Sponge: Fast Reactive Scaling for Stream Processing with Serverless Frameworks
summary
Sponge is a stream-processing runtime that pre-inserts router, transient and merge operators at compile time to avoid runtime state migration, redirects bursty partitions to AWS Lambda instances launched in a few hundred milliseconds, and merges lightweight partial state instead of moving full state. It monitors CPU with mpstat, targets 60–80% utilization, and aims to recover within one second. Implemented with Apache Beam/Nemo and evaluated on EC2 + Lambda using NEXMark, it cuts 99th‑percentile latency by 88% and cost to about 17% of naive overprovisioning.
AI-written summary · about
chapters (11)
- 0:00Sponge introduction
- 1:48Bursty input load challenges
- 3:54Serverless frameworks for scaling
- 5:19Serverless instance challenges
- 6:24Sponge design
- 9:13Resource calculation for scaling
- 12:07Adaptive scaling runtime
- 13:11Implementation setup
- 14:37Experimental results
- 17:03Cost and efficiency analysis
- 18:07Conclusion
related talks
40mThe State of Serverless Computing
38mBackground Processing, Serverless Style
32mEmbracing Serverless Ops
32mUsing Serverless Functions for Real-time Observability
34mA journey of scaling high-volume low-latency stateful processing
26mHow to Pick Your Next Streaming Architecture
35mServerless WebSockets at Scale
29mGoing Serverless