Systems

The C++ library is structured as a collection of subsystems under cpp/src/arrow/ plus two sister projects (cpp/src/parquet/ and cpp/src/gandiva/). This section provides a deep dive into each subsystem.

Pages

• Arrays and types — cpp/src/arrow/array/, cpp/src/arrow/type.h, scalars, tables, record batches.
• I/O and filesystem — cpp/src/arrow/io/, cpp/src/arrow/filesystem/, local + S3 + GCS + Azure + HDFS.
• Compute kernels — cpp/src/arrow/compute/, scalar/vector/aggregate kernels and dispatch.
• Acero — cpp/src/arrow/acero/, the streaming execution engine.
• Dataset — cpp/src/arrow/dataset/, partitioned multi-format datasets.
• Parquet — cpp/src/parquet/, the Parquet C++ implementation.
• Gandiva — cpp/src/gandiva/, LLVM-based expression compilation.
• CSV and JSON — cpp/src/arrow/csv/, cpp/src/arrow/json/.
• Substrait — cpp/src/arrow/engine/substrait/, plan-format adapter.
• Util — cpp/src/arrow/util/, the kitchen sink (bit ops, futures, threading, compression, telemetry, ...).

Subsystems by file count (cpp/src/arrow only)

A rough sense of weight, using subdirectory file counts inside cpp/src/arrow/:

Subsystem	Files	Highlights
compute/ (incl. kernels/ and row/)	~120	Function registry, ~150 kernels
acero/	~80	14 ExecNode types, hash-join engines, TPC-H generator
flight/ (incl. sql/, transport/)	~70	Client, server, gRPC + UCX transports, Flight SQL
dataset/	~50	Format adapters, scanner, partitioning
array/	~50	Array hierarchy, builders, validation
util/	~250	Bit ops, futures, threading, compression, telemetry, RLE, hashing
io/ + filesystem/	~60	Local + cloud filesystems, compressed/buffered streams
ipc/	~30	IPC reader/writer + Feather
csv/ + json/	~50	CSV and JSON readers and writers

Subsystems outside arrow/

Subsystem	Path	Focus
Parquet	cpp/src/parquet/	Parquet reader/writer, encryption, page index, bloom filters
Gandiva	cpp/src/gandiva/	LLVM-based expression compiler
ORC adapter	cpp/src/arrow/adapters/orc/	Apache ORC integration
TensorFlow adapter	cpp/src/arrow/adapters/tensorflow/	Legacy TensorFlow IO
GPU support	cpp/src/arrow/gpu/	CUDA buffers and contexts
Tensor I/O	cpp/src/arrow/tensor/	Sparse + dense tensors
Telemetry	cpp/src/arrow/telemetry/	OpenTelemetry plumbing

Integration

The cpp/src/arrow/integration/ directory hosts the cross-language integration test driver. It implements:

• integration_command.cc — the CLI used by dev/archery/archery/integration/.
• JSON ↔ IPC encoding so any implementation can dump or read data in JSON form for round-trip checks.
• A test harness that invokes the C++ Arrow library's IPC reader and writer against fixture files generated by other implementations.

Cross-cutting reads

A few pieces span multiple subsystems and are worth highlighting separately:

• Async machinery. cpp/src/arrow/util/future.h, async_generator.h, thread_pool.h, task_group.h. This is the runtime that backs Acero, datasets, and Flight.
• Validity bitmap utilities. cpp/src/arrow/util/bit_util.h, bitmap.h, bitmap_ops.cc. Touched by every array operation that supports nulls.
• Compression codecs. cpp/src/arrow/util/compression*.cc. Used by IPC, Parquet, ORC, and the dataset framework.
• Hashing. cpp/src/arrow/compute/key_hash_internal.cc (with AVX2 variant) and arrow/util/hashing.h. Used by hash aggregates, hash joins, and dictionaries.
• CPU dispatch. cpp/src/arrow/util/cpu_info.cc, dispatch_internal.h. Picks the best SIMD code path at runtime (SSE4.2, AVX2, AVX-512, NEON, SVE).