Replication as a First-Class Citizen: The Art of Global Resilience
summary
Argues that multi-region success depends on concrete tradeoffs: measure MTTD/MTTR (49s allows ~53 minutes downtime), avoid synchronous replication beyond ~100 km, and choose between read-local write-global or active-passive routing. Recommends replicating via change data capture or at the message-broker layer (Kafka) rather than only DB snapshots, handle conflicts beyond last-write-wins with app logic or manual queues, and use SDK features (region pinning, read-only annotations) to automate routing and replication while accepting higher cost and operational complexity.
AI-written summary · about
chapters (16)
- 0:00Introduction and speaker welcome
- 0:16Talk overview and goals
- 3:51Single-region resilience principles
- 6:19Reasons to adopt multi-region deployment
- 8:11Key considerations before going multi-region
- 12:07CAP theorem and replication choices
- 16:26Data access patterns and architecture strategies
- 22:27Understanding service dependencies
- 25:58Configuration and operational dependencies
- 26:41Operational challenges and failover types
- 34:10Global routing and internal traffic redirection
- 37:42Data replication strategies and event-driven systems
- 41:17Conflict resolution in multi-region systems
- 42:43Wacka platform and multi-region features
- 45:33Reasons not to go multi-region
- 46:14Summary and key takeaways
related talks
44mDistributed systems and scalability rules
18mLeveraging Multiple Regions to Improve Site Reliability:
01h 00mDistributed transactions implementation trade-offs
44mAvailability, Latency and Cost: Withstanding Outages
35mDesign patterns for application resiliency
56mFrom Unreliable to Resilient: Mastering Distributed Systems
39mTrust Deterministic Execution to Scale & Simplify Your Systems
20mLightning Talk: Multi-Datacenter Strategies for Data Services