promcraft

A Python library for building Prometheus QL queries programmatically. Instead of constructing raw query strings, use composable Python objects that serialize to valid PromQL syntax.

Installation

pip install promcraft
# or with Poetry
poetry add promcraft

Requires Python 3.11+.

Quick Start

from promcraft import InstantVector, RangeVector, Label, Duration

# Simple metric selector
query = InstantVector("up", [])
str(query)  # "up{}"

# With label filters
query = InstantVector(
    "http_requests_total",
    [
        Label.eq("job", "api-server"),
        Label.eq("env", "production"),
    ],
)
str(query)  # 'http_requests_total{ job = "api-server", env = "production" }'

# Range vector with rate()
from promcraft import rate

query = rate(
    RangeVector(
        "http_requests_total",
        [
            Label.eq("job", "api-server"),
        ],
        Duration(m=5),
    )
)
str(query)  # 'rate(http_requests_total{ job = "api-server" }[5m])'

# Binary operation
left = InstantVector("http_requests_total", [Label.eq("status", "500")])
right = InstantVector("http_requests_total", [])
ratio = left / right
str(ratio)  # 'http_requests_total{ status = "500" } / http_requests_total{}'

Raw values instead of wrapper types

Anywhere a String is expected (label values, count_values, label-manipulation function arguments), you can pass a plain str instead — it's converted with String.from_value() under the hood. Anywhere a Scalar is expected (Float-typed function/aggregation parameters, at, binary-operator operands), you can pass a plain float/int instead, converted with Float.from_value(). Range, resolution and offset values are converted with Duration.from_value() instead, so offset=300 renders as offset 5m. Passing an existing String/Scalar instance through is a no-op, so both styles compose freely:

Label.eq("job", "api-server")  # same as Label.eq("job", String("api-server"))
topk(5, v)  # same as topk(Float(5), v)

API Reference

Scalars

Float

Wraps a Python float for use in queries.

from promcraft import Float

Float(3.14)  # "3.14"
Float(0.0)  # "0.0"
Float(-2.5)  # "-2.5"
Float(5)  # "5" -- an int stays an int

Float.from_value(3)  # Float(3) -- "3"
Float.from_value(3.0)  # Float(3.0) -- "3.0"
Float.from_value(Float(3.14))  # passthrough, still "3.14"

Hex

Represents a hexadecimal integer literal.

from promcraft import Hex

Hex(255)  # "0xff"
Hex(0)  # "0x0"
Hex(16)  # "0x10"

Hex.from_value() accepts whole floats (3.0 → 0x3) but raises ValueError for a fractional value such as 3.7 instead of truncating it.

Duration

Represents a Prometheus duration literal. All parameters are keyword-only.

from promcraft import Duration

Duration()  # "0s"  (default when no units given)
Duration(s=5)  # "5s"
Duration(m=1, s=30)  # "1m30s"
Duration(h=1, m=30)  # "1h30m"
Duration(ms=500)  # "500ms"
Duration(y=1, w=2, d=3, h=4, m=5, s=6, ms=7)  # "1y2w3d4h5m6s7ms"
Duration(m=5, neg=True)  # "-5m"

Fields must be non-negative; use neg=True for a negative duration (Duration(m=-5) raises ValueError).

Every Scalar subclass (Float, Hex, Duration) also implements a from_value() classmethod used throughout the library to accept raw numbers wherever a scalar is expected. Passing an existing Scalar through from_value() returns it unchanged.

String

Represents a PromQL string literal with configurable quoting.

from promcraft import String

String("hello")  # '"hello"'   (double-quote default)
String("hello", '"')  # '"hello"'
String("hello", "'")  # "'hello'"
String("hello", "`")  # "`hello`"

String.from_value("hello")  # same as String("hello")
String.from_value(String("hello", "'"))  # passthrough, still "'hello'"

Special characters are escaped automatically for " and ' quotes. Backtick strings are not escaped.

