Time series
- A metric is a
time series - a stream of (timestamp, value) numbers
- Every unique combination of
name + labels is a different time series
- Prometheus collects and stores its metrics as
time series data, i.e. metrics information is stored with the timestamp at which it was recorded, alongside optional key-value pairs called labels
# time series name: http_requests_total
# time series labels: method="GET", status="200", service="checkout"
http_requests_total{method="GET", status="200", service="checkout"}
Metric types
Counter
- Cumulative metric (cannot decrease)
- Value can reset to zero
- You never read the raw value; you compute its rate:
rate(http_requests_total[5m] = requests/sec).
http_requests_total{method="GET", status="200", service="checkout"}
Gauge
- Single numerical value
- Can go up and down
- E.g., temperature, memory usage
Histogram
- Sample observations
- Histogram is cumulative
- E.g., request durations, response sizes
- An histogram generates 3 metrics
<metric_name>_sum<metric_name>_count<metric_name>_bucket: the number of measurement lower or equal to the value specified for each bucket. An +Inf bucket is automatically defined.
my_metric_bucket{le="1080"} # number of measurements lower or equal to 1080
my_metric_bucket{le="+Inf"} # total number of measurements (no upper bound)
my_metric_count # total number of measurements (always the same as my_metric_bucket{le="+Inf"})
my_metric_sum # sum of all measurements (the raw measurement value, not the highest fitted bucket)
Summary
- Similar to histogram, but percentiles are precomputed client-side (less flexible, histograms are usually preferred)