I realize I forgot to give source references: * metadata version: https://github.com/apache/arrow/blob/a811bcd6b50755cc489ee43d8ede70cc8b9a7a25/format/Schema.fbs#L31-L52 * format version: https://github.com/apache/arrow/blob/a811bcd6b50755cc489ee43d8ede70cc8b9a7a25/format/Schema.fbs#L20-L27 Regards Antoine. Le 09/09/2026 à 17:50, Antoine Pitrou a écrit :
Hello, I think it might be useful to have a comparison point for the versioning discussion we're having. In the Arrow IPC format there are two versioning mechanisms: 1. A "metadata version", which records semantically breaking changes to the IPC metadata format (such as: a change of layout for some datatype). The metadata version (currently V5) is recorded in the Flatbuffers-serialized metadata, such that a reader can reject a file with an unsupported metadata version. 2. A "format version", which records non-breaking feature additions, such as new Arrow datatypes. The format version (currently 1.5) is *not* recorded in the Flatbuffers-serialized metadata, and this doesn't seem to have bothered anyone until now. Of course, Parquet is not Arrow, and our constraints and use cases are different. For example, Parquet is much more often used for long-term storage and publishing of datasets, while the Arrow IPC format is better suited for short-term transmission and storage (caches etc.). Regards Antoine.
