Larm·April 18, 2026·Updated November 8, 2026monitoringreliabilityhow-it-works

How multi-probe voting eliminates false alerts

A check from one location can fail for reasons that have nothing to do with your service. How Larm uses majority voting across probes and confirmation windows to decide when a monitor is actually down.

A check that runs from one location has to cross the public internet to reach your server and back. A lot can go wrong on the way that has nothing to do with your service: a routing problem at the probe's provider, packet loss at a peering point, a slow CDN edge in that one region. Any of those turns into a timeout, and to a monitoring tool that only checks from one place, a timeout looks the same as an outage.

I've been on call for a lot of years, and I've seen plenty of false alerts like that. After enough of them, people start ignoring alerts, including the real ones. Larm is built to avoid that, and this post explains how it decides a monitor is down.

Retries from the same place don't help much

A common way to reduce false alerts is to retry the failed check before alerting. That helps with a one-off blip. But if the cause is a network problem near the probe, a retry from the same location runs into the same problem and fails again. Now the tool has two failures and treats it as confirmed.

Majority voting

Larm checks every monitor from several probe locations and lets them vote.

Every 5 seconds, the evaluator takes the most recent result from each probe within a lookback window. Each probe gets one vote: pass, or not pass. A failed check, an error and a timeout all count as not pass. Then:

  • If more probes fail than pass, the vote is a majority fail.
  • If more pass than fail, it's a majority pass.
  • If it's even, it's a tie, and the monitor doesn't change state.
Nuremberg   → timeout    ✗
Helsinki    → 200 OK     ✓
Virginia    → 200 OK     ✓
Singapore   → 200 OK     ✓

3 pass, 1 fail → majority pass → no alert

One probe timed out and three got a 200. That points at the network near one probe, not at your service.

The lookback window depends on the monitor's check interval: max(interval * 2 / 60 + 1, 3) minutes. A monitor that checks every 60 seconds gets a 3-minute window, and one that checks every 5 minutes gets 11 minutes. The window always covers at least two check cycles, which leaves room for probes that report a little late.

Confirmation windows

Voting handles a single probe having a bad moment. Sometimes a network problem affects several probes at once, for example a routing change that makes a server unreachable from part of the internet for half a minute. For a short time, most probes can report failure even though the service is fine.

That's what confirmation windows are for. When the vote turns into a majority fail, Larm doesn't mark the monitor down right away. It starts a timer. If the vote stays majority fail for the whole window, the monitor goes down. If the vote changes at any point, the timer resets.

The default window for going down is 1 minute, and for recovering it's 3 minutes. Recovery is slower on purpose: a monitor that comes back up too quickly and then goes down again sends a stream of alerts. You can set both per monitor, from 0 to 30 minutes.

So the two work together. Voting filters out problems at a single probe. Confirmation windows filter out short problems that hit several probes. When a monitor goes down, most probes have agreed on it for the whole window.

When a monitor goes down

When the window runs out and the monitor goes from up to down, Larm:

  1. Records the state change.
  2. Sends an alert to every alert channel on the monitor, like Slack, email or PagerDuty. The alert includes the last error a probe reported.
  3. Creates a disruption on any status page component linked to the monitor.
  4. Updates the dashboard for anyone looking at it.

The dashboard shows each probe's latest result and the request timing from each location: DNS, TCP connect, TLS, time to first byte. That's where you see the difference between "TCP connections time out from every location", which is a server problem, and "timeouts from Singapore and São Paulo, 200 from Europe and the US", which is probably routing in those regions.

The tradeoff

Larm doesn't mark a monitor down on the first failed check. With the default 1-minute confirmation window, a real outage is confirmed and alerted in about a minute. In return, an alert from Larm means most probes agreed the service was down for that whole minute.

That's the point of all of this. People respond to alerts they trust, and they learn to ignore the ones they don't.

Larm's free plan includes 15 monitors with multi-probe voting from all probe locations.

Start monitoring in minutes.

Free plan. 15 monitors. Multi-probe voting. No credit card.

Sign Up Free