[
https://issues.apache.org/jira/browse/THRIFT-1035?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12980359#action_12980359
]
Mathias Herberts commented on THRIFT-1035:
------------------------------------------
I see this issue as more than a backwards compatibility one. To me it's a
matter of making Thrift usable by a vast majority of programmers.
As we've seen in THRIFT-830 and THRIFT-882, using ByteBuffer can be tricky even
for expert programmers, to boot, ByteBuffer are not Thread Safe and thus would
impose extraneous precautions on their use.
For example, in the Cassandra API, SlicePredicate has a list<binary> field,
this will lead to List<ByteBuffer> objects being generated and thus their use
would impose extra synchronization.
I agree that my fix introduces an inconsistency internally, but on the user
facing side we've worked after THRIFT-830 so binary fields could be
set/retrieved using byte[], right now collections of binary fields cannot use
byte[], so we suffer from a lack of external consistency which, in my point of
view, is more important to have than internal consistency.
On the performance side, could you pinpoint cases where using byte[] in
collections would be less efficient that the ByteBuffers, since the retrieved
arrays will be those backing the ByteBuffers when deserializing and the
ByteBuffer passed to the serialization code are simply wrapping the byte[] in
the collections.
I'd like to have other people express their view on this issue, I think it
would be really sad to go with your second option.
> Container types containing binary data are parameterized with ByteBuffer in
> the generated Java code
> ---------------------------------------------------------------------------------------------------
>
> Key: THRIFT-1035
> URL: https://issues.apache.org/jira/browse/THRIFT-1035
> Project: Thrift
> Issue Type: Bug
> Components: Java - Compiler, Java - Library
> Affects Versions: 0.4, 0.5, 0.6, 0.7
> Environment: All
> Reporter: Mathias Herberts
> Attachments: THRIFT-1035-2.patch, THRIFT-1035.patch
>
>
> Since THRIFT-830, binary fields are internally handled using ByteBuffer.
> Release 0.4.0 was the first to expose the ByteBuffer to the outside world
> (replacing previous methods returning/accepting byte[]).
> THRIFT-882 lead to the methods accepting/returning byte[] being available
> again, as it was deemed more reasonable not to expose the ByteBuffer too much
> as their use could be cumbersome. This lead to 0.5.0 being backward
> compatible with 0.3.0 on the binary fields front.
> During that time, nobody noticed that container types that contained binary
> data had their generated Java code changed to collections parameterized with
> ByteBuffer instead of byte[].
> list<binary> -> List<ByteBuffer>
> set<binary> -> Set<ByteBuffer>
> map<binary,...> -> Map<ByteBuffer,...>
> map<...,binary> -> Map<...,ByteBuffer>
> This introduces confusion in the API and still exposes ByteBuffer when
> discussion on THRIFT-882 concluded this should be avoided.
> We need to provide a way to offer the original parameterization with byte[]
> as this will simplify working with that type of collection and thus will
> increase the odds of Thrift's adoption.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
