How do consistency levels and replication interact in Cassandra?

The coordinator contacts the replicas that own the token and waits for the requested number of responses. Consistency is chosen per query, so the same table can serve both fast, stale-tolerant reads and slower, stronger reads.

CONSISTENCY LOCAL_QUORUM;
SELECT * FROM orders WHERE customer_id = 7f1c2a3e;

CONSISTENCY LOCAL_ONE;
SELECT * FROM product_cache WHERE product_id = 42;

Levels and their trade-offs

• ONE: one replica, lowest latency, highest chance of stale data
• LOCAL_ONE: one replica in the local datacenter, useful for caches
• QUORUM: majority across all datacenters, stronger but pays cross-region latency
• LOCAL_QUORUM: majority within the local datacenter, the usual balance
• EACH_QUORUM: a quorum in each datacenter, stricter for multi-region writes
• ALL: every replica, highest consistency and least tolerant of failures

With RF=3 and LOCAL_QUORUM, two replicas must respond. If writes also use LOCAL_QUORUM, the write and read sets overlap within the datacenter. Cross-datacenter replication is asynchronous, so a LOCAL_QUORUM read in another region may not see the write until it replicates.