Kafka JSON Schema to Avro

Every constraint keyword disappears

minimum, maximum, pattern, minLength, maxItems, uniqueItems and the rest have no Avro equivalent, because Avro describes an encoding and not a set of rules. After this conversion nothing in the schema or the serializer enforces any of them. That is not a shortcoming of the converter, it is what the two formats are, and the honest thing is to say which ones you just stopped enforcing.

Optional means something different in each

In JSON Schema a property is optional by being absent from required, and the key can simply not be there. Avro has no absent: every field is written. So an optional property becomes ["null", T] with a default of null, and null has to be the first branch because a union default belongs to the first branch. That shape is the only one that makes an Avro field genuinely optional.

// JSON Schema
"properties": {"note": {"type": "string"}},
"required": []

// Avro
{"name": "note",
 "type": ["null", "string"],
 "default": null}

integer becomes long unless the schema says it fits

JSON Schema's integer has no width. Avro's int is 32 bits and its long is 64, and choosing int for an unbounded integer produces a schema that fails to serialize the first large value. So long is the default here, and a maximum at or below 2147483647 gets you an int.

An enum only converts when its values are legal names

Avro enum symbols follow the name rule: letters, digits and underscores, not starting with a digit. So ["NEW", "DONE"] becomes an enum and ["in-progress", "done"] cannot, and the field is a plain string with the permitted set no longer expressed anywhere. Renaming the values is the fix if the set is closed and worth enforcing.

allOf, tuples and patternProperties have no equivalent at all

allOf is intersection and Avro has nothing that intersects two shapes. A tuple-typed array gives each position its own schema and an Avro array has one item type for every element, so it has to become a record with one field per position. patternProperties cannot be expressed: a map takes every string key or a record names them all, and there is nothing in between. Each of these is reported rather than approximated.

What this handles and what it does not

Draft-07 keywords, and $ref only within the pasted document: a reference to another file or a URL is not followed, because this page never fetches anything. A $ref cycle is cut and reported, since a recursive Avro type has to be written by hand. The generated schema is run back through our own Avro validator on every change, so anything unregistrable is reported here rather than by the registry.