How to read a PromQL query

PromQL looks dense the first time you meet it. A line like histogram_quantile(0.99, sum by (le, route) (rate(http_request_duration_seconds_bucket[5m]))) reads like one long word, and the instinct is to scan it left to right like a sentence. That’s the wrong direction. PromQL is a functional language, so the meaning flows from the innermost expression outward — the same way you’d evaluate a nested formula in maths. Once you read it inside-out, almost every query decomposes into the same four layers.

The four layers

Section 1 of 8 · ~1 min

Nearly every non-trivial PromQL expression is built from these, stacked from the inside out:

1. A selector — which series you start from.
2. A range — over what window of time (only when you need history, not an instant).
3. A function — what transformation you apply to those samples.
4. An aggregation — how you collapse many series into fewer.

Read them in that order and the query explains itself.

Layer 1: the selector

Section 2 of 8 · ~1 min

The core of any query is a metric selector: a metric name plus optional label matchers in braces.

http_requests_total{job="api", status=~"5.."}

This selects every series named http_requests_total where the job label equals api and the status label matches the regex 5.. (any 5xx code). The matchers are the important part:

• = exact match
• != not equal
• =~ regex match
• !~ regex does not match

On its own a selector returns an instant vector — one current sample per matching series. That distinction matters for everything that follows.

Layer 2: the range

Section 3 of 8 · ~1 min

Append a duration in square brackets and the selector becomes a range vector — every sample in that window, per series, not just the latest one.

http_requests_total{job="api"}[5m]

You can’t graph a range vector directly; it’s raw material. You hand it to a function that knows what to do with a window of samples. The classic example is rate:

rate(http_requests_total{job="api"}[5m])

rate looks at the counter’s samples over the last 5 minutes and returns the per-second average rate of increase. This is the single most common pattern in Prometheus, and it’s worth internalising why it exists: http_requests_total is a counter that only ever goes up (until a restart resets it), so its raw value is meaningless on a dashboard. The rate of change is what you actually care about. rate also transparently handles counter resets, which is why you should never compute rates by hand.

A short note on window sizing: the range ([5m]) should comfortably cover at least a few scrape intervals. Too short and you get noisy, gappy results; too long and you smooth away the spikes you were trying to catch.

Layer 3: functions

Section 4 of 8 · ~1 min

Functions transform vectors. The ones you’ll see constantly:

• rate(...) — per-second average rate of a counter over a range.
• irate(...) — instant rate from the last two samples; spikier, good for fast-moving graphs.
• increase(...) — total increase over the range (essentially rate × seconds).
• histogram_quantile(φ, ...) — estimates a quantile (e.g. p99) from histogram buckets.
• rate(...[5m]) > 0 style comparisons — filtering, covered below.

So rate(http_requests_total{job="api", status=~"5.."}[5m]) reads, inside-out, as: take the 5xx request counter for the api job, over a 5-minute window, and give me the per-second error rate, per series.

Layer 4: aggregation

Section 5 of 8 · ~1 min

A selector with a job and a status label can still match dozens of series — one per instance, per pod, per status code. Aggregation operators collapse them.

sum by (job) (rate(http_requests_total{status=~"5.."}[5m]))

sum by (job) adds up the per-series rates, keeping only the job label and discarding the rest. The result is one error-rate line per job. The two clauses to know:

• by (labels) — keep these labels, aggregate away everything else.
• without (labels) — aggregate away these labels, keep everything else.

Other aggregators follow the same grammar: avg, max, min, count, topk, quantile. The mental model never changes — combine many series into fewer, grouped by the labels I name.

Putting it together

Section 6 of 8 · ~1 min

Now the intimidating query from the top decomposes cleanly. Read it from the inside out:

histogram_quantile(
  0.99,
  sum by (le, route) (
    rate(http_request_duration_seconds_bucket[5m])
  )
)

1. http_request_duration_seconds_bucket[5m] — the latency histogram buckets, over 5 minutes.
2. rate(...) — per-second rate of each bucket, so resets and scaling are handled.
3. sum by (le, route) (...) — add the rates across instances, keeping le (the bucket boundary, required by the next step) and route.
4. histogram_quantile(0.99, ...) — estimate the 99th-percentile latency from those buckets, per route.

In plain English: the p99 request latency per route over the last 5 minutes. One layer at a time, it’s not dense at all.

A few traps worth knowing

Section 7 of 8 · ~1 min

• Aggregating before rate-ing. rate(sum(...)) is almost always a bug. Take the rate first, then sum — summing counters across resets gives nonsense. The correct shape is sum(rate(...)).
• Dropping le. histogram_quantile needs the le label intact, so your by (...) clause must include it.
• Comparisons filter, they don’t just colour. rate(...)[5m]) > 0 doesn’t return booleans — it drops every series where the condition is false. That’s how you build alert expressions.
• Instant vs range mismatch. Passing an instant vector where a function wants a range vector (or vice versa) is the most common parse error. If a function complains, check your brackets.

Decode any query in seconds

Section 8 of 8 · ~1 min

The inside-out method works on every PromQL expression you’ll meet, but pulling apart a deeply nested production query by hand is still tedious — and easy to get subtly wrong under pressure. That’s exactly what the PromQL Explainer is for: paste any Prometheus query and get a plain-English, layer-by-layer breakdown of its selectors, ranges, functions, aggregations and comparisons. Everything runs client-side, so your queries never leave the browser.

Next time a dashboard panel or alert rule leaves you squinting, don’t guess at it.

Explain a PromQL query →