Variable

Represents a Grafana dashboard template variable ($name, or ${name} when braces are needed). Grafana interpolates the variable's selected value before the query text reaches Prometheus, so a Variable is accepted anywhere a query fragment, scalar, or string is expected — metric names, label names/values, offset/at/range/resolution, binary-operator operands, on/ignoring/group_left/group_right/by/without label lists, and aggregation/function arguments.

from promcraft import Variable

Variable("job")  # "$job"
Variable("job", braces=True)  # "${job}"

from promcraft import InstantVector, Label, sum_, topk

# dynamic metric name and label value
InstantVector(Variable("metric_name"), [Label.eq("env", Variable("environment"))])
# '$metric_name{ env = $environment }'

# dynamic offset and range
InstantVector("up", [], offset=Variable("offset_var"))[Variable("range_var")]
# 'up{}[$range_var] offset $offset_var'

# dynamic grouping labels
sum_(InstantVector("requests", [])).by([Variable("groupby")])
# 'sum(requests{}) by ($groupby)'

# dynamic aggregation parameter
topk(Variable("k"), InstantVector("requests", []))
# 'topk($k, requests{})'

A Variable is never coerced into a literal — Float.from_value(), Hex.from_value(), Duration.from_value(), and String.from_value() all pass an existing Variable through unchanged, the same way they already do for Scalar/String instances.

Label

Represents a label matcher used inside vector selectors. Use the factory methods for convenience. The value accepts either a raw str or a String instance.

from promcraft import Label, String

Label.eq("job", "prometheus")  # 'job = "prometheus"'
Label.neq("env", "prod")  # 'env != "prod"'
Label.re("name", "prom.*")  # 'name =~ "prom.*"'
Label.nre("name", String("test.*"))  # 'name !~ "test.*"'

InstantVector

Selects a set of time series at a single point in time.

InstantVector(metric, labels, *, offset=None, at=None)

from promcraft import InstantVector, Label, String, Duration, Float

# Basic
InstantVector("up", [])
# "up{}"

# With labels
InstantVector(
    "http_requests_total",
    [
        Label.eq("job", "api"),
        Label.re("status", "5.."),
    ],
)
# 'http_requests_total{ job = "api", status =~ "5.." }'

# With offset
InstantVector("up", [], offset=Duration(m=5))
# "up{} offset 5m"

# With @ modifier (Unix timestamp)
InstantVector("up", [], at=1609746000.0)
# "up{} @ 1609746000.0"

# With @ modifier (range function references)
InstantVector("up", [], at="start()")
# "up{} @ start()"

InstantVector("up", [], at="end()")
# "up{} @ end()"

RangeVector

Selects a range of samples over a time window. Required by functions like rate(), increase(), avg_over_time().

RangeVector(metric, labels, range, resolution=None, *, offset=None, at=None)

from promcraft import RangeVector, Label, String, Duration

# Basic range
RangeVector("http_requests_total", [], Duration(m=5))
# "http_requests_total{}[5m]"

# With resolution (subquery step)
RangeVector("up", [], Duration(m=5), resolution=Duration(s=30))
# "up{}[5m:30s]"

# With offset
RangeVector("up", [], Duration(m=5), offset=Duration(m=1))
# "up{}[5m] offset 1m"

# Combined
RangeVector(
    "http_requests_total",
    [Label.eq("job", "api")],
    Duration(m=5),
    resolution=Duration(s=30),
    offset=Duration(m=1),
    at="end()",
)
# 'http_requests_total{ job = "api" }[5m:30s] offset 1m @ end()'

InstantVector also supports subscript syntax as a shortcut for building a RangeVector that reuses its metric, labels, offset and @ modifier:

InstantVector("up", [])[Duration(m=5)]  # equivalent to RangeVector("up", [], Duration(m=5))
InstantVector("up", [])[Duration(m=5), Duration(s=30)]  # range + resolution

Subquery

A subquery evaluates any instant-vector expression over a range at a fixed resolution, producing a range vector for functions like max_over_time. Build one with .subquery() on any query, or construct Subquery directly. Leaving out resolution uses the global evaluation interval ([1h:]). Binary operations and negations are parenthesized automatically.

from promcraft import Duration, InstantVector, Subquery, max_over_time, rate, sum_

requests = InstantVector("http_requests_total", [])
per_second = rate(requests[Duration(m=5)])

per_second.subquery(Duration(h=1), Duration(m=1))
# "rate(http_requests_total{}[5m])[1h:1m]"

max_over_time(per_second.subquery(Duration(h=1)))
# "max_over_time(rate(http_requests_total{}[5m])[1h:])"

sum_(per_second).by(["job"]).subquery(Duration(h=1), Duration(m=1), offset=Duration(d=1))
# "sum(rate(http_requests_total{}[5m])) by (job)[1h:1m] offset 1d"

Subquery(requests + requests, Duration(h=1), at="end()")
# "(http_requests_total{} + http_requests_total{})[1h:] @ end()"

