Consistency

How it works

1. Choose model (strong/eventual)
2. Apply quorum or leader rules
3. Handle conflicts (CRDTs/outbox)
4. Validate with invariants

🎯 What is Consistency?

Consistency in distributed systems ensures that all nodes see the same data at the same time. It's like having a shared notebook - everyone writes on the same page, and no one can write on a different page until the first one is done.

Overview

• Governs what readers observe relative to recent writes across replicas.
• Strong: reads see latest committed write. Eventual: replicas converge over time.
• Tunable consistency lets callers pick (e.g., QUORUM vs ONE) per request.

When to use stronger guarantees

• Monetary transfers, inventory updates, auth/entitlements, uniqueness constraints.
• Workflows where dual-writes or replays cause significant harm.

Trade-offs

• Stronger consistency increases latency and reduces availability during partitions.
• Weaker consistency improves availability but requires idempotency and reconciliation.

Patterns

• Quorum reads/writes: W+R > N to ensure overlap.
• CRDTs/event sourcing for conflict-free merges.
• Sagas and outbox pattern for cross-service updates.
• Read your own writes via sticky routing or session tokens.

Anti-patterns

• Blind dual-writes to multiple stores without ordering/transactions.
• Assuming clock sync; prefer logical timestamps/vector clocks.
• Relying on caches without invalidation strategy.

📐 Quick Diagrams

  # Quorum (N=3, W=2, R=2)
  Client ▶ LB ▶ Replicas A,B,C
         └─ write to any 2
         └─ read from any 2
  

  # Outbox + Poller
  Service ▶ DB(outbox) ▶ Poller ▶ Broker ▶ Consumers
  

❓ Interview Q&A (concise)

• Q: CAP vs PACELC? A: CAP focuses on partitions; PACELC adds latency vs consistency trade-offs even without partitions.
• Q: Ensure idempotency? A: Use request IDs, upserts, de-dup tables, and at-least-once consumers with exactly-once semantics per key.
• Q: Read-after-write consistency? A: Route to primary, use sticky sessions, or quorum reads; avoid stale replicas.

📚 Consistency Models

ACID Properties

BASE Properties (Alternative to ACID)

🛠️ Achieving Consistency