ScyllaDB is a drop-in-compatible C++ reimplementation of Cassandra. Same CQL, same data model, same tools — but thread-per-core architecture, no JVM, and 3-10x per-node throughput. The 2026 question is no longer 'is it ready' but 'when does the migration pay back'.
Throughput delta
On the same hardware: Cassandra ~50K ops/sec/node, Scylla ~300-500K ops/sec/node. p99 latency: Cassandra 20-50ms, Scylla 2-5ms. The gap is real, not benchmark theater.
Why the delta exists
Seastar framework: thread-per-core, shared-nothing, async I/O. No GC pauses. Each shard owns a CPU. Compaction runs concurrently in the same process without preempting reads. Memory model is much more efficient than JVM off-heap dance.
CQL and tooling parity
Same CQL surface (with minor lag on new Cassandra-only features). cqlsh, drivers, nodetool-equivalent (nodetool itself is replaced by nodetool shim or scylla-tools). Existing schemas migrate as-is.
When to migrate
New project + Cassandra workload: start on Scylla. Existing Cassandra cluster: migrate when you'd otherwise add 3+ nodes for capacity. Migration tooling (Scylla Migrator) does live writes-replay-then-cutover. Plan a month for verification.
When NOT to migrate
Heavy use of Cassandra-specific features Scylla lags on (SAI indexes, some 5.x SSTable formats). Strong reliance on JVM ecosystem (Spark Cassandra connector — works on Scylla but with caveats). Tiny clusters (<5 nodes) — not enough gain to justify migration risk.