Customer Journey Observability: Start with Impact, Drill to Root Cause

# What this story describes

Capital One contrasts the traditional service-first observability model with a journey-first approach implemented in Observe by Snowflake. Instead of starting at infrastructure or service alerts and estimating customer impact afterward, teams begin with a defined customer interaction, measure business impact, then follow relationships in telemetry down to services, operations, traces, and logs to confirm root cause.

# Why this matters

# Snowflake

• Telemetry ingested into Observe's data lake includes real user monitoring (RUM) logs, clickstream data, and API access logs.
• After ingestion, interactions and services are modeled and enriched without requiring app-level instrumentation.
• Customer journey graphs are constructed where nodes represent interactions (for example: Home, Product Details, Add Item, Complete Order) and edges represent the user flow.
• RED metrics (rate, errors, duration) and unique session counts are aggregated and overlaid on the journey graph so the operational surface shows business-level impact first.

# The practical workflow (e-commerce example)

1) Start at the customer interaction

• Operators identify a degraded journey step such as Complete Order. The graph immediately shows how many sessions, error rate, and latency are affected so teams know scope and severity before troubleshooting technical layers.

2) Drill down to the failing operation

• From the failing interaction you pivot to service- and operation-level breakdowns. In the example, the EmptyCart operation in the cart service was causing latency and errors. This narrows ownership and the domain for remediation.

• Once a suspect operation is identified, engineers pivot into traces to inspect where time is spent and where errors originate in a specific request. They then drill into downstream span logs to confirm the explicit error message and complete the failure narrative: users can't complete checkout because a downstream operation cannot read cached cart data.

# Benefits observed

• Immediate, business-focused measurement of impact so executives and SREs share the same single source of truth.

• Implementation without adding application instrumentation because interactions are modeled after telemetry ingestion.

# Operational realities and constraints

• The structure and relationships are encoded in the data, but investigations remain human-driven. Engineers still decide where to pivot, which signals matter, and how to summarize findings under pressure.
• Synthesis of findings for the incident response room still requires time and coordination even when impact and root cause are surfaced more quickly.

# Bottom line