Availability, Latency and Cost: Withstanding Outages
summary
After a major load‑balancer outage the team builds multi‑region active‑active replication with Cassandra, global caches using Eevee, and DNS geo steering plus CDN edge SSL termination to steer traffic. Failover time drops from 45 minutes to 7 minutes, reducing operational burden and enabling routine evacuations. Cost modeling shows three regions achieve regional tolerance for about 150% cost for three regions when data is everywhere; alternative geo DNS and CDN edge SSL termination steering, central‑savior and ring sharding patterns balance latency, cost and complexity.
AI-written summary · about
chapters (34)
- 0:00Intro
- 0:22Agenda
- 1:06Show of hands
- 2:07Multiregion strategy
- 2:45Single point of failure
- 3:21Regional failover
- 4:57Evacuation
- 5:51Operational Advantage
- 6:28Recovery vs Prevention
- 7:52Peak Demand
- 11:16Example
- 11:47Models
- 14:13Latency
- 14:23Types of latency
- 14:44Different modes of latency
- 15:15Pure latency
- 15:34Pacific Ocean
- 17:14Cost
- 18:43Region
- 18:58Replication Factors
- 19:35Cost Dynamics
- 20:27What if data wasnt everywhere
- 21:15Summary costs
- 21:50Architecture multisite fault isolation
- 24:05No crossregion requests
- 25:39Rebalancing
- 26:27User Profiles
- 32:24Conclusion
- 35:25Questions
- 37:14Experts
- 37:30Organizations appetite for multiregion
- 39:39Reducing complexity
- 40:59Routing traffic to regions
- 43:42Thank you
related talks
33mAvailability, Latency and Cost: Withstanding Regional Outages
18mScaling APIs to Billions: Extreme-Scale Reliability
44mTechniques Netflix Uses to Weather Significant Demand Shifts
18mLeveraging Multiple Regions to Improve Site Reliability:
41mSeven patterns to increase application availability
27mAvailability and Performance Tradeoffs in Global Database Deployments
54mDistributed Caching in Real-Time Payment Systems
27mBuilding Ultra-Resilient Systems in Go