Apache Avro

Avro is a schema-on-the-side binary format from the Apache project. The writer and the reader both know a schema. The bytes are only the values, not the field names. The encoding rules are in the Avro specification (Apache-2.0). Cases are original.

A schema here is a small JSON description such as "int" or ["null", "string"]. The runner hands that schema to the decoder together with the bytes.

Versions

Version	Parser-visible difference
1.8	Primitive encodings: zigzag integers, length-prefixed strings and bytes.
1.11	Same primitives, plus a union of null and string.
1.12	Same encodings as 1.11 for these types.

A zigzag integer stores the sign in the least-significant bit so small negative numbers stay short. A union is a value that may be one of several types. The bytes start with the branch index (which arm), then the value of that arm.

What the cases cover

• Integers 0 and 7, string kelp, empty string, bytes
• Union [null, string] as a string and as null
• Rejects: truncated zigzag, union branch that does not exist

Catalog: compliance/data/avro/. The official Apache test_io encodings are vendored (avro-io-*, Apache-2.0). That is the same role JSONTestSuite plays for JSON: the project’s own binary-encoding table, stored as catalog JSON so nothing is downloaded at run time.

Python serializer

fastavro.schemaless_reader with the case’s schema. Schemaless here means “one Avro value, no object-container file header,” not “no schema.” The schema still comes from the catalog.

JavaScript avsc reads the same files. Mojo mojo-avro uses parse_avsc plus GenericDatum.decode on the same catalog. These primitives have been stable since 1.8, so the three columns should all pass on a current fastavro / avsc / mojo-avro.