Slice syntax is a shorthand for .subquery(range, resolution):

per_second[Duration(h=1) : Duration(m=1)]  # "rate(http_requests_total{}[5m])[1h:1m]"
per_second[3600:]  # "rate(http_requests_total{}[5m])[1h:]"
requests[Duration(h=1) :]  # "http_requests_total{}[1h:]"

On an InstantVector, a slice builds a Subquery, while plain and tuple indexing still build a RangeVector. Slices take no offset/@; use .subquery() for those.

mypy users: mypy rejects non-integer slice bounds (Slice index must be an integer, SupportsIndex or None), so slices with Duration or Variable bounds need # type: ignore[misc]. Integer bounds ([3600:60]) and pyright work as-is. Prefer .subquery() in mypy-checked code.

PromQL only allows offset and @ on selectors and subqueries. To shift a function or aggregation, wrap it in a subquery or put the modifier on the selector inside it.

Binary operators

Combine two query expressions with an arithmetic, comparison, logical/set, or trigonometric operator. Each operator has a module-level helper function, and this is the primary way to build binary expressions:

from promcraft import (
    add,
    sub,
    mul,
    div,
    mod,
    pow,
    eq,
    neq,
    lt,
    lte,
    gt,
    gte,
    and_,
    or_,
    unless,
    atan2,
    Float,
    InstantVector,
)

v1 = InstantVector("requests", [])
v2 = InstantVector("errors", [])

add(Float(1.0), Float(2.0))  # "1.0 + 2.0"
div(v2, v1)  # "errors{} / requests{}"
gt(v1, Float(0.0))  # "requests{} > 0.0"
and_(v1, v2)  # "requests{} and errors{}"

Python's own operators also work directly on any Query (+ - * / % **, == != < <= > >=, & for and, | for or). Plain numbers are accepted on either side:

requests + errors  # same as add(requests, errors)
errors / requests * 100  # "(errors{} / requests{}) * 100"
1 - errors / requests  # "1 - (errors{} / requests{})"

Because comparisons build PromQL expressions, Query objects have no truth value: bool(q), if q: and q in some_list raise TypeError. Comparing with anything that isn't a query or a number (q == "foo", q != None) raises TypeError too, rather than silently returning False/True. Use is None / is instead.

Nested expressions

Nested binary expressions are automatically parenthesized to preserve evaluation order (function calls and aggregations are self-delimiting and are left unwrapped):

inner = add(Float(1.0), Float(2.0))
outer = mul(inner, Float(3.0))
str(outer)  # "(1.0 + 2.0) * 3.0"

bool modifier

Comparisons filter by default. .bool() makes them return 0/1 per element instead, which PromQL requires when both sides are scalars. It is only valid on comparison operators:

(v1 > 0).bool()  # "requests{} > bool 0"
(Float(1) > Float(2)).bool()  # "1 > bool 2"
gt(v1, v2).bool().on(["job"])  # "requests{} > bool on (job) errors{}"

Unary minus

-rate(v1[Duration(m=5)])  # "-rate(requests{}[5m])"
-(v1 + v2)  # "-(requests{} + errors{})"

In PromQL -2 ^ 2 means -(2 ^ 2), because ^ binds tighter than unary minus. promcraft therefore parenthesises a negative base of ^, whether it is a literal or a negation:

Float(-2) ** 2  # "(-2) ^ 2"
(-v1) ** 2  # "(-requests{}) ^ 2"

Label matching

Use .on() or .ignoring() to control which labels are used for matching when combining two vector selectors:

requests = InstantVector("http_requests_total", [])
errors = InstantVector("http_errors_total", [])

# Match on specific labels
expr = div(errors, requests).on(["job", "env"])
# 'http_errors_total{} / on (job, env) http_requests_total{}'

# Ignore specific labels
expr = div(errors, requests).ignoring(["instance"])
# 'http_errors_total{} / ignoring (instance) http_requests_total{}'

Grouping

Use .group_left() or .group_right() for many-to-one or one-to-many matching:

# group_left: result has labels from the left side
expr = mul(requests, errors).on(["job"]).group_left(["env"])
# '... on (job) group_left (env) ...'

# group_right: result has labels from the right side
expr = mul(requests, errors).group_right([])

.on(), .ignoring(), .group_left(), and .group_right() are all chainable and each returns a new immutable instance, so they can be combined freely:

mul(requests, errors).ignoring(["env"]).group_right(["env"])

Aggregation operators

Apply a PromQL aggregation operator to a vector via a module-level helper function. Helpers for operators that require a parameter (topk, bottomk, quantile, limitk, limit_ratio, count_values) take it as their first argument.

