TopolearnNearly every useful graph and alert on counters, from request rates to error ratios, starts with rate().
A counter's raw value says little; what matters is how fast it grows.
rate(c[5m]): the per-second average rate over the last 5 minutes. It automatically handles counter resets (a process restart drops the value to 0) and extrapolates to the window edges. Use it for graphs and alerts.irate(c[5m]): the per-second rate between the last two samples in the window. More responsive, but spiky; useful only for fast-moving, zoomed-in graphs.increase(c[1h]): the total increase over the window, equal torate(c[1h]) * 3600. Because of extrapolation the result can be a non-integer even for integer counters.
Choosing the range: it must contain at least two samples, and should be at least four times the scrape interval to tolerate a missed scrape. With 15s scrapes, [1m] is the minimum and [5m] is common. In Grafana, $__rate_interval picks a safe value automatically.
The golden rule: rate first, then aggregate. sum(rate(x[5m])) is correct; rate(sum(x)[5m:]) is wrong because summing counters from instances that restart at different times hides resets.
An error ratio is just two rates divided: errors per second divided by requests per second. Never use rate() on a gauge; use deriv() or delta() for gauges instead.
Example
# Requests per second per series, averaged over 5 minutes
rate(http_requests_total{job="api"}[5m])
# Total requests per second for the whole job (rate first, then sum)
sum(rate(http_requests_total{job="api"}[5m]))
# Errors in the last hour (may be fractional because of extrapolation)
increase(http_requests_total{job="api", status=~"5.."}[1h])
# Error ratio: 5xx per second / all requests per second
sum(rate(http_requests_total{status=~"5.."}[5m]))
/
sum(rate(http_requests_total[5m]))
# e.g. 0.0042 (0.42 % of requests fail)Remember
- rate() = per-second average increase, reset-safe
- Use a range of at least 4x the scrape interval
- Always rate first, then sum
Watch out for
- Applying rate() to a gauge such as memory usage; it's meant for counters only.