PROVIDER Β· C++ rewritten low-latency wide-column database

ScyllaDB

ScyllaDB rewrites Cassandra in C++ on the Seastar framework, and its shard-per-core lock-free design delivers far better per-node throughput and tail latency than the JVM implementation, cutting cluster cost for high-write real-time systems.

Backend & Database Israel 2026.11.15
Backend & Database Israel C++ rewritten low-latency wide-column database
4pts

Pros

4 items
Shard-per-core lock-free architecture saturates cores and NUMA on a single node
CQL and Cassandra data model compatibility means swapping drivers then retesting
Vendor benchmarks show several times Cassandra throughput at lower P99 latency on equal hardware
Built-in workload prioritisation and per-shard memory avoid GC pauses

Cons

4 items
Tuning and operations tooling differ from Cassandra, forcing teams to rebuild expertise
A smaller community means fewer references for difficult issues
Backup and multi-tenancy live in commercial editions with open source gaps
Scarce Chinese material leans on vendor support

Product Information

scylladb
Category Backend & Database
Company ScyllaDB
Service Region Israel
Official Website https://www.scylladb.com/
Recommendation Score 4 pts
Published 2026-11-15

Detailed Notes

Overview

ScyllaDB launched in 2015 from Checkpoint Technologies, later renamed ScyllaDB, rewriting the Cassandra architecture in C++ with a shard-per-core lock-free design on its Seastar asynchronous framework, claiming several times the throughput and far lower tail latency on identical hardware.

For Backend & Database work the appeal is low migration cost plus resource efficiency, since CQL and the Cassandra data model stay compatible so switching drivers is enough to trial, one node saturates many cores, fewer nodes deliver the same throughput, and rack or cloud spend drops, fitting P99-sensitive recommendation and bidding systems.

Watch ecosystem and operations differences, because tuning parameters, metrics, and repair tooling differ from Cassandra so teams rebuild expertise, open source and enterprise editions split on backup, multi-tenancy, and support, and Chinese material is scarce.

Core Strengths

  • Shard-per-core lock-free architecture saturates cores and NUMA on a single node
  • CQL and Cassandra data model compatibility means swapping drivers then retesting
  • Vendor benchmarks show several times Cassandra throughput at lower P99 latency on equal hardware
  • Built-in workload prioritisation and per-shard memory avoid GC pauses

Product Ecosystem

ScyllaDB Open Source

The CQL-compatible core engine with multi-datacenter support and tunable consistency.

Enterprise and Cloud

Enterprise adds backup, multi-tenancy, and rolling upgrade guarantees while Cloud provides managed hosting.

Migration and ecosystem

Assessment tooling eases Cassandra migration alongside the Scylla Monitoring Stack.

Drawbacks

  • Tuning and operations tooling differ from Cassandra, forcing teams to rebuild expertise
  • A smaller community means fewer references for difficult issues
  • Backup and multi-tenancy live in commercial editions with open source gaps
  • Scarce Chinese material leans on vendor support

Use Cases

  • Cutting Cassandra cluster cost (β˜…β˜…β˜…β˜…β˜…) - Fewer nodes for equal throughput
  • P99-sensitive real-time systems (β˜…β˜…β˜…β˜…β˜…) - Low tail latency from a lock-free design
  • High-volume IoT and logging (β˜…β˜…β˜…β˜…) - Low-cost CQL migration
  • Projects needing a mature community (β˜…β˜…β˜…) - Smaller than Cassandra

Pricing

Plan Price Scope
Open Source $0 Self-hosted with full features, no support
Enterprise Annual per node Backup, multi-tenancy, and support
ScyllaDB Cloud Per instance and usage Managed with on-demand scaling

Value depends on nodes times unit price rather than a single node quote, since ScyllaDB costs more per node but needs fewer of them, so clusters above three nodes usually save once operations are counted.

FAQ

  • Does migrating from Cassandra require code changes? CQL and the data model are compatible so mostly drivers and configuration change, but consistency behaviour, paging, and lightweight transactions need retesting.

  • Why is per-node performance higher? Shard-per-core gives each core its own shard with no cross-core lock contention and no JVM GC pauses, saturating the machine.

  • Is it suitable for OLTP? It fits write-heavy primary-key access patterns, while multi-table and strongly transactional core ledgers belong in relational engines.

Visit ScyllaDB Website β†’ https://www.scylladb.com/
Visit Website β†—