For completeness here, and to wrap up this thread, the issue was identified as caused by THRIFT-1805, which regressed (a second time) an issue which suppressed TApplicationExceptions and simply caused the connection to be closed, resulting in what looked like a network issue (TTransportException) in the client.
ACCUMULO-2950 was opened to deal with this, and the patch is being tested, and is expected to restore the original behavior in 0.9.0. It does fix the issue identified in ACCUMULO-2935, but I made it a separate issue, so it was clearly documenting the underlying problem, separately from the symptom. Further, I reviewed the entire changeset from 0.9.0 to 0.9.1 in libthrift.jar and I didn't see any other changes that should cause any trouble for us. -- Christopher L Tubbs II http://gravatar.com/ctubbsii On Sun, Jun 22, 2014 at 5:41 PM, Josh Elser <[email protected]> wrote: > Since there was a discussion thread on this, I wanted to bring this up on > the list instead of just leaving a comment on the issue. > > I just bisect'ed a failing MetaSplitIT[1] and found that the changes > introduced by bumping to 0.9.1 were what supposedly introduced them. The > thing that worries me is that putting this in 1.6.1 came with a "it's > compatible" guarantee, when, to be perfectly honest, it's obviously not at > the level of compatibility that we want it to be for a bug-fix release (it > broke a test). > > Now, I haven't traced it back far enough to see what has exactly changed > with the exceptions (causing us to poll indefinitely instead of fail back > to the client), but that makes me worry that assumptions we have in the > implementation of our API, WRT exception handling, are suddenly invalid. > > I'll try to poke around at this some more to figure it out exactly. > > [1] https://issues.apache.org/jira/browse/ACCUMULO-2935 > > > On 5/23/14, 3:09 PM, Christopher wrote: > >> Given this conversation, and because I can't really think of a good >> reason not to, I'm going to proceed with applying this to 1.6.1, after >> addressing the issues in RB. That satisfies my desire to make a 1.7.0 >> minor release down the road, and satisfies busbey's concerns about >> mixing versioning semantics in 1.x.x (by avoiding 1.x.0 minor >> releases) >> >> -- >> Christopher L Tubbs II >> http://gravatar.com/ctubbsii >> >> >> On Fri, May 16, 2014 at 4:36 PM, Christopher <[email protected]> wrote: >> >>> Yes. They are 100% forward/backwards compatible on the wire. >>> >>> However, it is my understanding that there were some minor additions >>> (new API) to 0.9.1 which won't work in 0.9.0... but that won't affect >>> us since we are not using those features (and wouldn't be adding >>> anything that leverages those features in any bugfixes on 1.6.x), and >>> because we provide the jar. >>> >>> -- >>> Christopher L Tubbs II >>> http://gravatar.com/ctubbsii >>> >>> >>> On Fri, May 16, 2014 at 2:26 PM, Sean Busbey <[email protected]> >>> wrote: >>> >>>> Do we know if thirft 0.9.0 and 0.9.1 are forward compatible? >>>> >>>> >>>> On Fri, May 16, 2014 at 12:59 PM, Christopher <[email protected]> >>>> wrote: >>>> >>>> Correction: the current patch does *NOT* bump the wire version... I >>>>> thought I did that, but I did not. >>>>> >>>>> -- >>>>> Christopher L Tubbs II >>>>> http://gravatar.com/ctubbsii >>>>> >>>>> >>>>> On Fri, May 16, 2014 at 12:28 PM, Christopher <[email protected]> >>>>> wrote: >>>>> >>>>>> Devs, >>>>>> >>>>>> I'm considering whether or not it'd be appropriate to push in >>>>>> ACCUMULO-1691 into the 1.6.1-SNAPSHOT branch. >>>>>> This would effectively bump our dependency on libthrift to 0.9.1. >>>>>> However, thrift 0.9.1 and 0.9.0 are 100% wire-compatible (I've been >>>>>> assured by jfarrell and codesf in the #thrift IRC channel). >>>>>> >>>>>> Given that this we provide this dependency, and the bump would fix >>>>>> some thrift bugs, and that Thrift's own API is backwards-compatible in >>>>>> this version, I don't think this would impact our community except in >>>>>> the positive. >>>>>> >>>>>> (Note: currently, my patch for ACCUMULO-1691 bumps up the wire >>>>>> version, but I plan on changing that so it doesn't, now that I've been >>>>>> assured it is compatible... I've also done some manual tests to verify >>>>>> this, and haven't seen any issues across our tests, even without >>>>>> re-generating the thrift classes in ACCUMULO-2773; If I roll >>>>>> ACCUMULO-1691 in, I'd also include ACCUMULO-2773.) >>>>>> >>>>>> -- >>>>>> Christopher L Tubbs II >>>>>> http://gravatar.com/ctubbsii >>>>>> >>>>> >>>>> >>>> >>>> >>>> -- >>>> Sean >>>> >>>
