python-syntheticalert

Drive a synthetic alert metric from Python, so a Triple Pat check-in timer can verify your alerting pipeline end to end. Works with Prometheus and OpenTelemetry.

Why

A broken alerting pipeline looks exactly like a healthy system. No alerts might mean nothing is wrong, or it might mean your alerting is down, and your alerting system is the one thing that cannot alert you about itself.

This library provides a time-based callback to drive a synthetic alert metric. You register the callback as a gauge in your existing metrics setup, alert on the gauge like any other metric, and route the alert to a Triple Pat check-in timer. Every delivered alert then becomes a check-in, and every firing is another fire drill for the whole path from metric to notification. If the check-ins ever stop, your alerting pipeline is broken, and the Triple Pat app raises an alarm through a separate channel to tell you so. An example alert rule and Alertmanager route are below.

Usage

uv add triplepat-syntheticalert   # or: pip install triplepat-syntheticalert

The library has no dependencies and no background threads. It is a single callable that answers the question "should the synthetic alert be firing right now?", and you hand it to your metrics client as a gauge callback.

Prometheus

Alongside your existing Prometheus setup:

from prometheus_client import Gauge
from syntheticalert import SyntheticAlert

gauge = Gauge(
    "triplepat_synthetic_alert",
    "Set to 1 when the synthetic alert should fire and 0 otherwise. Alert on "
    "this metric and route the alert to a Triple Pat check-in timer to "
    "continuously test your alerting pipeline.",
)
gauge.set_function(SyntheticAlert())

This is for single-process servers. In a multi-process server the synthetic alert must come from exactly one process, and the next section shows how. The principle is worth stating on its own: a synthetic alert is a schedule, and a schedule has to have one owner. With one SyntheticAlert per worker there are as many schedules as workers, each scrape lands on a random one, and the alert flaps. Whether that one owner is the master process or a small sidecar exporter of its own is a deployment choice; what matters is that there is one.

Prometheus under gunicorn

Under gunicorn, prometheus_client runs in multiprocess mode: with PROMETHEUS_MULTIPROC_DIR set, every gauge keeps its value in a per-process file, and the scrape is answered by a MultiProcessCollector that reads those files. It never calls a gauge's function, so the set_function wiring above reports the 0 that was written when the gauge was constructed, on every scrape, forever. The alert never fires and the check-in timer raises an alarm for a pipeline that is fine.

The fix is to drive the gauge from the master, which is the one process that exists exactly once. when_ready runs in the master, once, before any worker forks. A daemon thread there sets the gauge from the schedule, which writes the master's file, and every scrape reports the master's value whichever worker answers it. In gunicorn.conf.py:

import threading
import time

from prometheus_client import Gauge
from prometheus_client.multiprocess import mark_process_dead
from syntheticalert import SyntheticAlert


def when_ready(server):
    alert = SyntheticAlert()
    gauge = Gauge(
        "triplepat_synthetic_alert",
        "Set to 1 when the synthetic alert should fire and 0 otherwise.",
        multiprocess_mode="livemostrecent",
    )

    def drive():
        while True:
            gauge.set(alert())
            time.sleep(1)

    threading.Thread(target=drive, daemon=True, name="synthetic-alert").start()


def child_exit(server, worker):
    mark_process_dead(worker.pid)

multiprocess_mode="livemostrecent" is what makes this work. Workers fork after when_ready, so they inherit the gauge, and the first metric a worker touches resets every inherited value and writes a placeholder 0 for the gauge into that worker's file, with no timestamp. Each set in the master stamps its write with the current time, and livemostrecent reports the sample with the newest timestamp, so the master's value always wins and the placeholders never show. The mode also collapses the series to one line without a pid label and drops processes that have exited. The scrape is at most one second stale, which is nothing against a ten-minute firing. The usual multiprocess rule still applies: empty PROMETHEUS_MULTIPROC_DIR when gunicorn starts, or a previous master's file lingers.

The library itself still starts no thread. The thread belongs in the deployment configuration, next to the decision about how many processes there are.

OpenTelemetry

The same callable serves OpenTelemetry through its observe method. The OTel-to-Prometheus exporter turns the dotted metric name into triplepat_synthetic_alert. The one-owner principle applies here too: register the observable gauge in exactly one process.

from syntheticalert import SyntheticAlert

meter.create_observable_gauge(
    "triplepat.synthetic.alert",
    callbacks=[SyntheticAlert().observe],
    description="Set to 1 when the synthetic alert should fire and 0 otherwise.",
)

The schedule

Each firing holds the gauge at 1 for exactly 10 minutes. The silent gap between firings, from the end of one to the start of the next, is memoryless: exponentially distributed with a mean of one hour.

