slo-kit

Stop hand-rolling SLO alerts. Define SLOs as code and get correct multi-window burn-rate Prometheus alerts for free.

slo-kit turns raw Prometheus / OpenTelemetry metrics into:

• SLO definitions as code — SLI query + target + rolling window, in YAML or Python.
• Live error-budget & burn-rate tracking — consumed, remaining, and time-to-exhaustion.
• Correct multi-window / multi-burn-rate Prometheus alert rules — per the Google SRE workbook.

SLO math (error budgets, multi-window multi-burn-rate alerting) is subtle and commonly implemented incorrectly. slo-kit packages it correctly and reusably, so you define reliability targets and get paged on budget burn without hand-rolling brittle PromQL.

Install

pip install slo-kit
# or
docker run --rm ghcr.io/slo-kit/slo-kit --help

Quickstart (< 5 minutes)

1. Define an SLO (checkout_slo.yaml):

apiVersion: slo-kit/v1
name: checkout-availability
service: checkout
description: Checkout API availability
objective: 0.999          # three nines
window: 30d               # rolling 30-day window
sli:
  # {window} is substituted per alert/budget window
  good_query: 'sum(rate(http_requests_total{job="checkout",code!~"5.."}[{window}]))'
  total_query: 'sum(rate(http_requests_total{job="checkout"}[{window}]))'
labels:
  team: payments

2. Validate it:

slo-kit validate checkout_slo.yaml

3. Check live budget & burn rate:

slo-kit status checkout_slo.yaml --source http://localhost:9090

4. Generate correct multi-window burn-rate alerts:

slo-kit gen-alerts checkout_slo.yaml > checkout_rules.yaml

5. Gate deploys on remaining budget (CI):

slo-kit gate checkout_slo.yaml --source http://localhost:9090 --min-budget 0.1
# exits non-zero when < 10% budget remains

From Python

from slo_kit import load_spec, evaluate_status, generate_rules_yaml
from slo_kit import PrometheusSource

slo = load_spec("checkout_slo.yaml")
source = PrometheusSource("http://localhost:9090")

status = evaluate_status(slo, source)
print(f"budget remaining: {status.budget_remaining_pct:.1%}")
print(f"1h burn rate:     {status.burn_rate('1h'):.2f}x")

print(generate_rules_yaml(slo))   # Prometheus alerting rules

Why multi-window multi-burn-rate?

Alerting on a single window forces a bad trade-off: short windows page fast but cry wolf; long windows are stable but slow to fire and slow to reset. The SRE workbook fix is to require two windows at once — a long window that measures how much budget is burning, and a short window (≈1/12 of it) that confirms the burn is still happening right now:

slo-kit implements this in alerts/multiwindow.py and emits the equivalent Prometheus rules from alerts/prometheus_rules.py — validated against an explicit truth table and golden files.

Core concepts

• SLI — good_events / total_events from a metric query.
• SLO target — e.g. 0.999 over a rolling window (e.g. 30d).
• Error budget — allowed failures = (1 - target) × total; track consumed, remaining, remaining %, and time-to-exhaustion.
• Burn rate — observed_error_rate / (1 - target); a burn rate of 1 exactly exhausts the budget over the window, >1 is too fast.
• Multi-window multi-burn-rate alerts — fast + slow burn condition sets to balance fast detection against low false positives.

See the documentation for concepts, the full API, and examples.

Development

pip install -e ".[dev]"
ruff check . && ruff format --check .
mypy src/slo_kit
pytest --cov=slo_kit --cov-report=term-missing

License

Apache-2.0.