No-parameter aggregations — sum_, avg, min_, max_, count, group, stddev, stdvar:

sum, min, and max are named sum_, min_, max_ in this library so they don't shadow the Python builtins.

from promcraft import sum_, avg, min_, max_, count, group, stddev, stdvar, InstantVector

v = InstantVector("http_requests_total", [])

sum_(v)  # "sum(http_requests_total{})"
avg(v).by(["job"])  # "avg(http_requests_total{}) by (job)"
stddev(v).without(["env"])
# "stddev(http_requests_total{}) without (env)"

Use .by(labels) to keep only the listed labels in the grouped result, or .without(labels) to drop the listed labels and keep everything else. Both are chainable and return a new instance.

labels must be non-empty — .by([])/.without([]) raise ValueError. To aggregate away all labels, omit .by()/.without() entirely rather than passing an empty list.

Scalar-parameter aggregations — topk, bottomk, quantile, limitk, limit_ratio. The parameter accepts a raw float/int or a Float:

from promcraft import topk, bottomk, quantile, limitk, limit_ratio

topk(5, v)  # "topk(5, http_requests_total{})"
bottomk(3, v)  # "bottomk(3, http_requests_total{})"
quantile(0.95, v)  # "quantile(0.95, http_requests_total{})"
limitk(10, v)  # "limitk(10, http_requests_total{})"
limit_ratio(0.1, v)  # "limit_ratio(0.1, http_requests_total{})"
topk(5, v).by(["job"])
# "topk(5, http_requests_total{}) by (job)"

String-parameter aggregation — count_values (label name, accepts a raw str or a String):

from promcraft import count_values

count_values("version", v)
# 'count_values("version", http_requests_total{})'

Functions

Every Prometheus query function has a typed module-level helper. Scalar and string arguments accept raw float/int/str values interchangeably with Float/String.

Rate / counter functions

from promcraft import RangeVector, Duration, rate, irate, increase, delta, deriv, resets

rv = RangeVector("http_requests_total", [], Duration(m=5))

rate(rv)  # "rate(http_requests_total{}[5m])"
irate(rv)  # "irate(http_requests_total{}[5m])"
increase(rv)  # "increase(http_requests_total{}[5m])"
delta(rv)  # "delta(http_requests_total{}[5m])"
deriv(rv)  # "deriv(http_requests_total{}[5m])"
resets(rv)  # "resets(http_requests_total{}[5m])"

*_over_time aggregations

from promcraft import (
    avg_over_time,
    min_over_time,
    max_over_time,
    sum_over_time,
    count_over_time,
    stddev_over_time,
    first_over_time,
    last_over_time,
    present_over_time,
    absent_over_time,
    quantile_over_time,
)

avg_over_time(rv)  # "avg_over_time(http_requests_total{}[5m])"
quantile_over_time(0.95, rv)  # "quantile_over_time(0.95, http_requests_total{}[5m])"
first_over_time(rv)  # "first_over_time(http_requests_total{}[5m])"

Math & rounding

from promcraft import InstantVector, abs, ceil, floor, sqrt, round, clamp, clamp_min, clamp_max

v = InstantVector("cpu_usage", [])

abs(v)  # "abs(cpu_usage{})"
ceil(v)  # "ceil(cpu_usage{})"
sqrt(v)  # "sqrt(cpu_usage{})"
round(v)  # "round(cpu_usage{})"
round(v, 0.5)  # "round(cpu_usage{}, 0.5)"
clamp(v, 0.0, 1.0)  # "clamp(cpu_usage{}, 0.0, 1.0)"
clamp_min(v, 0.0)  # "clamp_min(cpu_usage{}, 0.0)"
clamp_max(v, 1.0)  # "clamp_max(cpu_usage{}, 1.0)"

Exponential & logarithm

from promcraft import exp, ln, log2, log10

exp(v)  # "exp(cpu_usage{})"
ln(v)  # "ln(cpu_usage{})"
log2(v)  # "log2(cpu_usage{})"
log10(v)  # "log10(cpu_usage{})"

Trigonometry

from promcraft import sin, cos, tan, asin, acos, atan, sinh, cosh, tanh, deg, rad, pi, sgn

sin(v)  # "sin(cpu_usage{})"
deg(v)  # "deg(cpu_usage{})"
rad(v)  # "rad(cpu_usage{})"
pi()  # "pi()"
sgn(v)  # "sgn(cpu_usage{})"

