Products

OML Anomaly

OML Anomaly runs automatically as your application sends telemetry. It learns normal behavior, detects departures from that baseline, and feeds the evidence directly into OML’s incident management system. The same pipeline opens, updates, and resolves incidents as conditions change.

Signal extraction

OML extracts three kinds of evidence from incoming telemetry. Backend logs describe service activity and recurring errors. Browser request summaries describe request counts, durations, HTTP failures, and transport failures. Grouped JavaScript errors capture failures inside the browser.

Each signal has an identity that includes its project, source, environment, and monitored entity. This keeps the history of a browser request separate from the history of a backend operation. A browser failure can produce evidence even when the request never reaches the server.

Observation windows

The engine aggregates observations over time. For browser latency, it adds the durations across compatible request summaries and divides by the total request count. A summary of 900 requests therefore carries nine times the weight of a summary of 100. For event frequency, the denominator is the time actually observed.

Before a window can confirm a finding, OML checks its timing, continuity, and stored evidence. A gap in collection leaves the result unknown. Treating that gap as zero failures would let a telemetry interruption look like a recovery.

How baselines learn

OML builds a separate historical profile for each signal and metric. Eligible windows contribute to hourly summaries. Readiness depends on sample volume, coverage across distinct periods and days, and freshness. Replaying one busy hour cannot give a new signal several days of history.

The profile tracks a typical value and how far measurements usually vary. Its center moves in bounded steps, and its variability estimate uses median absolute deviation to limit the influence of isolated spikes. Suspicious observations are withheld from learning so a sustained failure cannot quickly become the expected behavior.

Anomaly evaluation

The detector compares a current observation with an earlier, sufficiently mature baseline for the same metric. A first pass flags values outside the expected range. Separate checks then test changes in frequency, mean latency, or failure proportion using the available sample size.

For example, 10 failures in 1,000 checkout requests and 20 in 2,000 both give a 1% failure rate. The raw count doubled, but the failure proportion held steady. If failures reach 100 in 2,000 requests, the proportion reaches 5%. Whether that becomes a finding depends on historical variation and the quality of the observation.

Incident creation and updates

The detector passes its findings directly to OML’s incident management system. An incident is keyed by the project, signal source, signal identity, metric, direction of change, and cause identifier when available. Repeated findings for that condition update the same incident with the observed value, expected range, sample count, and references to the supporting windows.

A new episode starts as suspected. Two independent, adjacent windows with confirmed anomaly evidence make it active. Reprocessing the same window cannot advance that count. The Incidents dashboard reads the resulting state automatically.

Recovery and recurrence

Recovery requires fresh, qualified observations of the same condition. Three consecutive independent healthy windows move the incident to recovering. Five resolve it. An unknown observation breaks the streak, and a confirmed anomaly during recovery returns the incident to active.

If the same condition recurs within 30 minutes of resolution, the system reopens the incident. A later recurrence starts a new episode. This preserves the history of a brief relapse while giving a separate occurrence its own timeline.

Design philosophy

An incident closes when the signal shows sustained recovery.

OML learns each signal’s baseline and keeps repeated findings in one incident. If telemetry goes quiet, that incident stays open until fresh observations confirm recovery.

Found something unexpected? Let us know.