Yes, I think it's inevitable that there is some kind of registry or database of schemas, since a fingerprint by itself does not contain enough information to be able to decode a record. https://issues.apache.org/jira/browse/AVRO-1704 calls this a SchemaStore.
In some cases, I imagine an application might be able to embed all schema versions in circulation into the application code. Then the SchemaStore could simply be an in-memory hash map. In this case, you'd have to be careful in deploying the application: when you want to add a new schema version, you first have to update all running instances to support the new schema (but not generate any records encoded with the new schema), and then in a second phase you update the application to start encoding records with the new schema. In general, the SchemaStore might have to be a remote database, or something like the Confluent schema registry (http://docs.confluent.io/1.0/schema-registry/docs/intro.html). In that case, if an instance of the application encounters a schema fingerprint that it has not seen before, it can fetch the full schema from that database/registry. The fingerprint could be a hash of the schema, bit it would also be ok to assign sequential numbers to schema versions, as long as you have some way of globally keeping track of those version numbers (this is what the Confluent schema registry does). Martin > On 23 Dec 2015, at 12:19, Niels Basjes <ni...@basjes.nl> wrote: > > Yes, as a generic crossplatform solution this makes a lot of sense. It s > easy to build and stops consuming the messages as soon as the fingerprint > changes. > In my corporate reality I see that a source of messages puts them into > Kafka, then several consumers read and deserialize them in near-realtime. > In the static situation of "the schema doesn't change" having a fingerprint > of the schema would suffice. > But now the schema changes. I would like to deploy the new version of the > source system, and consume these new messages using the schema migration > idea. I.e. read the records as long as they are backwards compatible. > As a general assumption: it is impossible to deploy a new version of all > consuming applications at the same time. > I want as little down time as possible and I do not want to lose any > messages. > > Is this possible with a simple fingerprint? > The only direction I can think of right now is by having a central registry > that contains all schemas ever used, indexed by their fingerprint. > > What alternative solutions can we think of? > > Niels > > On Wed, 23 Dec 2015, 01:42 Sean Busbey <bus...@cloudera.com> wrote: > >> Including a schema fingerprint at the start >> >> 1) reuses stuff we have >> 2) gives a language independent notion of compatibility >> 3) doesn't bind how folks get stuff in/out of the single record form. >> >> -- >> Sean Busbey >> On Dec 22, 2015 06:52, "Niels Basjes" <ni...@basjes.nl> wrote: >> >>> I was not clear enough in my previous email. >>> What I meant is to 'wrap' the application schema in a serialization >> wrapper >>> schema that has a field indicating the "schema classname". >>> That (generic setup) combined with some generated code in the schema >>> classes should yield a solution that supports schema migration. >>> >>> Niels >>> >>> On Tue, Dec 22, 2015 at 11:55 AM, Niels Basjes <ni...@basjes.nl> wrote: >>> >>>> Thanks for pointing this out. >>>> This is exactly what I was working on. >>>> >>>> The way I solved the 'does the schema match' question at work is by >>>> requiring that all schema's start with a single text field "schema >>>> classname" being the full class name of the class that was used to >>> generate >>>> it. >>>> That way we can have newer versions of the schema and still be able to >>>> unpack them. In this form the classname is essentially an indicator if >>>> schema migration is possible; even though the schemas are different. >>>> >>>> What do you think of this direction? >>>> >>>> Niels >>>> >>>> >>>> On Mon, Dec 21, 2015 at 11:30 PM, Ryan Blue <b...