Ballista is a distributed compute platform primarily implemented in Rust, and powered by Apache Arrow and DataFusion. It is built on an architecture that allows other programming languages (such as Python, C++, and Java) to be supported as first-class citizens without paying a penalty for serialization costs.
The foundational technologies in Ballista are:
- Apache Arrow memory model and compute kernels for efficient processing of data.
- Apache Arrow Flight Protocol for efficient data transfer between processes.
- Google Protocol Buffers for serializing query plans.
- Docker for packaging up executors along with user-defined code.
Ballista can be deployed as a standalone cluster and also supports Kubernetes. In either case, the scheduler can be configured to use etcd as a backing store to (eventually) provide redundancy in the case of a scheduler failing.
Refer to the core Ballista crate README for the Getting Started guide.
Ballista uses the DataFusion query execution framework to create a physical plan and then transforms it into a distributed physical plan by breaking the query down into stages whenever the partitioning scheme changes.
Specifically, any RepartitionExec operator is replaced with an UnresolvedShuffleExec and the child operator
of the repartition operator is wrapped in a ShuffleWriterExec operator and scheduled for execution.
Each executor polls the scheduler for the next task to run. Tasks are currently always ShuffleWriterExec operators
and each task represents one input partition that will be executed. The resulting batches are repartitioned
according to the shuffle partitioning scheme and each output partition is streamed to disk in Arrow IPC format.
The scheduler will replace UnresolvedShuffleExec operators with ShuffleReaderExec operators once all shuffle
tasks have completed. The ShuffleReaderExec operator connects to other executors as required using the Flight
interface, and streams the shuffle IPC files.
Ballista implements a similar design to Apache Spark, but there are some key differences.
- The choice of Rust as the main execution language means that memory usage is deterministic and avoids the overhead of GC pauses.
- Ballista is designed from the ground up to use columnar data, enabling a number of efficiencies such as vectorized processing (SIMD and GPU) and efficient compression. Although Spark does have some columnar support, it is still largely row-based today.
- The combination of Rust and Arrow provides excellent memory efficiency and memory usage can be 5x - 10x lower than Apache Spark in some cases, which means that more processing can fit on a single node, reducing the overhead of distributed compute.
- The use of Apache Arrow as the memory model and network protocol means that data can be exchanged between executors in any programming language with minimal serialization overhead.