You may want to check out this Avro feature request: https://issues.apache.org/jira/browse/AVRO-1124
which has a lot of nice motivation and usage patterns. Unfortunately, its not yet a resolved request.
There are really two broad use cases.
1) The data are "small" compared to the schema (perhaps because its a collection or stream of records encoded by different schemas)
2) The data are "big" compared to the schema. (very big records or lots of records that share a schema)
Case (1) is often a candidate for a schema registry. Case (2) not as much.
Examples from my own usage:
For Kafka we include an MD5 digest of the writer's schema with each Message. It is serialized as a concatenation of the fixed-length MD5 and the binary Avro-encoded data. To decode we read off the MD5, look up the schema and use it to decode the remainder of the Message.
[You could also segregate data written with different schemas into different Kafka topics. By knowing which topic a message is under you then arrange a way to look up the writer's schema. That lets you avoid even the cost of including the MD5 in the Messages.]
In either case consumer code needs to look up the full schema from a "registry" in order to do the actual decode the Avro-encoded data. The registry serves the full schema that corresponds to the specified MD5 digest.
We use a similar technique for storing MD5-tagged Avro data in "columns" of Cassandra and so on.
Case (2) is pretty well handled by just embedding the full schema itself.
For example, for Hadoop you can just use Avro data files which include the actual schema in a header. All the record in the file then adhere to that same schema. In this case using a registry to get the writer's schema is not necessary.
Note: As described in the feature request linked above, some people use a schema registry as a way of coordinating schema evolution rather than just as a way of making schema access "economical".
On Aug 20, 2013, at 9:19 AM, Mark wrote:
> Can someone break down how message serialization would work with Avro and a schema registry? We are planning to use Avro with Kafka and I've read instead of adding a schema to every single event it would be wise to add some sort of fingerprint with each message to identify which schema it should used. What I'm having trouble understanding is, how do we read the fingerprint without a schema? Don't we need the schema to deserialize? Same question goes for working with Hadoop.. how does the input format know which schema to use?