Observability Is Eating Your Cores: Fine-Grained Analysis of Microservice Metrics with...
summary
μView implements an on-accelerator observability data plane that runs streaming IPU-hosted sketches to analyze microservice metrics locally and ship only informative data. It claims sub-second visibility for short-lived anomalies (<30s) and biased trace sampling that increases symptomatic-trace coverage versus head-based samplers, matching or exceeding dimensionality-reduction baselines. Evaluation shows a single BlueField2 IPU core sustains 256 pods, processing fine-grained metrics with low host CPU impact and reducing the remote ingestion and long-term storage burden from massive pipelines (e.g., Uber’s 500M/s → $21M/mo example).
AI-written summary · about
chapters (14)
- 0:00Microservice observability challenges
- 1:02Microservices failure complexity
- 2:26Costs of metrics collection
- 4:11Slow sampling limitations
- 5:15Requirements for monitoring systems
- 6:41Centralized monitoring intelligence
- 7:23Local failure detection analysis
- 9:08Fine-grained metrics costs
- 10:13Low-overhead anomaly detection
- 13:20Offloading sketches to IPUs
- 14:02Evaluating MicroView sampler
- 15:51MicroView scalability on IPUs
- 16:33MicroView meets monitoring needs
- 17:16Q&A and open-source
related talks
52mIs your JVM app flying blind? Unmask hidden issues with Observability superpowers!
22mUnified Observability
22mAI Observability At Meta Scale
50mIntuit Journey To Unified Observability at Scale: Challenges, Be...
26mProfiling Mastery: Decode Your Data and Elevate Observability
38mObservability 101 with Spring and Micrometer
48mAPI Observability by Design: Building Systems That Tell Their Own Stories
31meBPF for Observability: The Good, the Bad, and the Ugly