G-Research’s Adam Reeve describes Vortex, the columnar format originating at SpiralDB. A small postscript lets a reader fetch the file’s tail, find schema and segment locations, then request the relevant pieces. Columns can be chunked independently; logical types are separate from physical encodings, which can vary by chunk.
Trial compression samples approximately 1% of a chunk, with at least 1,024 values, and can stack lightweight encodings three levels deep. Zone maps cover roughly 8,000-row blocks, while filtering can operate on compressed data. Arrow-to-Vortex conversion is zero-copy; conversion back is zero-copy only for fully decompressed arrays. Embedded WebAssembly decoders remain a proposal. The account explains the format, without establishing G-Research production adoption.