Apache Parquet vs New File Formats: A New Era in Analytics

Apache Parquet Vs New File Formats. There’s no denying the impact Apache Parquet has had on the analytics ecosystem. - Columnar layout - Strong compression & encoding - The de facto format in data lakes and warehouses for over a decade But there are also a lot of new things happening with our analytical workloads. ❌ Workloads are no longer just batch analytics - they involve AI pipelines, low-latency use cases, and hardware-accelerated compute ❌ Hardware is not only CPU but also GPU, from x86 to ARM/RISC-V, with wide SIMD instruction sets becoming the norm ❌ And performance bottlenecks are seen in decompression speed, memory pressure and lack of vectorized execution paths As a result, we’re seeing a wave of innovation around next-gen file formats. Efforts like: BtrBlocks – introduces cascaded compression with lightweight codecs Nimble (Meta) – optimized for fast scan & inference LanceDB – tailored for vector search and ML use cases And the past week I read about “Fastlanes” Here are some highlights from their VLDB paper ✅ Expression Encoding: Flexible chains of lightweight codecs (FFOR, DELTA, DICT, FSST) that outperform heavyweight compression like Zstd ✅ Multi-Column Compression: Exploits inter-column correlations (e.g. equality, one-to-one mappings) to go beyond traditional columnar compression ✅ Segmented Layout: Decompresses in small vectors (1024 values) rather than rowgroups, reducing memory pressure and improving cache efficiency ✅ Compressed Execution Support: Returns compressed vectors to engines like DuckDB/Velox for SIMD/GPU-friendly query execution Here are some numbers I picked up on Performance: - 800× faster random access than Parquet+Zstd - 41% better compression than Parquet+Snappy - 2% better compression than Parquet+Zstd (without using heavyweight codecs) - Decoding accelerates with AVX512 (up to 40% faster) I think we are going to see more of fit-for-purpose formats in the future as we tackle specific use cases around AI workloads and to target performance. Paper link in comments. #dataengineering #softwareengineering