Analyze & Correlate

Patterns, anomalies, and correlations across your telemetry.

AI Insights runs over the whole lake rather than a sampled recent window. It collapses repetitive log lines into the handful of patterns actually behind them, flags where behaviour departs from its own history, and connects movements across logs, metrics, and traces that would otherwise be three separate investigations.

Pattern compression

Millions of log lines, a handful of patterns

Production emits the same lines endlessly. AI Insights groups them by structure — variable parts abstracted, counts preserved — so a night of noise becomes a short list a person or an agent can actually read. This is CtrlB Decompose, running continuously over the lake.

Raw log lines0
GET /api/checkout 200 142ms user=8831
conn pool acquire wait=0ms host=db-3
GET /api/checkout 200 138ms user=2290
cache miss key=cart:8831 region=us-east
GET /api/checkout 200 151ms user=4517
retry attempt=2 svc=payments backoff=200ms
GET /api/checkout 200 147ms user=9902
conn pool acquire wait=0ms host=db-1
GET /api/checkout 500 err=timeout user=1043
cache miss key=cart:2290 region=eu-west
GET /api/checkout 200 144ms user=7781
retry attempt=1 svc=payments backoff=100ms
GET /api/checkout 200 142ms user=8831
conn pool acquire wait=0ms host=db-3
GET /api/checkout 200 138ms user=2290
cache miss key=cart:8831 region=us-east
GET /api/checkout 200 151ms user=4517
retry attempt=2 svc=payments backoff=200ms
GET /api/checkout 200 147ms user=9902
conn pool acquire wait=0ms host=db-1
GET /api/checkout 500 err=timeout user=1043
cache miss key=cart:2290 region=eu-west
GET /api/checkout 200 144ms user=7781
retry attempt=1 svc=payments backoff=100ms
Structural patterns5 total
GET /api/checkout %d %dms user=%s×1.9M
cache miss key=%s region=%s×412K
conn pool acquire wait=%dms host=%s×88K
retry attempt=%d svc=payments backoff=%dms×6.2K
GET /api/checkout 500 err=timeout user=%s×914
0 lines5 patterns to read~100,000× fewer things to look at

Anomaly detection

Measured against everything, not a recent window

Most tools compare today to the last two weeks, because that is all they keep hot. AI Insights holds full history as open Parquet, so “unusual” is judged against a service’s real baseline — seasonality, deploy cycles, quiet weekends and all — instead of a fortnight that happened to look calm.

checkout · error-rate baseline
365-day baselinedeparture
LAST 15 DAYS

Recent-window view

Inside the last 15 days the spike looks like ordinary variance.

Full-history view

Against a year of baseline it is three standard deviations out.

Cross-signal correlation

One movement, seen across logs, metrics, and traces

Because all three signals land in the same lake, a latency step in metrics, a burst of a log pattern, and a slow span in traces are one finding at one timestamp — not three tools open in three tabs, reconciled by hand.

Correlated signals09:12 UTC · payments-api v84
Logspattern rate
Metricsp99 latency
Tracesslow spans

One finding, three signals. The log-pattern burst, the latency step, and the slow spans share a timestamp and a deploy — so AI Insights reports them as a single correlated event, not three unrelated alerts.

Transparent findings

Every finding is a query, not a black box

A finding opens into the exact query that produced it, the sample it ran on, and the window it covers. Adjust the query, widen the window, or drop it into the Data Lake — none of the reasoning hides behind a score you cannot reproduce.

HighResource Exhaustion

Analytics processor queue backing up

0Impact
Pattern logs4,120
Window6h
Signalslogs · metrics
the query behind the finding
SELECT pattern_id, count(*) AS n
FROM logs
WHERE service = 'analytics-processor'
  AND ts > now() - 6h
GROUP BY pattern_id
HAVING n > baseline(pattern_id) * 3
reproducibleeditable windowOpen in Data Lake

What you get

  • Repetitive logs compress into patterns a human or an agent can reason over.
  • Anomalies are measured against full history, not the last fifteen days.
  • Correlations span signals, because all three live in one data layer.
  • Findings are queries you can open, inspect, and adjust — not opaque scores.

Other core components

All four run over the same lake of open Parquet in object storage you own.

Back to platform overview

See what your own lake is already telling you.

Point AI Insights at live data and read the patterns, anomalies, and correlations back as queries you can open — not scores you have to trust.