@cloudera.com> wrote: >>>> >>>>> Niels, >>>>> >>>>> This sounds like a good idea to me to have methods like this. I've had >>> to >>>>> write those methods several times! >>>>> >>>>> The idea is also related to AVRO-1704 [1], which is a suggestion to >>>>> standardize the encoding that is used for single records. Some >> projects >>>>> have been embedding the schema fingerprint at the start of each >> record, >>> for >>>>> example, which would be a helpful thing to do. >>>>> >>>>> It may also be a good idea to create a helper object rather than >>>>> attaching new methods to the datum classes themselves. In your example >>>>> below, you have to create a new encoder or decoder for each method >>> call. We >>>>> could instead keep a backing buffer and encoder/decoder on a class >> that >>> the >>>>> caller instantiates so that they can be reused. At the same time, that >>>>> would make it possible to reuse the class with any data model and >> manage >>>>> the available schemas (if embedding the fingerprint). >>>>> >>>>> I'm thinking something like this: >>>>> >>>>> ReflectClass datum = new ReflectClass(); >>>>> ReflectData model = ReflectData.get(); >>>>> DatumCodec codec = new DatumCodec(model, schema); >>>>> >>>>> # convert datum to bytes using data model >>>>> byte[] asBytes = codec.toBytes(datum); >>>>> >>>>> # convert bytes to datum using data model >>>>> ReflectClass copy = codec.fromBytes(asBytes); >>>>> >>>>> What do you think? >>>>> >>>>> rb >>>>> >>>>> >>>>> [1]: https://issues.apache.org/jira/browse/AVRO-1704 >>>>> >>>>> >>>>> On 12/18/2015 05:01 AM, Niels Basjes wrote: >>>>> >>>>>> Hi, >>>>>> >>>>>> I'm working on a project where I'm putting Avro records into Kafka >> and >>> at >>>>>> the other end pull them out again. >>>>>> For that purpose I wrote two methods 'toBytes' and 'fromBytes' in a >>>>>> separate class (see below). >>>>>> >>>>>> I see this as the type of problem many developers run into. >>>>>> Would it be a good idea to generate methods like these into the >>> generated >>>>>> Java code? >>>>>> >>>>>> This would make it possible to serialize and deserialize singles >>> records >>>>>> like this: >>>>>> >>>>>> byte [] someBytes = measurement.toBytes(); >>>>>> Measurement m = Measurement.fromBytes(someBytes); >>>>>> >>>>>> Niels Basjes >>>>>> >>>>>> P.S. possibly not name it toBytes but getBytes (similar to what the >>>>>> String >>>>>> class has) >>>>>> >>>>>> public final class MeasurementSerializer { >>>>>> private MeasurementSerializer() { >>>>>> } >>>>>> >>>>>> public static Measurement fromBytes(byte[] bytes) throws >>>>>> IOException { >>>>>> try { >>>>>> DatumReader<Measurement> reader = new >>>>>> SpecificDatumReader<>(Measurement.getClassSchema()); >>>>>> Decoder decoder = >>> DecoderFactory.get().binaryDecoder(bytes, >>>>>> null); >>>>>> return reader.read(null, decoder); >>>>>> } catch (RuntimeException rex) { >>>>>> throw new IOException(rex.getMessage()); >>>>>> } >>>>>> } >>>>>> >>>>>> public static byte[] toBytes(Measurement measurement) throws >>>>>> IOException { >>>>>> try { >>>>>> ByteArrayOutputStream out = new ByteArrayOutputStream(); >>>>>> Encoder encoder = >> EncoderFactory.get().binaryEncoder(out, >>>>>> null); >>>>>> SpecificDatumWriter<Measurement> writer = new >>>>>> SpecificDatumWriter<>(Measurement.class); >>>>>> writer.write(measurement, encoder); >>>>>> encoder.flush(); >>>>>> out.close(); >>>>>> return out.toByteArray(); >>>>>> } catch (RuntimeException rex) { >>>>>> throw new IOException(rex.getMessage()); >>>>>> } >>>>>> } >>>>>> } >>>>>> >>>>>> >>>>>> >>>>>> >>>>> >>>>> -- >>>>> Ryan Blue >>>>> Software Engineer >>>>> Cloudera, Inc. >>>>> >>>> >>>> >>>> >>>> -- >>>> Best regards / Met vriendelijke groeten, >>>> >>>> Niels Basjes >>>> >>> >>> >>> >>> -- >>> Best regards / Met vriendelijke groeten, >>> >>> Niels Basjes >>> >>
