The binary compat is a slippery slope. It'd be a bummer if we couldn't
take advantage of all the innovation you guys are doing. At the same
time, it's tough to require the Phoenix user community, for example, to
upgrade their HBase servers to be able to move to the latest version of
Phoenix. I don't know what the right answer is, but here are a couple of
concrete cases:
- between 0.94.3 and 0.94.4, new methods were introduced on
RegionScanner. Often coprocessors will have their own implementation of
these so that they can aggregate in the postScannerOpen. Though this
broke binary compat, it improved scan performance as well. Where does
the binary compatible line stop?
- between 0.94.3 and 0.94.4, the class loading changed for coprocessors.
If a coprocessor was on the classpath, it didn't matter what the jar
path was, it would load. In 0.94.4, that was no longer the case - if the
jar path was invalid, the coprocessor would no longer load. Though this
broke compatibility, it was a good cleanup for the class loading logic.
Call it a bug fix or a change in behavior, but either way it was an
incompatible change. Is a change in behavior that causes
incompatibilities still meet the binary compatible criteria?
- between 0.94.4 and 0.94.5, the essential column family feature was
introduced. This is an example of one that is binary compatible. We're
able to take advantage of the feature and maintain binary compatibility
with 0.94.4 (in which case the feature simple wouldn't be available).
Maybe if we just explicitly identified compatibility issues, that would
be a good start? We'd likely need a way to find them, though.
James
On 04/04/2013 03:59 PM, lars hofhansl wrote:
I agree we need both, but I'm afraid that ship has sailed.
It's not something we paid a lot of attention to especially being
forward-binary-compatible. I would guess that there will be many more of these
issues.
Also, we have to qualify this statement somewhere. If you extend HRegionServer
you cannot expect compatibility between releases. Of course that is silly, but
it serves the point I am making.
For client visible classes (such as in this case) we should make it work, we
identifies issues with Filters and Coprocessors in the past and kept them
binary compatible on a best effort basis.
TL;DR: Let's fix this issue, and be wary of more such issues.
-- Lars
________________________________
From: Andrew Purtell <[email protected]>
To: "[email protected]" <[email protected]>
Sent: Thursday, April 4, 2013 3:21 PM
Subject: Re: Does compatibility between versions also mean binary compatibility?
"Compatible" implies both to my understanding of the term, unless
qualified.
I don't think we should qualify it. This looks like a regression to me.
On Thu, Apr 4, 2013 at 1:20 PM, Jean-Daniel Cryans <[email protected]>wrote:
tl;dr should two compatible versions be considered both wire and
binary compatible or just the former?
Hey devs,
0.92 is compatible with 0.94, meaning that you can run a client for
either against the other and you can roll restart from 0.92 to 0.94.
What about binary compatibility? Meaning, can you run user code
compiled against 0.92 with 0.94's jars?
Unfortunately, the answer is "no" in this case if you invoke setters
on HColumnDescriptor as you'll get:
java.lang.NoSuchMethodError:
org.apache.hadoop.hbase.HColumnDescriptor.setMaxVersions(I)V
HBASE-5357 "Use builder pattern in HColumnDescriptor" changed the
method signatures by changing "void" to "HColumnDescriptor" so it' not
the same methods anymore.
I don't think we really had talks about binary compatibility before so
this is why I'm raising it up now.
Should "compatible" versions be just wire compatible or both wire and
binary compatible? The latter means we need new tests. I think it
should be both.
What do you guys think?
J-D
