use for dbsp_step_latency_seconds).

With a gauge, you only get history as detailed as your metrics polling frequency

agree my thinking was you already put it in samply for low-level debugging
whereas metrics is also used by users that are more interested in long term perf

With a gauge, you only get history as detailed as your metrics polling frequency

agree my thinking was you already put it in samply for low-level debugging whereas metrics is also used by users that are more interested in long term perf

That's reasonable.

So the gauge(s) would report the number of bytes/records in the most recent input batch, I guess.