Operators Developers Contributors Reference
Sidebar updates per section — preview limitation
Apache Cassandra

Performance Tuning Guide

This page consolidates performance tuning guidance for Apache Cassandra, linking to detailed documentation for each area.

Hardware and Deployment

Start with the right foundation:

  • Production Recommendations — hardware sizing, disk layout, network requirements

  • Use SSDs for data directories; spinning disks are not recommended for production workloads

  • Separate commitlog and data directories onto different physical devices when possible

JVM and Garbage Collection

Cassandra 6 supports JDK 17 and JDK 21. JDK 21 with Generational ZGC provides the best GC pause behavior:

-XX:+UseZGC -XX:+ZGenerational

Key JVM tuning areas:

  • Heap size: Set -Xms and -Xmx equal; typically 8-16 GB. Avoid heaps larger than 32 GB.

  • GC algorithm: Generational ZGC (JDK 21+), G1GC (JDK 17), or CMS (legacy)

  • Off-heap memory: Cassandra uses significant off-heap memory for memtables, bloom filters, and compression metadata. Monitor total process RSS, not just heap.

JVM options are configured in jvm-server.options, jvm17-server.options, or jvm21-server.options under the Cassandra conf/ directory.

Compaction Strategy Selection

The compaction strategy has a major impact on read/write performance:

Compaction Overview

Understanding how compaction works and when it runs.

Unified Compaction Strategy (UCS)

New in Cassandra 5.0+. A single strategy that adapts to different workloads. Recommended as the default for new deployments.

Size Tiered Compaction (STCS)

Good for write-heavy workloads. Default in older Cassandra versions.

Leveled Compaction (LCS)

Good for read-heavy workloads with low write amplification requirements.

Time Window Compaction (TWCS)

Optimized for time-series data with TTL-based expiration.

Compression

Compression reduces disk usage and I/O at the cost of CPU:

  • LZ4 (default) — best balance of speed and compression ratio

  • Zstd — better compression ratio, slightly more CPU

  • ZstdDictionary — new in Cassandra 6, provides improved compression for small SSTables

Tune chunk_length_in_kb based on your read patterns: smaller chunks improve random read performance, larger chunks improve sequential scan throughput.

Read Path Tuning

Key cassandra.yaml settings affecting read performance:

  • concurrent_reads — number of concurrent read threads (default: 32). Increase for SSD-backed storage.

  • read_request_timeout — timeout for read operations

  • key_cache_size — cache for partition key lookups (reduces disk seeks)

  • row_cache_size — cache for entire rows (use sparingly; most workloads benefit more from OS page cache)

See cassandra.yaml Reference for all configuration options.

Write Path Tuning

Key settings affecting write performance:

  • concurrent_writes — number of concurrent write threads (default: 32)

  • commitlog_sync — periodic (default, batches syncs) vs. batch (sync per write, higher durability)

  • memtable_heap_space and memtable_offheap_space — memory allocated to memtables before flushing

Monitoring for Performance

Use Cassandra Metrics to identify bottlenecks:

  • Read/write latency histograms — ReadLatency and WriteLatency per table

  • Compaction pending tasks — sustained high values indicate compaction can’t keep up

  • GC pause time — long pauses affect tail latencies

  • Memtable flush frequency — frequent flushes may indicate insufficient memtable memory

  • Dropped messages — indicates timeouts or overload

Use async-profiler for CPU and allocation profiling of running Cassandra instances.