Anoop Sam John commented on HBASE-20078:

getXXX() calls at the boundary of a BB is rare case any way (As per how the 
design of BC BBs and block sizes).  So when getXXX() is called and all the 
bytes of this primitive is not in single BB, multiple calls to MBB#get() is ok. 
That is having perf implication of doing the index calc etc. But as rare case, 
lets make the impl perfect than worrying for very rare case perf.  WDYT 

> MultiByteBuff : bug in reading primitives when individual buffers are too 
> small
> -------------------------------------------------------------------------------
>                 Key: HBASE-20078
>                 URL: https://issues.apache.org/jira/browse/HBASE-20078
>             Project: HBase
>          Issue Type: Bug
>    Affects Versions: 2.0.0-alpha-1
>            Reporter: stack
>            Assignee: stack
>            Priority: Major
>             Fix For: 2.0.0, 3.0.0
>         Attachments: HBASE-20078.patch
> Copied from the bottom of HBASE-13916, the JIRA that introduced MultiByteBuff:
> [~anoop.hbase] This bit of code needs more test. It does not seem to be doing 
> the right thing....
> Doing something like the below, it does not seem to be returning the right 
> answers....
> {code}
>   @Ignore // This test fails.
>   @Test
>   public void testGetInt() {
>     ByteBuffer bb1 = ByteBuffer.allocate(1);
>     bb1.put((byte)1);
>     ByteBuffer bb2 = ByteBuffer.allocate(1);
>     bb2.put((byte)0);
>     ByteBuffer bb3 = ByteBuffer.allocate(1);
>     bb3.put((byte)0);
>     ByteBuffer bb4 = ByteBuffer.allocate(1);
>     bb4.put((byte)1);
>     MultiByteBuff mbb = new MultiByteBuff(bb1, bb2, bb3, bb4);
>     // Value is wrong here ... needs adjusting but code is doing wrong thing.
>     assertEquals(256, mbb.getIntAfterPosition(0));
>   }
> {code}
> Ignore the expected answer in the above... just a place holder as i messed 
> with the return... but if I make buffers of one byte, we do not seem to be 
> moving into the next buffer properly... If I make the buffers of two bytes, 
> similar. This stuff normally works because we are not spanning buffer 
> boundaries... but if we do, response seems like it could be off sir. I'll 
> have a go at this. You might have input meantime.
> The issue is the individual BBs are too small. Ideally that would never 
> happen. Still better to solve this corner cases.

This message was sent by Atlassian JIRA

Reply via email to