Memoryless gaps make the firings an attempt at a Poisson process, which cannot synchronize with cron jobs or scrape cycles, and which by the PASTA theorem sees your pipeline as it typically is rather than at some special moment.

As a nod to practicality the gap is truncated. It is never less than 10 minutes, so the alert visibly resolves between firings, and never more than two hours, so the check-in timer can be sized. The truncation pulls the realized mean gap down to about 49 minutes and makes the process only roughly Poisson. If you need the PASTA property and can tolerate wider variation in start times, set a lower min and a higher max, then size the timer for the larger max. That recovers most of the Poisson behavior; for the last few percent, use a mean much longer than the firing duration, since the interval between firing starts is the firing plus the gap.

The schedule advances lazily, at scrape time, from time.monotonic(). If nobody scrapes for a while, the next scrape replays every transition it missed, so the process stays honest whatever your scrape interval.

There is no magic here: one line is a serviceable substitute, firing for the first ten minutes of every hour:

import time

gauge.set_function(lambda: 1.0 if time.localtime().tm_min < 10 else 0.0)

But that version fires at the top of every hour, exactly when your cron jobs are doing something interesting. The memoryless schedule cannot synchronize with anything, and that is the point of the library. If you want a deterministic schedule anyway, the line above is all you need.

Options

Every duration is a datetime.timedelta, Python's standard duration type, so the unit is in the code rather than in the documentation. The type hints are the contract; nothing checks types at runtime:

from datetime import timedelta

SyntheticAlert(mean_interval=timedelta(minutes=30), max_interval=timedelta(hours=1))

Keyword	Effect	Default
mean_interval	Mean silent gap between firings	timedelta(hours=1)
min_interval	Lower bound on the silent gap	timedelta(minutes=10)
max_interval	Upper bound on the silent gap	timedelta(hours=2)
firing_duration	How long each firing holds the gauge at 1	timedelta(minutes=10)
clock	Time source in seconds, for tests	time.monotonic

The firing duration must be shorter than the mean interval, and the min and max intervals must bracket the mean. Bad options raise ValueError at construction, never at scrape time. Setting all three intervals equal is allowed: every gap is then exactly that long and the schedule is periodic, which is pointless in production but handy for deterministic debugging. The defaults are exported as DEFAULT_MEAN_INTERVAL, DEFAULT_MIN_INTERVAL, DEFAULT_MAX_INTERVAL, and DEFAULT_FIRING_DURATION.

Alert on the metric

groups:
  - name: synthetic
    rules:
      - alert: SyntheticAlert
        expr: triplepat_synthetic_alert == 1
        labels:
          severity: synthetic
        annotations:
          summary: Synthetic alert exercising the alerting pipeline.

Route the alert to a check-in timer

Create a check-in timer at Triple Pat, then point the alert at it. Prefer email delivery: mail transfer agents queue, retry, and try every backend listed in DNS, so a check-in email is more likely to arrive than a single webhook request to a single destination. Send to the same timer at both the .com and .net addresses. The two domains are served by independent DNS providers, so if one zone cannot be resolved the other address still delivers, and extra simultaneous check-ins are harmless. Merge this into your existing Alertmanager config (the fragment assumes you already have a default receiver and working smtp_* defaults):

route:
  routes:
    - matchers:
        - alertname="SyntheticAlert"
      receiver: triplepat
      group_wait: 0s
receivers:
  - name: triplepat
    email_configs:
      - to: YOUR-TIMER-UUID@checkin.triplepat.com
        send_resolved: false
      - to: YOUR-TIMER-UUID@checkin.triplepat.net
        send_resolved: false

send_resolved: false keeps the resolve notification from counting as an extra check-in, so each firing checks in when it starts and not again when it resolves.

If you cannot send email, replace the triplepat receiver above with this webhook receiver instead. Alertmanager rejects a configuration that defines the same receiver name twice:

receivers:
  - name: triplepat
    webhook_configs:
      - url: https://triplepat.com/api/v1/checkin/YOUR-TIMER-UUID
        send_resolved: false

Sizing the timer

Set the check-in timer's interval to at least max interval + firing duration + your alerting pipeline's latency. With the defaults (silent gaps of at most two hours, plus 10 minutes of firing), a three-hour timer is comfortable.

License

Apache-2.0. See LICENSE.