Data Locality

The Physics You Can’t Negotiate With

 speed of light sets floor latency:

 Tokyo → Virginia:      ~170ms RTT (fiber path reality)
 Frankfurt → Virginia:  ~90ms
 within-region AZ-AZ:   ~1–2ms

 a page needing 10 sequential cross-region round trips:
 10 × 100ms = 1 full second of pure transit. before compute.

 LOCAL data + local compute: sub-10ms interactions.

 locality isn't an optimization; at global scale it's
 THE product-feel difference. users feel distance directly.

The Homing Model

 assign each datum a HOME region:

 user profile:        home = signup region (EU-user → EU)
 their posts/likes:   home follows profile home
 sessions:            region-local by construction

 [EU store: EU-users' data]──replicate──►[US][APAC read-only]
 [US store: US-users' data]─────────────►[EU][APAC read-only]

 WRITES go home (single-writer = no conflicts);
 READS served locally from replicas.
 = active-active's pragmatic pattern, applied per-entity.

Read Paths vs Write Paths

 LOCAL reads (the common case):
   profile views, feeds, product pages → regional replicas ✓
   
 WRITE-home routing:
   user edits → route to home region → replicate out
   
 CROSS-HOME reads:
   EU-user views US-friend's post:
     option A: read US replica in EU (laggy-tolerant content)
     option B: async fan-out — friend's post ALREADY replicated
               into EU follower timelines at write time ✓
   social platforms choose B aggressively: pre-computed local
   copies of everything you'll want to see.

 design instinct: move DATA to readers ahead of time;
 don't drag readers to data across oceans.

When Locality Breaks

Scenario	Consequence	Mitigation
User travels regions	far-from-home writes	accept write latency OR re-home tooling
Region failure	homes stranded	failover re-homing runbooks
Cross-region features (shared docs)	no single home fits	partition per-OBJECT or accept one-region latency for that feature
Global uniqueness needs	cross-region coordination	pre-allocated ID ranges / UUIDs

 RE-HOMING: migrating a user's data between regions
 (they moved to Tokyo permanently). background copy +
 cutover + reverse-sync — treat as a project feature,
 not an afterthought.

Interview Framing

“Global collaboration app feels slow for Asia-Pacific users” scored shape: light-latency arithmetic first, homing model proposed (writes-home/reads-local), replication topology drawn, cross-home interaction pattern chosen (fan-out vs remote-read), travel/re-homing edge acknowledged. Data locality questions test whether candidates think in PHYSICS first and architecture second — start with milliseconds.