New in 0.8 · What changed · Pro

Anomaly detection for Kubernetes and Docker: See what changed first in every incident.

Each metric in WatchLog’s catalog learns its usual week. When something breaks, WatchLog lists everything that left its range, in the order it happened, with the likely cause underneath. No thresholds to write, no AI guessing.

Part of Cloud Pro · $1 per node a month, free during the beta · Kubernetes and Docker

What changed
Platform
Outside the usual week · each metric in WatchLog’s catalog

acme-prod · around the 14:02 alert

  1. 13:58reports-worker · connections to postgresusual 0–1/snow 12/s
  2. 13:58postgres · connectionsusual 18–30now 97
  3. 14:00checkout · p95 response timeusual 120–180 msnow 2.4 s
  4. 14:01checkout · errorsusual 0–1 %now 38 %
Likely cause · where it starts

Connections are at 97% of max_connections (97 of 100). Most are held by reports-worker.

docker-prod · around the 14:02 alert

  1. 13:58storefront-reports-1 · connections to dbCompose storefront · reportsusual 0–1/snow 12/s
  2. 13:58postgres · connectionsusual 18–30now 97
  3. 14:00storefront-checkout-1 · p95 response timeCompose storefront · checkoutusual 120–180 msnow 2.4 s
  4. 14:01storefront-checkout-1 · errorsCompose storefront · checkoutusual 0–1 %now 38 %
Likely cause · where it starts

Connections are at 97% of max_connections (97 of 100). Most are held by storefront-reports-1.

An example, in each platform’s words. The live demo shows its own: search’s response time drifts up every 2 hours, at 50 minutes past every odd hour UTC.

How it works

Anomaly detection you can read in one glance.

  1. 01

    It learns your week

    Each metric in WatchLog’s catalog gets its own usual range for each of the 168 hours of the week, from the last 14 days.

  2. 02

    It notices what leaves it

    Something counts only after 10 minutes far outside its range. One spike isn’t an incident.

  3. 03

    It shows what moved first

    One list, ordered by when each thing changed, with probable cause naming where it starts.

Thresholds vs. your usual week

A fixed number can’t know it’s Tuesday lunchtime.

Threshold alertp95 > 500 ms for 5 min
  • Fires every busy lunchtime, so people mute it
  • Misses 120 → 450 ms at 3 a.m., the slow drift before the outage
  • Someone has to choose the number, for every service
WatchLog anomaly alertcheckout’s p95 response time is unusual for a Tuesday at 14:00
  • Each metric learns its own 168 hours of the week
  • Alerts only after 10 minutes far outside its range, one per workload or container and metric
  • Quiet for 30 minutes after a rollout, and while another alert fires on it
  • “No, this is normal for us” widens that hour’s range
The alert and its message →
  • No black boxA median and a spread per hour of the week. Every row shows its usual range, so you can see why it was listed.
  • Nothing sent to an AIWorked out inside WatchLog. Your metrics never leave it to be analysed, and the same data gives the same answer.
  • Probable cause underneathFixed rules with their evidence name where it starts. The list itself is ordered by time, not guessed.
  • Kubernetes and DockerWorkloads, nodes and the service map, or containers, Compose services and hosts. Each in its own words.

Questions

What is Kubernetes anomaly detection?

Finding the metrics that behave differently from their own usual pattern, instead of comparing them with a fixed threshold. WatchLog learns the usual week of each metric in its catalog for every workload, node and service map connection, and lists the ones that stay far outside it for 10 minutes.

Does it do Docker container anomaly detection too?

Yes. On a Docker fleet each container, its Compose service and each host learn their usual week the same way: CPU, memory, network, and the requests, errors and response time a container serves. The page names them in Docker’s words.

Is it Kubernetes root cause analysis?

It shows what changed first, in the order it happened, and puts probable cause underneath: fixed rules that name where an incident starts, with their evidence. It doesn’t claim one anomaly caused another; the order and the evidence are there to check.

How is it different from threshold alerts?

A threshold is one number for every hour. What changed compares each metric with what it usually does at that hour of that weekday, so a busy lunchtime stays quiet and a slow drift at night is caught. Anomaly alerts need no threshold, wait 30 minutes after a rollout, and stay quiet while another alert already fires on the same workload or container.

Does it use AI, or send my data anywhere?

No. It is statistics, a median and a spread for each hour of the week, worked out inside WatchLog. Nothing is sent to an AI model, and the same data always gives the same answer.

How soon does it work?

It starts learning when your organization gets Pro and lists changes after 7 days of history; each hour of the week has two weeks behind it after 14. The page says how long is left. The live demo has learned already.

Which plan includes it?

WatchLog Cloud Pro: $1 per node a month with a $10 minimum, free during the beta. Alerts, the service map, probable cause and Egress watch stay free on every plan, including the self-hosted Community edition.

See what changed in your next incident.

Start free with alerts, the service map and probable cause. Add What changed with Pro: $1 per node a month.