avro-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Pankaj Shroff <shro...@gmail.com>
Subject Re: Bypassing "handshake" in Responder
Date Mon, 25 Feb 2013 19:15:23 GMT
Doug - thanks. I think I see what you are saying - but I was trying to
leverage Avro a bit more than perhaps currently designed.

There is a lot of great code in  SpecificResponder that uses JsonDecoder
(for example) to do the data binding between a SpecificData derived object
and its Json representation. The great thing about this is that it
validates against the protocol schema specified (or installed upon
construction of a SpecificResponder object). To be clear, here is an

1) I define Avro protocol file to describe my protocol (.avpr)
2) In the generate-sources phase, avro-maven plugin helps me to generate
Class types for my protocol objects and the interface for the message
exchange (my protocol API). These extend the SpecificRecord object in Avro

3) I write an implementation of the protocol API by implementing the
interface generated by Avro

-- so far everything's great ----

4) Now I want to use the avro-ipc package to create client and server
classes (Requestor and Responder in avro-ipc)

5) My protocol does not "REQUIRE" Avro RPC framing - surely it CANNOT -
otherwise I would be forcing every client and server implementation to have
a dependency on Avro. They might just want to use direct Jackson or direct
Protobuf or direct Thrift support for encoding/decoding - or they may have
legacy code already in place which they would rather not refactor - just
because of the dependency I created.

6) I therefore need hooks to provide my own implementation of
SpecificDecoder and SpecificEncoder.

There is no way to do this.

If, as you suggest, I extend from the DatumReader/Writer classes directly,
I have to write all the code for codehaus.jackson's ObjectMapper and
JsonParser usage, which is currently nicely encapsulated in SpecificData. I
also have to write the invocation logic (that invokes the method
"respond()" in my implementation instance of my protocol API) which I would
have been just loved to reuse.

Also the hooks mentioned in (6) above would have allowed me to leverage in
the future any other encoder/decoder support Avro may provide (if at all) -
such as Protobuf, Thrift, etc.

In general what I was looking for is an "option" really - to tell Avro's
IPC API whether I am using Avro RPC framing or not! If not, it should just
let me handle the framing bits in a derived class or something and bypass
the handshake.

What do you think? :)



On Mon, Feb 25, 2013 at 1:38 PM, Doug Cutting <cutting@apache.org> wrote:

> This sounds like a different RPC wire format than Avro's.  Avro's
> Requestor and Responder implement Avro's RPC wire format.  Avro's
> Encode/Decoder and DatumReader/DatumWriter APIs should facilitate
> implementation of other RPC wire formats that include Avro data.
> Avro's Transceiver API may or may not be reusable, since it assumes
> Avro-style framing.  Parts of Requestor and Responder *might* be
> reusable and some refactoring of those classes *might* make such reuse
> easier, but there's not that much logic there that's not specific to
> Avro's wire format, so it might be just as easy to reimplement this
> layer for a different wire format.  It's hard for me to say without
> seeing a patch with a proposed refactoring.  Does that make sense?
> Doug
> On Mon, Feb 25, 2013 at 9:08 AM, Pankaj Shroff <shroffg@gmail.com> wrote:
> > Hi
> >
> > We are using Avro for implementing an open source reference
> implementation
> > of the OpenRTB protocol.
> >
> > We have made a design goal to model the protocol using Avro protocol
> files
> > (.avpr) and generate types defined in the protocol schema using Avro .
> The
> > challenge is that the protocol does not necessarily require the use of
> Avro/
> > Binary wire encoding - or even the use of Avro/ RPC context. In fact many
> > counter parties have proprietary implementations supporting either
> Protobuf
> > or Json encoding.
> >
> > Now, there is a Json encoder/decoder in the Avro package but it seems
> that
> > the approach is a "schema-first" approach. The JsonEncoder assumes that
> the
> > encoding on the wire still follows the Avro Json encoding - which
> includes a
> > handshake followed by schema confirmation on both sides (client and
> server).
> >
> > For the protocol we are implementing - this presents 2 problems if Avro/
> > binary is not the chose encoding type for both sides - and if instead,
> lets
> > say, raw Json encoding is being used
> >
> > 1) the handshake is rather Avro specific - and we would like to
> completely
> > skip it if both sides have agreed on using raw json encoding - there may
> be
> > a separate info-exchange phase in the protocol that is protocol specific
> and
> > does not need to use Avro handshake. Is it possible to use Avro RPC
> without
> > the handshake?
> >
> > 2) we would like to use the data binding and schema resolution as
> > implemented by the SpecificResponder class in Avro - but extend it to use
> > raw JSON - not Avro JSON - encodings.
> >
> > 3) We would prefer not to have to override the "respond(List<buffers>)"
> > method of the base class Responder. This implementation always performs
> > handshake and always uses BinaryEncoder/Decoder which removes any
> > flexibility of using a different encoder /decoder in a derived class. We
> > would prefer if the Responder or some derived class saves the chosen
> > Decoder/ encoder as a protected property of the Responder object.
> Instead of
> > instantiating BinaryEncoder/ Decoder objects on the fly within the
> respond
> > method, it would be great if this was made more extensible and if the
> > Encoder/Decoder can be specified during construction.
> >
> > 4) For future flexibility it would be great to have the handshake
> > functionality available in sub-classes of Responder as an inherited
> method
> > (instead of private scope right now).
> >
> > I would welcome any suggestions/corrections.
> >
> > Pankaj
> >
> > --
> > Pankaj Shroff
> > twitter: @chompi
> > http://github.com/chompi
> > http://github.com/openrtb/openrtb2x
> >

Pankaj Shroff

View raw message