Cassandra Read Path — MemTable, Row Cache, Bloom, SSTables

Why it matters

Read latency in Cassandra is usually dominated by SSTable access. Understanding what happens on read explains p99 latency and helps identify optimization opportunities.

The architecture

Coordinator forwards read to replicas per CL. Each replica reconstructs row from local storage.

Local read path: check MemTable, check row cache (if enabled), check SSTables (using bloom filter to skip empty ones).

How it works end to end

Bloom filter check per SSTable: skip if row provably absent. Otherwise, check partition summary + index to find offset.

Read from SSTable: seek to offset, read row block, apply row filter.

Reconcile: gather versions from all sources, take latest timestamp per column, apply tombstones.