Hardwood 1.0, being a lightweight, multi-threaded alternative, bypasses bloated legacy pipelines to process an impressive 16.5 million rows per second across concurrent CPU cores.
On 22 June 2026, Hardwood launched as General Availability (GA) version 1.0.0. Kickstarted by Java Champion Gunnar Morling, the creator of the One Billion Row Challenge, the project was built AI-natively over a five-month development cycle, though code review and software design remained entirely human-owned.
Hardwood solves ‘heavyweight’ JVM constraints. Traditional Apache Parquet Java implementations, like parquet-java, introduce massive dependency footprints, pulling in Hadoop and Avro, along with classpath vulnerabilities, while relying on a single-threaded reader. Written from scratch to explore different engineering trade-offs, Hardwood serves as a fast, simpler, multi-threaded alternative for Java 21 or newer.
Unlike older sequential implementations, Hardwood spreads Parquet page decoding across all available CPU cores by default, saturating the host’s I/O and CPU bandwidth. The core framework features a zero-mandatory-dependency profile to minimise supply-chain risks; uncompressed or Gzip files require no third-party libraries. Optional codecs like LZ4, Snappy, Brotli, Zstd, or AWS S3 utilities attach as single-JAR extensions.
Filtered scans leverage branchless, batch-at-a-time evaluation to minimise CPU branch mispredictions. Data access is split into two clean layers: a structured Row Reader API for general record mapping, and a batch-oriented Column Reader API engineered for high-throughput analytics pipelines.
In performance scans on 8 vCPUs, the reader achieved a throughput of 16.5 million rows per second. Version 1.0 also ships with an interactive text-based user interface (TUI) command-line tool to inspect file schemas and structures directly without boilerplate code. Hardwood 1.0 is currently reader-only, with file-writing capabilities and Apache Arrow zero-copy interop on the roadmap.
