SE-Radio Episode 354: Avi Kivity on ScyllaDB.mp3

Performance Metrics

The comparison between ScyllaDB and Cassandra reveals significant differences in throughput and latency metrics. Avi Kivity explains that ScyllaDB can achieve a million operations per node, scaling horizontally with additional nodes, while Cassandra requires a larger cluster to reach similar performance levels 1. This efficiency translates to substantial cost savings for enterprises, as fewer nodes are needed to achieve the same throughput 1.

The less hardware you need, then the more savings you get. Or sometimes conversely, with the same amount of hardware you can do more or achieve lower throughput.

--- Avi Kivity

The choice of programming language also plays a role, with ScyllaDB implemented in C for greater control and performance compared to Cassandra's Java implementation 1.

   

Garbage Collection

Garbage collection presents a significant challenge for databases like Cassandra, impacting latency and performance. Avi Kivity highlights that garbage collection can pause nodes for hundreds of milliseconds, disrupting distributed environments and affecting latency-sensitive applications 2. ScyllaDB addresses these issues with a log-structured allocator that manages heap fragmentation without traditional garbage collection, allowing for more predictable performance 2.

We already know what the garbage is, it's explicitly freed.

--- Avi Kivity

This approach ensures that ScyllaDB maintains control over memory management, avoiding the unpredictability associated with Java's garbage collection process 3.