hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Edward Capriolo <edlinuxg...@gmail.com>
Subject Re: [Discuss] project chop up
Date Fri, 16 Aug 2013 15:58:38 GMT
Summary from hive-irc channel. Minor edits for spell check/grammar.

The last 10 lines are a summary of the key points.

[10:59:17] <ecapriolo> noland: et all. Do you want to talk about hive in
maven?
[11:01:06] smonchi [~
ronin@host34-189-dynamic.23-79-r.retail.telecomitalia.it] has quit IRC:
Quit: ... 'cause there is no patch for human stupidity ...
[11:10:04] <noland> ecapriolo: yeah that sounds good to me!
[11:10:22] <noland> I saw you created the jira but haven't had time to look
[11:10:32] <ecapriolo> So I found a few things
[11:10:49] <ecapriolo> In common there is one or two testats that actually
fork a process :)
[11:10:56] <ecapriolo> and use build.test.resources
[11:11:12] <ecapriolo> Some serde, uses some methods from ql in testing
[11:11:27] <ecapriolo> and shims really needs a separate hadoop test shim
[11:11:32] <ecapriolo> But that is all simple stuff
[11:11:47] <ecapriolo> The biggest problem is I do not know how to solve
shims with maven
[11:11:50] <ecapriolo> do you have any ideas
[11:11:52] <ecapriolo> ?
[11:13:00] <noland> That one is going to be a challenge. It might be that
in that section we have to drop down to ant
[11:14:44] <noland> Is it a requirement that we build both the .20 and .23
shims for a "package" as we do today?
[11:16:46] <ecapriolo> I was thinking we can do it like a JDBC driver
[11:16:59] <ecapriolo> Se separate out the interface of shims
[11:17:22] <ecapriolo> And then at runtime we drop in a driver implementing
[11:17:34] Wertax [~wertax@wolfkamp.xs4all.nl] has quit IRC: Remote host
closed the connection
[11:17:36] <ecapriolo> That or we could use maven's profile system
[11:18:09] <ecapriolo> It seems that everything else can actually link
against hadoop-0.20.2 as a provided dependency
[11:18:37] <noland> Yeah either would work. The driver method would
probably require use to use ant build both the drivers?
[11:18:44] <noland> I am a fan of mvn profiles
[11:19:05] <ecapriolo> I was thinking we kinda separate the shim out into
its own project,, not a module
[11:19:10] <ecapriolo> to achive that jdbc thing
[11:19:27] <ecapriolo> But I do not have a solution yet, I was looking to
farm that out to someone smart...like you :)
[11:19:33] <noland> :)
[11:19:47] <ecapriolo> All I know is that we need a test shim because
HadoopShim requires hadoop-test jars
[11:20:10] <ecapriolo> then the Mini stuff is only used in qtest anyway
[11:20:48] <ecapriolo> Is this something you want to help with? I was
thinking of spinning up a github
[11:20:50] <noland> I think that the separate projects would work and
perhaps nicely.
[11:21:01] <noland> Yeah I'd be interested in helping!
[11:21:17] <noland> But I am going on vacation starting next week for about
10 days
[11:21:27] <ecapriolo> Ah cool where are you going?
[11:21:37] <noland> Netherlands
[11:21:42] <noland> Biking around and such
[11:23:52] <noland> The one thing I was thinking about with regards to a
branch is keeping history. We'll want to keep history for the files but
AFAICT svn doesn't understand git mv.
[11:24:16] Wertax [~wertax@wolfkamp.xs4all.nl] has joined #hive
[11:31:19] jeromatron [~textual@host90-152-1-162.ipv4.regusnet.com] has
quit IRC: Quit: My MacBook Pro has gone to sleep. ZZZzzz…
[11:35:49] <ecapriolo> noland: Right I do not play to suggest that we will
do this in git
[11:36:11] <ecapriolo> I just see that we are going to have to hack stuff
up and it is not the type of work that lends itself well to branches.
[11:36:17] <noland> Ahh ok
[11:36:56] <ecapriolo> Once we come up with a solution for the shims, and
we have something that can reasonably build and test hive we can figure out
how to apply that to a branch/trunk
[11:36:58] <noland> yeah so just do a POC on github and then implement on
svn
[11:37:05] <noland> cool
[11:37:29] <ecapriolo> Along the way we can probably find things that we
can do like that common test I found and other minor things
[11:37:41] <noland> sounds good
[11:37:50] <ecapriolo> Those we can likely just commit into the current
trunk and I will file issues for those now
[11:37:58] <noland> cool
[11:38:41] <ecapriolo> But yea man. I just cant take the project as it is
now
[11:38:51] <ecapriolo> in eclipse everytime I touch a file it rebuilds
everything!
[11:38:53] <ecapriolo> Its like WTF
[11:39:09] <ecapriolo> Running one tests takes like 3 minutes
[11:39:12] <ecapriolo> its out of control
[11:39:23] <noland> LOL
[11:39:29] <noland> I agree 110%
[11:39:32] <ecapriolo> eclipse was not always like that I am not sure how
the hell it happened
[11:39:51] <noland> The eclipse sep thing is so harmful
[11:40:08] <noland> dep thing that is
[11:40:12] <ecapriolo> I mean command line ant was always bad, but you used
to be able to work in eclipse without having to rebuild everything every
change/test
[11:40:39] <noland> Yeah the first thing I do these days is disable the ant
builder
[11:40:52] <ecapriolo> Ow... I did not really know that was a thing
[11:40:55] <noland> it starts compiling while you are still working and
blocks for minutes
[11:41:02] <ecapriolo> Right that is what I mean
[11:41:11] <ecapriolo> Everyone has like 10 hacks to work on the project
[11:41:14] <noland> yeah you can remove it in project…one sec
[11:41:17] <ecapriolo> perm gen
[11:41:20] <ecapriolo> ant builder
[11:41:32] <noland> project -> properties -> builders
[11:41:34] <ecapriolo> hive does not build offline anymore
[11:41:37] <noland> yeah
[11:41:47] <ecapriolo> Im not sure when this stuff went bad, but it has
gotten really really bad
[11:42:09] <ecapriolo> Also what I plan on doing is stripping out
non-essentials
[11:42:25] <ecapriolo> like serde has all this thrift and avro stuff to
support custom formats
[11:42:30] <ecapriolo> that is going into its own module
[11:42:43] <ecapriolo> Going to rip out all the udfs accept between and or.
[11:43:50] <noland> yeah it'd be nice to have those items in their own
modules so you can just build/test them when you want
[11:44:12] <ecapriolo> hbase zookeeper locking
[11:44:31] Wertax [~wertax@wolfkamp.xs4all.nl] has quit IRC: Remote host
closed the connection
[11:44:44] <noland> yeah for sure
[11:45:04] <noland> I think the default for testing should be the in
process locking
[11:45:10] <ecapriolo> Absolutely.
[11:45:40] <ecapriolo> The other issue I want to tackle is hive-exec.jar
[11:45:54] <ecapriolo> I want to jar-jar all the dependencies.
[11:46:46] <ecapriolo> I run into to many conflicts with log4j and guava,
and commons-utils all those things need to be packaged into non-conflicting
packages
[11:46:58] <noland> I haven't looked at how we build that yet but I agree
it'd be nice if we could jar-jar things like guava
[11:47:12] <noland> so we can actually use them on server side
[11:47:16] <ecapriolo> We dont really need quava. its probably just used
for one tiny thing
[11:47:43] <ecapriolo> People are forgetting/do not understand that
hive-exec needs to get sent via the distributed cache
[11:47:57] <noland> Wen we implement range joins they have a RangeMap that
we'll need.
[11:47:57] <ecapriolo> so making it hulkingly fat just slows everything down
[11:48:11] <noland> Do we ship it every time?
[11:48:25] <noland> Cause we only have to ship it once per version of the
jar.
[11:48:42] <ecapriolo> Recently you need the jackson jars on the auxlib as
well
[11:48:46] <ecapriolo> hive will not work without it
[11:49:11] <ecapriolo> People are just focused
feature-feature-feature...bigger...bigger bigger
[11:49:24] rubensayshi [drakie@nat/hyves.nl/x-uxywnflkbberbzhq] has quit
IRC: Quit: Leaving
[11:49:27] <noland> yeah maven modules will definitely help us understand
who depends on what.
[11:49:28] <ecapriolo> Next up kyro
[11:49:51] <noland> I agree there is a lot of tech debt that needs paying
[11:50:30] <ecapriolo> So those are all the high level things I want to
tackle
[11:50:59] <ecapriolo> shims, general cleanup, break out non-essential
code, build a better non conflicting hive-exec jar
[11:51:10] <noland> That sounds good. Once we hack on github for a while
it'd be nice to develop a brief high level plan on how to implement
[11:51:26] <ecapriolo> Also get maven artifacts with correct depencency
scopes like provided etc
[11:51:40] <ecapriolo> Right now pulling a hive jar from maven is like
pulling in the world
[11:52:08] bvanhoy [~Adium@64.124.34.34] has joined #hive


