ARROW-6315: [Java] Make change to ensure flatbuffer reads are aligned

[tianchen92]

Implements the IPC message format alignment changes for ARROW-6315.
i. MessageReader can read messages with the old alignment
ii. ArrowWriter could choose produces messages with the new alignment or the old format.

[emkornfield]

@tianchen92 it looks like this breaks the ArrowTools test

[tianchen92]

@tianchen92 it looks like this breaks the ArrowTools test

It's really strange, I believe it works fine with maven build on my local machine, is there any difference with travis?

[emkornfield]

That is interesting, your change passes on my box as well. Let me try rerunning CI.

[emkornfield]

@ursabot build

[BryanCutler]

Thanks for the pr @tianchen92 . I'll have to look at it more closely again, but I think when you deserialize you check for the continuation that is not there because it has already been read.

[tianchen92]

Thanks for the pr @tianchen92 . I'll have to look at it more closely again, but I think when you deserialize you check for the continuation that is not there because it has already been read.

@BryanCutler Thanks a lot for your feedback, I left some comments above and updated this PR.
Besides, this PR always test fail in travis but works fine in local machine, the error log seems useless, do you have any idea how to find the cause? @emkornfield

org.apache.arrow.tools.EchoServerTest
[ERROR] basicTest(org.apache.arrow.tools.EchoServerTest) Time elapsed: 0.023 s <<< ERROR!
java.net.SocketException: Connection reset
at org.apache.arrow.tools.EchoServerTest.testEchoServer(EchoServerTest.java:115)
at org.apache.arrow.tools.EchoServerTest.basicTest(EchoServerTest.java:147)

[BryanCutler]

Looks like I can reproduce the error locally, I'll look into it.

[tianchen92]

Looks like I can reproduce the error locally, I'll look into it.

Awesome, thanks a lot for your help!

[emkornfield]

@BryanCutler do you have a sense if as part of this change the old/new writing/reading code paths in Java need to be controllable by an environment variable for use with Spark? I think we need something like that for Python?

[BryanCutler]

I just found the root cause of the error and can post a patch soon.

@BryanCutler do you have a sense if as part of this change the old/new writing/reading code paths in Java need to be controllable by an environment variable for use with Spark? I think we need something like that for Python?

Maybe, I'm not entirely sure. The case that I think might fail is when a user has pyarrow 0.15.0 and using Spark <= 2.4.4 with Arrow Java 0.14.1. It should work going Java -> Python, but then going Python -> Java, it will write the 8-byte prefix by default and Java will choke. So an env var would help in that case I think.

[BryanCutler]

@tianchen92 #5288 is merged, please rebase when you can and hopefully this will be good to go

[BryanCutler]

Hmm, this doesn't seem quite right. The padded bytes are included in the messageLength, then that length is read using ReadChannel.readFully so reader.bytesRead() should include the padded bytes as well

[BryanCutler]

Oh wait, is this the additional 4 bytes sent at the end of stream? What happens if the socket isn't closed and then another stream or other data is sent over it? I think that extra 4 bytes will still need to be read and cause problems for the next stream.

@wesm @emkornfield , does C++ ignore the last 4-bytes at the EOS, since now it is an 8-byte value?

[tianchen92]

Good catch, I looked again and I think you are right, the additional 4 bytes should be read.
In this case, may be we should also pass option to reader and read the additional 4 bytes if necessary?

[eerhardt]

It looks like the C++ implementation only reads the first 4 bytes:

https://github.com/apache/arrow/blob/ARROW-6313-flatbuffer-alignment/cpp/src/arrow/ipc/message.cc#L327-L345

How else are we supposed to know if we should read 4 or 8 bytes? Maybe for the new format, an 8-byte EOS token should look like {0xFFFFFFFF, 0x00000000}, so we read the continutation token first, and then know to read the next 4 bytes, which are then 0 to signal EOS.

[BryanCutler]

Could the reader just remember the state, so if it reads the continuation
token from the beginning, then read all 8 bytes at the end?

[tianchen92]

@tianchen92 #5288 is merged, please rebase when you can and hopefully this will be good to go

Thanks Bryan, I have rebased and let's see if it could pass :)

Ah, I made a mistake, this branch is not based on master...

[BryanCutler]

Looks like the previous test failure is passing now. Integration tests have lots of failures with JS and Go, but Java <-> C++ look fine. Are there known issues with JS/Go integration, or maybe they don't have the protocol changes yet?

Also, I think from the discussion here #5229 (comment) this is going to be a problem. We could fix this as a followup, since it seems C++ is doing the same thing? What do you guys think?

[emkornfield]

JS hasn't been checked in. Go has been checked so we might need to investigate if that is failing.

[emkornfield]

It seems like all the errors on the integration test are due to JS.

[emkornfield]

I'm going to merge, lets follow-up on the terminal 8 bytes on the mailing list.

[BryanCutler]

@emkornfield did you merge this yet? The PR still looks open, but I see you made a commit

NVM, I missed that the changes were going into an integration branch.

[emkornfield]

Closing to avoid confusinon.