Interesting that the two main "alternatives" mentioned here are columnar data storages. The problem I see with both Arrow and Parquet are that they are binary-

Interesting that the two main "alternatives" mentioned here are columnar data storages. The problem I see with both Arrow and Parquet are that they are binary-based files. Sure, binary files can be VERY efficient. But the "magic" behind files like CSV, JSON or YAML is that they are both "sufficiently" both human and machine readable. Maybe a CSV killer would be a human readable columnar based file format. Nevertheless, the article basically discusses the issues encountered with the manual "editability" of CSV files, not so much with its performance. It also mentions parquet or arrow and concedes that they require specialized format to read/write. If we are looking to that, then there are a lot of options such as sqlite format, BerkleDB (used by some cryptocurrency projects) among plenty of others.