On Thu, Aug 15, 2013 at 11:14 PM, Edward Capriolo <edlinuxguru@gmail.com>wrote:

> I have opened https://issues.apache.org/jira/browse/HIVE-5107 because I
> am growing tired of how long hive's build take.
>
> I have started playing with this by creating a simple multi-module project
> and copying stuff as I go. I have ported a minimal shims and common and I
> have all the tests in common almost running.
>
> Q. This is going to be ugly hacky work for a while, I was thinking it
> should be a branch but it is just going to be a mess of moves and copies
> etc. Not really something you can diff etc.
>
> Is anyone else interested in working on this as well. If so I think we can
> just setup a github and I can arrange for anyone to have access to it.
>
> Thanks,
> Edward
>
>
> On Wed, Aug 7, 2013 at 5:04 PM, Edward Capriolo <edlinuxguru@gmail.com>wrote:
>
>> "Some of the hard part was that some of the test classes are in the wrong
>> module that references classes in a later module."
>>
>> I think the modules will have to be able to reference each other in many
>> cases. Serde and QL are tightly coupled. QL is really too large and we
>> should find a way to cut that up.
>>
>> Part of this problem is the q.tests
>>
>> I think one way to handle this is to only allow unit tests inside the
>> module. I imagine running all the q tests would be done in a final module
>> hive-qtest. Or possibly two final modules
>> hive-qtest
>> hive-qtest-extra (tangential things like UDFS and input formats not core
>> to hive)
>>
>>
>> On Wed, Aug 7, 2013 at 4:49 PM, Owen O'Malley <omalley@apache.org> wrote:
>>
>>> On Wed, Aug 7, 2013 at 12:55 PM, kulkarni.swarnim@gmail.com <
>>> kulkarni.swarnim@gmail.com> wrote:
>>>
>>> > > I'd like to propose we move towards Maven.
>>> >
>>> > Big +1 on this. Most of the major apache projects(hadoop, hbase, avro
>>> etc.)
>>> > are maven based.
>>> >
>>>
>>> A big +1 from me too. I actually took a pass at it a couple of months
>>> ago.
>>> Some of the hard part was that some of the test classes are in the wrong
>>> module that references classes in a later module. Obviously that prevents
>>> any kind of modular build.
>>>
>>> As an additional plus to Maven is that Maven includes tools to correct
>>> the
>>> project and module dependencies.
>>>
>>> -- Owen
>>>
>>
>>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message