Date / time

All accept an optional instant vector; when omitted they default to vector(time()) in PromQL.

from promcraft import hour, minute, day_of_week, day_of_month, month, year, time

hour()  # "hour()"
hour(v)  # "hour(cpu_usage{})"
time()  # "time()"

Sorting

from promcraft import sort, sort_desc, sort_by_label, sort_by_label_desc

sort(v)  # "sort(cpu_usage{})"
sort_desc(v)  # "sort_desc(cpu_usage{})"
sort_by_label(v, "job")  # 'sort_by_label(cpu_usage{}, "job")'
sort_by_label_desc(v, "job", "env")
# 'sort_by_label_desc(cpu_usage{}, "job", "env")'

Label manipulation

from promcraft import label_join, label_replace

label_join(v, "addr", ":", "host", "port")
# 'label_join(cpu_usage{}, "addr", ":", "host", "port")'

label_replace(v, "job", "${1}", "job", "(.+)")
# 'label_replace(cpu_usage{}, "job", "${1}", "job", "(.+)")'

Type conversion

from promcraft import scalar, vector

scalar(v)  # "scalar(cpu_usage{})"
vector(1.0)  # "vector(1.0)"

Histogram functions

from promcraft import (
    histogram_quantile,
    histogram_fraction,
    histogram_count,
    histogram_sum,
    histogram_avg,
    histogram_stddev,
    histogram_stdvar,
)

histogram_quantile(0.95, v)  # "histogram_quantile(0.95, cpu_usage{})"
histogram_fraction(0.0, 1.0, v)  # "histogram_fraction(0.0, 1.0, cpu_usage{})"
histogram_count(v)  # "histogram_count(cpu_usage{})"

Prediction & smoothing

from promcraft import predict_linear, double_exponential_smoothing

predict_linear(rv, 3600.0)
# "predict_linear(http_requests_total{}[5m], 3600.0)"

double_exponential_smoothing(rv, 0.1, 0.5)
# "double_exponential_smoothing(http_requests_total{}[5m], 0.1, 0.5)"

Absence detection

from promcraft import absent, absent_over_time

absent(v)  # "absent(cpu_usage{})"
absent_over_time(rv)  # "absent_over_time(http_requests_total{}[5m])"

Low-level: Function

All of the helpers above build a Function instance internally. It isn't part of the top-level promcraft public API — import it from its submodule if you need to call a custom or newly-added Prometheus function that doesn't have a helper yet:

from promcraft import Function
from promcraft import InstantVector, Float

Function("my_custom_func", [InstantVector("up", []), Float(42.0)])
# "my_custom_func(up{}, 42.0)"

Composing Complex Queries

Every Query implements to_string() (which __str__ delegates to), so you can compose arbitrarily deep expressions and print them with str():

from promcraft import (
    InstantVector,
    RangeVector,
    Label,
    Duration,
    rate,
    div,
)

job_label = Label.eq("job", "api")

# rate of 5xx errors divided by rate of total requests
error_rate = rate(
    RangeVector(
        "http_requests_total",
        [
            job_label,
            Label.re("status", "5.."),
        ],
        Duration(m=5),
    )
)

total_rate = rate(RangeVector("http_requests_total", [job_label], Duration(m=5)))

ratio = div(error_rate, total_rate).on(["job"])
str(ratio)
# 'rate(http_requests_total{ job = "api", status =~ "5.." }[5m])
#   / on (job) rate(http_requests_total{ job = "api" }[5m])'

Pretty printing

Every to_string() accepts an optional indent argument to render multi-line, indented output instead of the single-line compact form. Pass an int (spaces per level) or a whitespace string (e.g. "\t"). A bool is rejected rather than treated as 1/0:

from promcraft import InstantVector, Duration, sum_

vec = InstantVector("metric", [])[Duration(m=1)]
expr = sum_(vec).by(["job", "env"])

print(expr.to_string(indent=4))
# sum(
#     metric{}[1m]
# ) by (job, env)

Omitting indent (or passing None) keeps the existing single-line output — str(query) is unaffected either way.

Development

# Install dependencies
poetry install

# Run tests (includes type checking and linting)
poetry run pytest

# Run only unit tests
poetry run pytest -o addopts=""

The test suite uses pytest with pytest-mypy for static type checking and pytest-ruff for linting. All checks run together by default.