Zero Copy Streaming in Apache Cassandra® 4.0

One of the very important qualities of a distributed database is the ability to quickly transfer data from one node to another. There are many scenarios in which this becomes pivotal:

• Bootstrapping a new server to scale out the cluster and handle the increased workload
• Repairing a cross-node inconsistency
• Decommissioning a redundant server to reduce infrastructure costs
• Quickly rebuilding or replacing a failed machine to decrease Mean Time To Recover

The last case may need an additional explanation. While Cassandra’s replication feature allows losing one of the nodes, a single failed server creates a potentially dangerous situation; where the next failed server will lead to data unavailability on the higher consistency levels. The longer this issue stays unresolved, the higher the chance of potential downtime!

To help build a dynamic infrastructure and decrease a potentially dangerous time span, Dinesh Joshi, Senior Engineering Manager at Apple, reworked the streaming mechanism of Apache Cassandra resulting in a whole new level of performance.

While it sounds simple, did it make any significant change? Yes, definitely! The Netflix team did extensive research, investigating different scenarios and hardware setups, and discovered an incredible time boost, such as streaming time being reduced from 70-90 minutes to approximately 25 minutes (scenario: failed node replacement, 500 GB per node, 6-node cluster of i3.4xl AWS instances).

The conclusions of their research were:

• In version 3.0, streaming times are inconsistent (high variability) and highly affected by the instance type (more powerful servers perform better)
• With Cassandra 4.0, streaming is approximately 5x faster and shows little variability in its performance
• Cassandra 4.0 streaming performance is not CPU-bound and remains consistent across different instance types (so even cheaper servers perform better)

We talk a lot about zero-copy streaming, but what does it actually mean? What exactly do we copy and why is it somehow zero? The answer requires some understanding of how memory works in modern computers. For security reasons, there are two different regions known as Kernel Space and User Space. When you retrieve data from a device or a disk, you first pass it to the Kernel Space and then it has to be copied to the User Space where it can be accessed by an application. The previous approach to streaming partitions followed this same path in order to serialize data and prepare it for transfer. This led to unnecessary operations and overhead costs, drastically raising memory usage and keeping the Garbage Collector busy as the object in memory wasn’t reused after the transmission. In version 4.0, the data is streamed directly from the Kernel Space without copying it to the User Space, which greatly decreases workloads on both writer and reader.

Note: Internode Encryption, when enabled, requires data to be copied to the User Space for encryption purposes, forcing Cassandra to follow the former pathway and automatically disables Zero Copy Streaming!

This improvement was so efficient that it led to an unforeseen consequence. Previously, streaming was limited by CPU and RAM performance, causing a bottleneck. Since the disks normally used with Cassandra are lightning-fast, simultaneous Zero Copy Streaming of multiple nodes could have saturated the entire network throughput. This would interfere with the normal data queries; in other words, the new streaming feature is too good! To address this issue, throttling was introduced and it is enabled by default limiting outbound streams to 200 Mbps or 25 MB/s.

Note: Throttling should be adjusted carefully as it may be potentially harmful if disabled.

Zero copy streaming is enabled by default but can be disabled in cassandra.yaml using the stream_entire_sstables configuration option.

Other important options are:

• stream_throughput_outbound_megabits_per_sec used to throttle the overall outbound streaming; the default value is 200 Mbps
• inter_dc_stream_throughput_outbound_megabits_per_sec used to throttle the outbound streaming traffic to non-local datacenters; the default value is 200 Mbps

It is also important to note that internode encryption, which may prevent Zero Copy Streaming from working, is configured via the server_encryption_options option.