Staying Alive: Connection Path Reselection at the Edge
summary
Connection Path Reselection (CPR) augments TCP at Fastly edge POPs to detect stalled connections by monitoring transport-layer forward progress and to reroute them via alternate upstreams using per-connection reroute counter + ECMP-hash selection. Implementation is a Linux kernel patch that switches routing tables (main vs transits) and retries randomized egresses on an RTO, i.e., seconds (RTO) timescale, not minutes like BGP. Deployed in production for >2 years, CPR restores a small but measurable share of flows, handling ~120 impairments/day (median 8 min) and protecting terabits/sec.
AI-written summary · about
chapters (7)
- 0:00Introduction to CPR
- 1:00Limitations of current tools
- 2:27Challenges in edge cloud
- 4:06The CPR design approach
- 5:50Detection and rerouting
- 7:51Path selection strategy
- 9:35Performance and conclusion
related talks
19mTackling DC Congestion and Bursts
24mHarp: Improving VPC Network Availability via Efficient Failure Detection and Rerouting...
05mJump-starting back-office connections
14mA large-scale deployment of DCTCP
24mBlink: Fast Connectivity Recovery Entirely in the Data Plane
22mAdapting TCP for Reconfigurable Datacenter Networks
17mUnlocking ECMP Programmability for Precise Traffic Control
16mR-TCP: A Framework to Optimize TCP Performance Over Rate-Limiting Networks