Design Multi-Region Deployment - System Design | WinJob

Problem Statement: Multi-Region Deployment

Design a multi-region deployment for a global SaaS API (Netflix/Google/Uber style). Users expect low latency anywhere, 99.99% availability during regional outages, and controlled consistency for writes. The platform runs identical regional cells (Kubernetes or VM fleets) behind GeoDNS / global load balancing, with async or multi-master replication for shared data and regional evacuation playbooks.

Interview Focus

Tie every decision to RPO/RTO, regional blast radius, and data consistency class
Quantify cross-region bandwidth, DNS failover time, and replication lag before picking topology
Explain why active-active needs conflict handling while active-passive simplifies writes
Describe failure modes: split brain, stale reads, regional evacuation, and config drift

javaOne Dark Pro

1 public final class RegionHealth {
2   public boolean routeTraffic(String regionId, boolean healthy, int errorRateBps) {
3     return healthy && errorRateBps < 50;
4   }
5 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class RoutingPolicy:
5     home_region: str
6     weight: int
7  
8 def pick_region(policies: list[RoutingPolicy], client_region: str) -> str:
9     local = [p for p in policies if p.home_region == client_region]
10     return local[0].home_region if local else policies[0].home_region

typescriptOne Dark Pro

1 interface RegionCell {
2   id: string;
3   healthy: boolean;
4   weight: number;
5 }
6  
7 export function weightedPick(cells: RegionCell[]): string | null {
8   const live = cells.filter((c) => c.healthy && c.weight > 0);
9   if (live.length === 0) return null;
10   return live.sort((a, b) => b.weight - a.weight)[0].id;
11 }

How to open this one

The framing that lands for multi-region is identical regional cells behind global load balancing, with the hard question being data consistency, not compute. Lead with the consistency choice — async replication for low-latency reads versus multi-master for write availability, and what you give up either way — and the failure story that proves it: a region fails, GeoDNS evacuates traffic to a healthy cell, and the replication lag determines how much (if any) data is lost. That shows you understand the CAP trade-off is the real design, not the box diagram.

Key Highlights

•Regional cells bound blast radius
•Explicit RPO/RTO per datastore
•Layered failover beats DNS-only
•Replication class drives write topology

1	public final class RegionHealth {
2	public boolean routeTraffic(String regionId, boolean healthy, int errorRateBps) {
3	return healthy && errorRateBps < 50;
4	}
5	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class RoutingPolicy:
5	home_region: str
6	weight: int
7
8	def pick_region(policies: list[RoutingPolicy], client_region: str) -> str:
9	local = [p for p in policies if p.home_region == client_region]
10	return local[0].home_region if local else policies[0].home_region

1	interface RegionCell {
2	id: string;
3	healthy: boolean;
4	weight: number;
5	}
6
7	export function weightedPick(cells: RegionCell[]): string \| null {
8	const live = cells.filter((c) => c.healthy && c.weight > 0);
9	if (live.length === 0) return null;
10	return live.sort((a, b) => b.weight - a.weight)[0].id;
11	}

Problem Statement: Multi-Region Deployment

Interview Focus

How to open this one

Key Highlights

Section Rescue Kit