Conflict Resolution

The Inevitable Moment

 multi-master writes meet:

 t=0  EU: cart = {apple}
 t=0  US: cart = {banana}     (same user, racing writes)

 replicas exchange → CONFLICT: which cart is true?
 {apple}? {banana}? {apple, banana}?

 every active-active design either:
 AVOIDS conflicts by routing, RESOLVES them deterministically,
 or DIVERGES silently (the bug you discover in court).

Strategy 1: Avoidance (the best resolution)

 single-writer-per-entity eliminates conflicts by design:

 user's data: only HOME region accepts writes
              others replicate read-only.
 cross-region flows via events/sagas, not dual-writes.

 [EU]──owns user-912──►writes ✓
 [US]──replica─────────►writes REJECTED/routed

 cost: home-region outage blocks those users' writes
       until re-homing/failover. bounded blast radius —
       usually the RIGHT tradeoff.

 most systems should stop here. the rest of this lesson
 is for designs that genuinely can't.

Strategy 2: Deterministic Rules

 when concurrent writes must BOTH apply:

 LAST-WRITE-WINS (LWW):
   keep higher timestamp. 
   ⚠ clock skew lies: EU's slow clock loses real-later writes.
   mitigate with hybrid clocks + skew monitoring;
   accept silent-lost-updates as the semantic.

 DOMAIN RULES (better):
   inventory: MIN(stock) wins (conservative)
   bids:      MAX wins;  balances: never below floor
   settings:  explicit merge policy per field
   
   domain rules encode INTENT — always preferable to
   time-based arbitration when a sensible rule exists.

Strategy 3: Mergeable Types (CRDTs)

 data structures whose states MERGE mathematically,
 order-independently:

 G-Counter: {EU: 5, US: 3} merge = per-node sum → 8 ✓
 LWW-Map:   per-key LWW semantics  
 OR-Set:    add/remove with unique tags; removes win ties

 properties: commutative + associative + idempotent merges
 → all replicas converge to identical state regardless
 of delivery order. no central authority needed.

 limits: works for specific shapes (counters/sets/registers);
 arbitrary relational logic doesn't fit. libraries exist;
 don't hand-roll the math.

Choosing Per Entity Type

Entity	Concurrent-write reality	Choice
User profile	rare, low harm	LWW + version UI
Cart	additive mostly	merge-union or home-region
Stock/balance	correctness-critical	single-writer / sync
Presence/counters	mergeable naturally	CRDT counters
Documents	human editing	OT/CRDT text engines

 decision tree:
 can one region own it? → route there (done!)
 truly concurrent? → domain rule if expressible
                    → CRDT if shape fits
                    → LWW with eyes open
                    → manual review queue (last resort)

Interview Framing

“Two regions update the same shopping cart simultaneously” scored answer: FIRST offer write-homing as the avoidance design (and why that suffices), THEN demonstrate resolution fluency: LWW-with-clock-caveats, domain-merge for carts specifically, CRDT mention scoped honestly to fitting shapes. Leading with avoidance shows judgment; the fallback taxonomy shows depth. Both together is the complete answer.