On Thu, Feb 23, 2012 at 11:02 AM, Santhosh Srinivasan <[email protected]> 
wrote:
> On average how long have unit tets failures been tolerated in Hadoop?

It depends. Some nondeterministic tests were tolerated for awhile, but
most are fixed relatively quickly (hours/days) because, when the CI
build + test patch script isn't trusted, broken code gets committed
over the "noise". Contrib modules can stay broken for awhile, if
nobody cares to fix them.

> For point 2, till CI is setup - I would venture to say that Linux is a must 
> for committers. Once CI is in place then this requirement can be relaxed as 
> the nightly will catch it within 24 hours.

There's no enforcement (or even detection) without CI/nightly builds,
so I'm not sure what we're discussing. The project relies on the
devs/users who use Oozie on Linux to file bugs. The obvious
inefficiency there should be sufficient motivation to set up
CI/nightly builds.

Besides, the project is (mostly) in Java; how often does this even come up? -C

> -----Original Message-----
> From: Chris Douglas [mailto:[email protected]]
> Sent: Wednesday, February 22, 2012 8:35 PM
> To: [email protected]
> Cc: Mohammad Islam; Alan Gates; Christopher Douglas
> Subject: Re: Feedback requested: Building in Linux/Mac/Doesn't matter
>
> On Wed, Feb 22, 2012 at 7:53 PM, Santhosh Srinivasan <[email protected]> 
> wrote:
>> AFAIK there are no Mac installations of the Hadoop, Pig, Oozie, etc. Given 
>> that the cost of catching a failing unit test during release time (given 
>> that we don't have CI in Apache) is fairly high. Most of you must have 
>> noticed unit test case failures on the recent release.
>
> CI resources are available, if interest drives someone to work on it.
> What makes a failing unit test expensive during release time?
>
>> Use of a Mac (or Windows ;) is a convenience that we can't forego. What are 
>> the best practices in the other Apache projects?
>
> In Hadoop, Windows+Cygwin was supported while developers who used it cared to 
> fix those bugs. When they didn't, some unit tests consistently failed on that 
> platform. Most projects either rely on CI or an auditing phase during 
> release. It's not the end of the world if corners of trunk, or any 
> development branch, are broken for awhile.
>
> YMMV, but nightly builds are usually sufficient to catch regressions.
> Does this need a policy? Surely no release would go out if it didn't work on 
> Linux. -C
>
>> Santhosh
>>
>> -----Original Message-----
>> From: [email protected] [mailto:[email protected]] On Behalf Of
>> Roman Shaposhnik
>> Sent: Wednesday, February 22, 2012 4:34 PM
>> To: [email protected]
>> Cc: Mohammad Islam
>> Subject: Re: Feedback requested: Building in Linux/Mac/Doesn't matter
>>
>> On Wed, Feb 22, 2012 at 4:22 PM, Alejandro Abdelnur <[email protected]> 
>> wrote:
>>> To run tests on a given platform? No, but if doing hadoop native
>>> stuff, only linux is supported at the moment.
>>
>> I must add to that: there is absolutely a hard requirement for test-patch 
>> not failing to build. And that build is a Linux one.
>>
>> Thanks,
>> Roman.

Reply via email to