AI-Driven O11y for Massive AWS Serverless Workflows
summary
APIX wires OpenTelemetry spans from every AWS Step Functions branch into a Fargate collector, feeds 60s windows to a two‑layer graph neural network (about 1.2 million parameters), and patches running map-state chunk sizes with guardrails (25% update cap, concurrency limit, rollback on rising 95th‑percentile latency). Measured overhead is ~2 ms per invocation; results cut throttle errors from 47→5 on a skewed 10 TB workload while lowering p95 latency and overall cost.
AI-written summary · about
chapters (13)
- 0:00Introduction and Welcome
- 0:12Understanding Adaptive Parallelism Insights
- 0:46Challenges in Serverless Fan-Outs
- 1:28Guiding Questions for Adaptive Parallelism
- 2:04Technical Deep Dive: Telemetry and Graph Neural Networks
- 3:09Implementation of A PIX
- 3:22Data Flow and Control Mechanisms
- 4:41Governance and Safety Measures
- 6:52Performance Evaluation and Metrics
- 8:21Lessons Learned and Best Practices
- 9:26Limitations and Future Work
- 10:17Conclusion and Next Steps
- 10:49Q&A and Closing Remarks
related talks
29mGoing Serverless
47mAdvanced Serverless Architectural Patterns on AWS
52mServerless Compute at the Heart of Your EDA
34mObservability in Serverless Application
42mServerless Tales from the Trenches
34mBuilding scalable end-to-end deep learning pipelines in the cloud
51mArchitecting for Hyperscale with AWS: Navigating Real-world Serverless Challenges
41mBoost your serverless application availability with AIOps