Eventual Consistency

If no new updates are made to a piece of data, all nodes will eventually converge to the same value — given enough time.

Consistency Models (Spectrum)

Model	Guarantee	Performance
Strong Consistency	All reads see latest write immediately	Slower
Read Your Writes	You always see your own writes	Medium
Monotonic Read	Never see older data than before	Medium
Causal Consistency	Causally related ops seen in order	Medium-fast
Eventual Consistency	Nodes converge eventually	Fastest

How It Works Internally — Cassandra Example

Step 1: Write balance = 800 → Node 1 (acknowledged)
Step 2: Node 1 asynchronously syncs to Node 2, Node 3
Step 3: Read from Node 2 immediately → might see 500 (not synced yet)
Step 4: Read from Node 2 after sync → sees 800

Window of inconsistency = milliseconds usually

Conflict Resolution Strategies

• Last Write Wins (LWW): Highest timestamp wins. Simple but can lose data. Used by Cassandra default.
• Vector Clocks: Track causality between writes. More accurate but complex. Used by DynamoDB, Riak.
• CRDT (Conflict-free Replicated Data Types): Data structures that auto-merge. Counter: add all increments. Set: union of all sets. Used by Redis, Google Docs.
• Application-level Resolution: DB surfaces the conflict, application decides how to merge. Used by Amazon Dynamo.

When Acceptable vs Not Acceptable

Acceptable (AP)	Not Acceptable (CP needed)
Social media likes/views	Bank balance
DNS propagation	Flight seat booking
Product catalog	Stock trading
User profile updates	Authentication tokens
Shopping cart (Amazon)	Financial transactions