Schema Registry and Avro

Analogy categoryAnalogies

Why Kafka Needs a Schema Registry

Kafka itself treats every message as an opaque byte array; it has no built-in concept of structure, field names, or types. That's fine until multiple independent producers and consumers need to agree on what those bytes mean, at which point an undocumented or silently changed message format becomes a serious operational risk: a producer that renames a field or changes its type can break every downstream consumer without Kafka itself ever noticing or objecting. The Schema Registry, most commonly Confluent's, solves this by storing versioned schemas centrally and enforcing compatibility rules whenever a producer tries to register a new version.

🏏

Cricket analogy: Without a shared scorecard format, one scorer might record 'wide' as a boolean and another as a run count, and nobody notices until the two feeds disagree on the final total, which is exactly the ambiguity a standardized scoring rulebook prevents.

Avro Serialization and the Wire Format

Avro is a compact binary serialization format built around a JSON-defined schema; unlike formats like JSON that embed field names in every single message, Avro-encoded data contains no field names at all, just the values in schema-defined order, which makes it significantly smaller on the wire. When a producer using the Confluent serializer sends an Avro record, it doesn't embed the full schema either, instead it registers the schema with the Schema Registry (or reuses an existing registration), receives back a small integer schema ID, and prefixes the message with that ID; consumers then look up the schema by ID to decode the payload.

🏏

Cricket analogy: It is like referring to a bowler by their jersey number instead of spelling out their full name on every scorecard line, the roster (schema registry) maps number to name once, and every subsequent entry is compact.

Compatibility Modes and Schema Evolution

The Schema Registry enforces one of several compatibility modes when a new schema version is registered: BACKWARD compatibility means new consumers can read data written with the previous schema (the common default, achieved by only adding optional fields with defaults or removing fields), FORWARD compatibility means old consumers can read data written with the new schema, and FULL compatibility requires both directions to hold simultaneously. Picking the wrong mode, or the wrong kind of change for that mode, causes the registry to reject the schema registration outright at write time, which is a deliberate design choice: failing fast at schema registration is far cheaper than discovering a broken consumer in production hours later.

🏏

Cricket analogy: BACKWARD compatibility is like a new scoring app being able to read last season's archived scorecards even though it now supports an extra optional stat field, while FORWARD compatibility is like an old scoring app still being able to parse this season's scorecards despite the new field it doesn't understand.

json

{
  "type": "record",
  "name": "OrderEvent",
  "namespace": "com.shop.orders",
  "fields": [
    { "name": "orderId", "type": "string" },
    { "name": "customerId", "type": "string" },
    { "name": "amount", "type": "double" },
    { "name": "currency", "type": "string", "default": "USD" }
  ]
}

Adding the 'currency' field with a default value of "USD" is a BACKWARD-compatible change: consumers built against the old schema simply won't see it, and consumers built against the new schema fall back to the default when reading old records that lack it.

Removing a required field, or changing a field's type from string to int, is almost never a compatible change under any mode; the Schema Registry will reject that registration, and the correct fix is usually to introduce a new field and deprecate the old one gradually rather than mutating it in place.

Kafka stores raw bytes with no built-in structure; the Schema Registry adds a central, enforced contract for message shape.
Avro's binary format omits field names from each message, using a schema-registry-assigned ID for compact encoding.
BACKWARD compatibility (the common default) lets new consumers read old data; FORWARD lets old consumers read new data; FULL requires both.
Compatible evolution typically means adding optional fields with defaults or removing already-optional fields.
Incompatible changes are rejected at registration time, catching problems before they reach production consumers.
The registry assigns each schema an integer ID, which is what's actually embedded in each Avro-encoded message.