For one of our key Kafka-based applications, we ensure that all messages in the stream have a common binary format, which includes (among other things) a version identifier and a schema identifier. The version refers to the format itself, and the schema refers to the "payload," which s the data for the application itself.
Because we have a small number of schemas (50-100) and we only introduce 5-10 per year, we stash the mapping from schema identifier to schema details for easy access in ZooKeeper. This does basically create 2 levels of serialization but most processing we do occurs just by reading the common format, and not deserializing the payload. Only specialized code has to do that extra step. On Jun 12, 2013, at 12:50 PM, "Shone Sadler" <shone.sad...@gmail.com> wrote: > Jun, > I like the idea of an explicit version field, if the schema can be derived > from the topic name itself. The storage (say 1-4 bytes) would require less > overhead than a 128 bit md5 at the added cost of managing the version#. > > Is it correct to assume that your applications are using two schemas then, > one system level schema to deserialize the schema id and bytes for the > application message and a second schema to deserialize those bytes with the > application schema? > > Thanks again! > Shone > > > On Wed, Jun 12, 2013 at 11:31 AM, Jun Rao <jun...@gmail.com> wrote: > >> Actually, currently our schema id is the md5 of the schema itself. Not >> fully sure how this compares with an explicit version field in the schema. >> >> Thanks, >> >> Jun >> >> >> On Wed, Jun 12, 2013 at 8:29 AM, Jun Rao <jun...@gmail.com> wrote: >> >>> At LinkedIn, we are using option 2. >>> >>> Thanks, >>> >>> Jun >>> >>> >>> On Wed, Jun 12, 2013 at 7:14 AM, Shone Sadler <shone.sad...@gmail.com >>> wrote: >>> >>>> Hello everyone, >>>> >>>> After doing some searching on the mailing list for best practices on >>>> integrating Avro with Kafka there appears to be at least 3 options for >>>> integrating the Avro Schema; 1) embedding the entire schema within the >>>> message 2) embedding a unique identifier for the schema in the message >> and >>>> 3) deriving the schema from the topic/resource name. >>>> >>>> Option 2, appears to be the best option in terms of both efficiency and >>>> flexibility. However, from a programming perspective it complicates the >>>> solution with the need for both an envelope schema (containing a "schema >>>> id" and "bytes" field for record data) and message schema (containing >> the >>>> application specific message fields). This requires two levels of >>>> serialization/deserialization. >>>> Questions: >>>> 1) How are others dealing with versioning of schemas? >>>> 2) Is there a more elegant means of embedding a schema ids in a Avro >>>> message (I am new to both currently ;-)? >>>> >>>> Thanks in advance! >>>> >>>> Shone >>>> >>> >>> >>
