Avro Basics

What Is Avro#

Avro is a schema format developed by Apache, commonly used with Kafka. Unlike Protobuf which uses field numbers, Avro uses field names — same as JSON.

{
  "type": "record",
  "name": "Order",
  "fields": [
    { "name": "order_id", "type": "int" },
    { "name": "amount",   "type": "float" }
  ]
}

No field numbers. Fields are identified by name.

The Problem With Field Names#

If identity is based on name, renaming a field is a breaking change:

{ "name": "amount" }   →   { "name": "price" }

Old messages have amount. New consumer looks for price. No match — field is lost.

This is the fundamental tradeoff vs Protobuf:

Writer Schema vs Reader Schema#

Avro solves compatibility differently from Protobuf. Instead of field numbers, it uses schema resolution — comparing the schema the producer used (writer schema) against the schema the consumer has (reader schema).

Writer schema (producer):   { order_id, amount }
Reader schema (consumer):   { order_id, amount, currency }

Avro compares them field by field and resolves differences:

• Field in reader but not in writer → use the default value defined in reader schema
• Field in writer but not in reader → ignore it
• Field in both → deserialize normally

Default Values Are Mandatory for New Fields#

Since Avro relies on defaults for missing fields, every new field must have a default value:

{
  "name": "currency",
  "type": "string",
  "default": "USD"    ← required for backward compatibility
}

If you add a field without a default, old messages (which don't have this field) cannot be read by the new consumer. Schema Registry will reject it.

Why Schema Registry Is Critical for Avro#

Protobuf embeds field numbers in the binary — the consumer can partially decode even without the exact writer schema.

Avro binary is just values, no field names or numbers:

Binary: [123] [50.0]   ← just raw values, no labels

Without the writer schema, the consumer has no idea what these bytes mean. It cannot deserialize at all.

This is why Schema Registry is not optional for Avro — the consumer must fetch the exact writer schema to make sense of the binary data.