[
https://issues.apache.org/jira/browse/AVRO-769?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12997649#comment-12997649
]
Scott Carey commented on AVRO-769:
----------------------------------
Some extra comments here:
I did not move in configuration items on Json, Validating, or Resolving
decoder/encoders.
The intention is that Encoder/Decoder and the two factories will be stable.
Json : requires a Schema, and can't change its schema (currently), only its
input/output. I left reconfiguring its output out of the factory and put it on
the objects. This could be moved in, and would seem to be very stable.
Validating: requires a schema, and an input/output decoder/encoder. We could
expose the reconfiguration of the encoder/decoder, but it wasn't in the old
API, and I'm not sure how useful that is versus just making a new one.
Resolving: This is where things get more tricky. Not only does this require
two schemas, but it involves an API I'm not sure I can say will be stable on
the resolver side. Having it swap out the underlying decoder/encoder is only
sometimes used, and that use is questionable as is. So I did not want to make
that part of this 'stable' api yet. I did make one constructor form private
which is reduced functionality.
In the long run I want to get rid of this class and separate resolution of
schema pairs from decoders entirely. The current scheme requires that a user
traverse the schema AND use the parser to get the job done which is inefficient
and error prone.
So the summary is that although we could add more in the factories, I would
rather leave them tied to the basics as much as possible -- abstracting away
binary encoder/decoder subtypes, and providing construction facilities in
general. Configuration is only provided for the binary use cases. The binary
encoder/decoder have to hide their configuration options better because of the
implementation types. Only ResolvingDecoder has similar configuration
complexity and that should be dealt with if and when it changes.
> Java: Align Decoder/Encoder APIs for consistency and long term stability
> -------------------------------------------------------------------------
>
> Key: AVRO-769
> URL: https://issues.apache.org/jira/browse/AVRO-769
> Project: Avro
> Issue Type: Improvement
> Components: java
> Reporter: Scott Carey
> Assignee: Scott Carey
> Priority: Blocker
> Fix For: 1.5.0
>
> Attachments: AVRO-769.v1.patch
>
>
> As part of AVRO-753, we modified the Encoder API to be more like the current
> Decoder API. This issue tracks related changes to solidify the API of both
> Encoder and Decoder to be more stable and consistent. It is expected that
> the result will be long-lived and not require major changes in the future for
> the following reasons:
> * Instantiation and configuration will be funneled through EncoderFactory and
> DecoderFactory. Individual implementation types and constructors are not
> exposed. With this abstraction we could, for example, put the features of
> BlockingBinaryEncoder into BufferedBinaryEncoder and not break any user code.
> We already have some of this distinction on the Decoder side, but not all
> BinaryDecoders are going through the factory.
> * The core Encoder and Decoder abstract classes will not declare
> configuration methods or constructors. This makes them 'pure' low level Avro
> read/write API constructs. This separation of concerns means, for example,
> that not all encoder implementations need wrap an OutputStream because of
> init(OutputStream out).
> * The core Encoder and Decoder API does not know or care about Schemas,
> resolution, or any other 'higher order' Avro concept. This is the pure
> separation of concern for writing/reading primitive Avro types to/from
> somewhere.
> * Implementations have been heavily performance tuned on both sides, so
> changes to the API necessary for high performance will not be likely.
> The Factories will adhere to the following general principles:
> * configuration options that do not affect the semantics of a type can be set
> through the factory. i.e. buffer sizes.
> * configuration that affects the semantics or changes the output or supported
> input will have separate factory methods. For example, choosing between an
> implementation that requires calling flush() and one that does not, requires
> choosing a different factory method to instantiate. This is important
> because it generally means that client code explicitly requests the
> behavioral type, and that helps prevent bugs caused by accidentally
> configuring a factory to return an object that is incompatible with the use
> case.
--
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira
