Attribute Correlation Analysis

Finding the right traces is only half the problem — understanding why those traces are failing is the other half. Without automated analysis, diagnosing root cause requires manually comparing span attributes across hundreds of traces, building complex queries, and forming hypotheses through trial and error.

Attribute Correlation Analysis (ACA) automates this process. It statistically compares the anomalous traces identified by Intelligent Exemplar against a healthy baseline from the same time window, surfaces the attribute-value pairs that most strongly differentiate the two groups, and presents findings as a plain-language AI summary with a confidence score. The result is a directed starting point for remediation, not a raw data dump to interpret.

How ACA works

ACA runs against the outlier trace set produced by Intelligent Exemplar. It does not operate on arbitrary trace queries — it requires an active Intelligent Exemplar investigation as its input.

Target group vs baseline group

ACA constructs two groups from your trace data for the selected time window:

• Target (outlier) traces: the anomalous traces identified by Intelligent Exemplar — the ones exhibiting high latency or elevated error counts relative to baseline behavior.
• Baseline traces: healthy, non-anomalous traces from the same entity and time window that represent normal behavior.

Statistical comparison

ACA compares the attribute value distributions across both groups. Attributes that appear disproportionately in the target group — for example, a specific deployment version, host, geographic region, or customer segment — are surfaced as correlated with the anomaly. Each attribute is assigned a probability score that indicates how strongly it differentiates the outlier group from the baseline.

AI synthesis

An AI model synthesizes the statistical signals into a plain-language explanation, naming the primary culprit and explaining why it is implicated. For example: "The latency spike in Payment Service is primarily driven by requests handled by entity-status-daemon in the production environment running release-953, which accounts for 78% of outlier traces."

Requirements

• Intelligent Exemplar must be initiated first. ACA operates on the outlier trace set it produces — you cannot run ACA independently.
• Distributed tracing must be enabled for your services. See Introduction to distributed tracing.
• New Relic AI must be enabled on your account to generate AI summaries.

Run Attribute Correlation Analysis

ACA is initiated from the Intelligent Exemplar landing page after an investigation is underway.

1. Complete an Intelligent Exemplar investigation to reach the landing page with an outlier trace list.
2. On the landing page, select Analyze patterns & attributes.
3. ACA runs automatically against the outlier trace set. Results appear in the same view.

Review ACA results

ACA results consist of three components: the AI-generated summary, the detected patterns cross-reference, and the attribute distribution charts.

AI-generated summary

The summary panel identifies the primary driver of the anomaly in plain English. It names the specific attribute-value pairs — such as a deployment version, environment tag, host, or customer account ID — that appear disproportionately in the outlier traces.

Each named attribute in the summary is a clickable filter. Selecting it narrows the outlier trace list to traces that match that attribute value, allowing you to drill further or verify the finding.

Detected patterns cross-reference

The patterns detected by Intelligent Exemplar remain visible during ACA results. Use them to cross-reference and corroborate the AI summary:

• If ACA cites a downstream service as the primary culprit and the Downstream latency or Downstream errors pattern is active, the two signals are in agreement — high confidence.
• If ACA cites a focal entity attribute but the Downstream latency pattern is active, there may be multiple contributing factors. Explore the attribute charts for additional signals.

Attribute distribution charts

Below the AI summary, ACA renders bar charts for each attribute it evaluated. Each chart shows the frequency distribution of attribute values in the outlier (target) group compared to the baseline group.

Use the attribute charts to:

• Validate the AI summary: Confirm that the named attribute appears disproportionately in the outlier bar relative to the baseline bar.
• Explore secondary causes: Look for other attributes with large distribution gaps that the AI summary may not have surfaced as the primary driver. A spike may have multiple contributing factors.
• Identify blast radius: Attribute charts for dimensions like customer.id, aws.region, or deployment.environment reveal which user segments, regions, or environments are affected and can be used to scope the incident's impact.

Interpret confidence scores

Each attribute surfaced by ACA is assigned a probability score indicating how strongly it differentiates the outlier group from the baseline.

• High scores (roughly 70% and above) indicate the attribute value appears in a significantly larger proportion of outlier traces than baseline traces. These are strong candidates for root cause.
• Moderate scores (40–70%) indicate a meaningful but less definitive correlation. Worth investigating, especially if corroborated by a detected pattern.
• Low scores (below 40%) indicate the attribute appears only marginally more often in outlier traces. The anomaly may have multiple contributing factors, or the trace volume may be insufficient for reliable statistical differentiation.

What's next

• Navigate to the failing span: Use the attribute filter in the AI summary to narrow the outlier trace list, then select a trace to land directly on the anomalous span.
• Correlate with a deployment: If ACA cites a deployment version, cross-reference it with Change Tracking to see what was deployed at that time.
• Manage trace volume: See Control distributed tracing data ingestion to manage trace ingestion costs after an incident.
• Tune sampling: See Configure sampling to increase trace coverage for the affected service or transaction.