World's Fair 2026
Using RL Agent to Detect and Remediate ETL Pipeline Failures - Anna Marie Benzon
Overview
This talk presents an RL-guided system designed to detect and remediate failures in ETL pipelines. The core idea is to enable an AI agent to act usefully, explainably, and within operational trust boundaries, significantly reducing the time and effort required for manual incident resolution. The system aims to automate responses for routine failures while escalating complex or high-risk situations.
Who should watch
- AI Engineers
- Data Engineers
- Site Reliability Engineers (SREs)
- Product Managers overseeing data infrastructure
- Builders working on automated incident response
- Those facing challenges with slow, manual ETL failure remediation
Key takeaways
- A system architecture using AWS Lambda and EventBridge triggers an RL agent upon ETL job failure events.
- The agent gathers evidence from read-only sources like CloudWatch logs and Glue Data Catalog to classify failures and assess risk.
- A deterministic anomaly detection layer identifies observable facts, while a Q-learning policy handles contextual action selection.
- A safety layer enforces operational constraints, overriding the learned policy for critical or uncertain cases, ensuring actions are both safe and appropriate.
- The system separates concerns into deterministic rules for facts, learning for bounded choices, and guards for authority, making it inspectable and explainable.
- Performance benchmarks show a simulated success rate of 74.63% and an 88.63% non-escalation rate, representing a significant reduction in Mean Time To Resolution (MTTR) compared to manual processes.
- Reliability stems from standardized logic, external safety constraints, and the agent's ability to recognize its own boundaries, rather than solely from RL sophistication.
- Key engineering principles include using deterministic logic for measurable facts, learning for contextual action selection, placing safety constraints outside the learned policy, and treating escalation and validation as first-class outcomes.
Notable quotes
The central question is not simply whether an agent can act, but whether it can act usefully, explainably, and within boundaries that an operation would actually trust.
*A practical self-healing system does not need the largest possible model; it needs a clear state, bounded actions, reproducible evaluation, observable decisions, and the discipline to stop when uncertainty exists.*
The goal is not to eliminate human judgment, but to stop spending that judgment on the same recognizable failures at two in the morning.
Unofficial community note. Prefer the recording for nuance.