Geographically Distributed Systems#

When your users span continents, a single-region deployment creates a terrible experience. A user in Tokyo hitting a server in Virginia faces 150-200ms of network latency on every request — before the server even starts processing.

Geographically distributed systems solve this by placing compute and data closer to users. But distribution introduces hard problems: consistency, conflict resolution, and compliance.

Why Geo-Distribution?#

Three forces push systems toward multi-region:

1. Latency — Physics limits speed of light. Cross-ocean round trips add 100-300ms
2. Availability — A single region is a single point of failure. Cloud regions go down
3. Compliance — GDPR, data sovereignty laws, and industry regulations require data to stay in specific geographies

Single region:
  Tokyo user → Virginia server → 180ms RTT
  Frankfurt user → Virginia server → 90ms RTT

Multi-region:
  Tokyo user → Tokyo server → 5ms RTT
  Frankfurt user → Frankfurt server → 3ms RTT

Data Replication Across Regions#

The core challenge is keeping data consistent across regions. There are two fundamental approaches.

Synchronous Replication#

Every write is confirmed by all replicas before acknowledging the client.

Client → Write to US-East
         → Replicate to EU-West (wait)
         → Replicate to AP-Southeast (wait)
         → All confirmed → ACK to client

Pros: Strong consistency — every read sees the latest write. Cons: Write latency equals the slowest replica. One slow region penalizes everyone.

Asynchronous Replication#

The primary region acknowledges immediately; replicas catch up in the background.

Client → Write to US-East → ACK (immediate)
         → Replicate to EU-West (background, ~50-200ms)
         → Replicate to AP-Southeast (background, ~100-300ms)

Pros: Fast writes. No region blocks another. Cons: Reads from replicas may return stale data (eventual consistency).

Consistency Models#

Most applications use session consistency — users always see their own writes, even if other regions lag slightly.

Geo-Routing: Getting Users to the Right Region#

Latency-Based DNS#

DNS resolves to the region with lowest measured latency:

Route 53 latency-based routing:
  User in Japan → ap-northeast-1
  User in Germany → eu-west-1
  User in Brazil → sa-east-1

GeoDNS#

Routes based on the user's geographic location (IP geolocation):

GeoDNS rules:
  IP in EU → eu-west-1.app.com
  IP in APAC → ap-southeast-1.app.com
  IP in Americas → us-east-1.app.com

Anycast#

A single IP address is announced from multiple locations. BGP routing sends packets to the nearest one. Cloudflare and most CDNs use this approach.

Application-Level Routing#

For more control, route at the application layer:

// Middleware determines user's home region
function routeRequest(req) {
  const userRegion = getUserHomeRegion(req.userId);
  if (userRegion !== LOCAL_REGION) {
    return proxy(req, regionEndpoints[userRegion]);
  }
  return handleLocally(req);
}

Active-Active vs Active-Passive#

Active-Passive (Primary-Secondary)#

One region handles all writes. Other regions serve reads and stand by for failover.

US-East (Primary):  reads + writes
EU-West (Passive):  reads only, receives replicated data
AP-South (Passive): reads only, receives replicated data

Failover: promote EU-West to primary (~30-60s)

Best for: Applications where write conflicts are unacceptable (banking, inventory).

Active-Active (Multi-Primary)#

Every region accepts both reads and writes. Conflicts are resolved after the fact.

US-East (Active):  reads + writes → replicates to other regions
EU-West (Active):  reads + writes → replicates to other regions
AP-South (Active): reads + writes → replicates to other regions

Best for: Low-latency writes globally, high availability requirements.

The cost: You must handle conflicts when two regions modify the same data simultaneously.

Conflict Resolution#

When two regions write to the same record, you need a resolution strategy.

Last-Writer-Wins (LWW)#

The write with the latest timestamp wins. Simple but can lose data.

Region A: SET user.name = "Alice"  (t=100)
Region B: SET user.name = "Bob"    (t=101)
Result: "Bob" wins. Alice's change is silently dropped.

Conflict-Free Replicated Data Types (CRDTs)#

Data structures that mathematically guarantee convergence without coordination:

• G-Counter — Grow-only counter (each region has its own counter, sum on read)
• OR-Set — Observed-remove set (add/remove without conflicts)
• LWW-Register — Last-writer-wins for single values

Application-Level Merge#

Custom logic for your domain:

// Shopping cart merge: union of items, max of quantities
function mergeCart(cartA, cartB) {
  const merged = new Map();
  for (const [item, qty] of [...cartA, ...cartB]) {
    merged.set(item, Math.max(merged.get(item) || 0, qty));
  }
  return merged;
}

Version Vectors#

Track causality to detect true conflicts vs sequential updates. Only flag genuine conflicts for resolution.

Compliance and Data Residency#

GDPR Data Residency#

EU regulation requires that personal data of EU citizens can be stored and processed only in approved jurisdictions.

Architecture pattern:

EU user data → EU region ONLY (never replicated outside)
US user data → US region (can replicate to EU if needed)
Metadata/analytics → Any region (if anonymized)

Implementation Strategies#

1. Region-pinned tables — User data stays in the user's home region
2. Data classification — Tag data as "restricted" or "global," replicate accordingly
3. Encryption boundaries — EU data encrypted with EU-managed keys
4. Audit logging — Track every cross-region data movement

Multi-Region Databases#

CockroachDB#

Distributed SQL with configurable replication zones:

-- Pin EU user data to EU nodes
ALTER TABLE users CONFIGURE ZONE USING
  constraints = '{"+region=eu-west-1": 3}',
  num_replicas = 3;

-- Global tables replicated everywhere for low-latency reads
ALTER TABLE products CONFIGURE ZONE USING
  num_replicas = 5,
  constraints = '{"+region=us-east-1": 1, "+region=eu-west-1": 1, "+region=ap-southeast-1": 1}';

Google Cloud Spanner#

Globally consistent with TrueTime (atomic clocks + GPS):

• Externally consistent reads and writes
• Automatic sharding and replication
• 99.999% availability SLA

DynamoDB Global Tables#

Multi-region, multi-active with last-writer-wins:

• Automatic replication across selected regions
• Sub-second replication latency
• Conflict resolution via timestamps

Comparison#

Architecture Checklist#

Before going multi-region, verify these:

• Identify data locality requirements — What data must stay in which region?
• Choose consistency model per data type — Not everything needs strong consistency
• Design conflict resolution — What happens when two regions write the same row?
• Plan failover procedures — Automated detection, DNS TTL, connection draining
• Set up cross-region observability — Replication lag dashboards, per-region error rates
• Test with chaos engineering — Simulate region failures, network partitions
• Compliance audit — Verify data residency constraints are enforced

Key Takeaways#

1. Latency is physics — The only way to reduce it is to move compute closer to users
2. Consistency is a spectrum — Choose the weakest model your use case can tolerate
3. Active-active is powerful but expensive — Conflict resolution adds real complexity
4. Compliance drives architecture — Data residency requirements may dictate your region topology
5. Start active-passive — Graduate to active-active only when latency or availability demands it

---

Geo-distributed systems are among the hardest problems in backend engineering. If you're designing systems that serve users across regions, explore our architecture deep-dives at codelit.io.

This is article #169 in the Codelit engineering blog series.