The Full Picture: Visualizing Service "Fullness" To Rethink Saturation Prevention
summary
Proposes exposing each instance’s live concurrency limit and its current in‑flight requests by extending Envoy and leveraging eBPF and Cilium to read per‑request data, using the existing adaptive concurrency filter to derive a dynamic capacity gauge. Prototype detects saturation per instance (P99 latency drift) within ~100 ms, exports both capacity and active‑request counts to Prometheus, and feeds admission control or autoscalers (Keda) to cut MTTR and shrink safety margins.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:04Saturation prevention rethink
- 3:08Defining saturation
- 5:38Indirect metrics fail
- 8:06Human vs AI SRE limits
- 9:32Service mesh gaps
- 13:04Adaptive concurrency filter
- 16:15Dynamic capacity per instance
- 18:43In-flight requests tracking
- 20:08Capacity vs demand metrics
- 21:12Real-time saturation visualization
- 22:54Admission control applications
- 24:16Fine-grained metrics implementation
- 25:00AI for metric quality
- 26:27Q&A
related talks
37mMonitoring Service Architecture and Health with BPF
35mPractical and Useful Latency Analysis using Istio and OpenCensus - Varun Talwar
20mService Mesh: A Hole in the Pocket?
20mNetwork Monitor: A Tale of ACKnowledging an Observability Gap
37mPractical Capacity Planning using Prometheus
27mAchieving Fault Tolerance in Istio with Observability-Driven Load Management
27mInsightful Traffic Monitoring: Harnessing Cilium for Comprehensive Network Observability
18mMeasuring Congestion in High Performance Datacenter Interconnects