It's sure 10 year after SEDA is quite smelly :-) In my mind the codec code should be used by a filter for transforming the bytes into pojos (like today) but really not dependent of MINA. IMHO demux handler is a piece of s..t, you should use a visitor pattern. Much more testable.
I like the loop until it's decoded idea, it very simple to understand. Le 31 déc. 2012 18:13, "Emmanuel Lécharny" <[email protected]> a écrit : > Le 12/31/12 7:55 AM, Julien Vermillard a écrit : > > Hi, > > > > Since few year, I stopped to use the MINA ProtocolCodecFilter and > > associated stuff (CumulativeCodec..). for implementing my own codec > > independent of MINA. > > > > it's just a service consuming ByteBuffer and pushing decoded POJO in a > > callback. The point is to be independent of MINA for example, parse & > save > > files using the codec, or simply implement an HTTP version of the > transport > > using old style servlet. > > > > Basically a decoder looks like : https://gist.github.com/4417934 > > One is instantiated by session. > > > > I'm quite happy with that and I think we should not port the old > > ProtocolCodeFilter to MINA 3.0 and replace it with a independent MINA > async > > decoder framework (consuming BB, accumulating if needed and producing > pojo). > It sounds a reasonnable proposal. > > If we think about it, decoding is not part of a filter chain : it > introduces a change of data type being passed from one filter to the > other, and if we have to cumulate data, we will just stop processing the > incomming data in the middle of the chain, the handler being unaware of > this fact. > > > Julien's proposal seems way better : the Handler would have a common > interface for encoding and decoding, used as a service when a > MessageReceived or a Write events are to be processed. This way, the > handler is fully in charge of all the aspects of the data processing, > including the accumulation of data. > > It won't either eliminate the existence of pre-written codec, like the > HttpCodec, or the Textline codec. We can even think about a chain of > codecs : one codec depends on the result of the previous codec. > > As far as I can tell, changing MINA this way will not impact ApacheDS, > even if we are using a DemuxIoHandler (the handler called depends on the > received message) : I don't see such a handler as a simplification over > a simple switch... > > > Keep in mind that the exisiting MINA logic depends on an idea which is > 10 years old : SEDA, and has not proven any advantage against simpler > implementations. It's also important to notice that SEDA implies that > each process part communicates with the next process (read : filter) by > the use of queues. This is highly costly and memory consuming. I'm not > sure that SEDA has anything to do with MINA implementation anwyay... > > On more thing : the current codec supposes that we pass a callback which > is called as soon as something has been decoded. This make the code > extremely complicated to debug. I'd rather have a system where we can > loop on the decoder, until it produces nothing. In other words, instead > of having something like : > > void myCallback( IoSession session, Object message ) { > // Do something > } > > void decode( IoSession session, ByteBuffer buffer, callback ) { > // Decode and call the callback > } > > > void messageReceived( IoSession session, ByteBuffer buffer ) { > decode( session, myCalback ); > ... > } > > > I would prefer something like : > > Object decode( IoSession session, ByteBuffer buffer ) { > // Decode > > return decoded; > } > > > void messageReceived( IoSession session, ByteBuffer buffer ) { > while ( ( Object decoded = decode( session ) ) != null ) { > // Do something > } > } > > > > > > > > Julien > > > > > -- > Regards, > Cordialement, > Emmanuel Lécharny > www.iktek.com > >
