Re: backporting HBASE-3777 to 0.90

[Ted Yu <yu...@gmail.com>]

Bright will upload patch v4 which would disable
TestHCM.testManyNewConnectionsDoesnotOOME (disabled in TRUNK)

Otherwise patch v3 is ready for wider validation effort.

Thanks in advance for people who put this backport in their clusters.

On Wed, Sep 28, 2011 at 3:45 PM, Gary Helmling <gh...@gmail.com> wrote:

> On Wed, Sep 28, 2011 at 1:39 PM, Andrew Purtell <ap...@apache.org>
> wrote:
>
> > > I'd switch from -1 to +1 if we can get +1s from people who have tried
> > > it on clusters with several different real existing apps written by
> > > several different teams.
> >
> >
> > This makes sense. My +1 was partly an agreement that I'd try it.
> >
> >
> I think I can agree to this as well.  Despite my previous messages, my
> bigger concern is subtler side effects from scope of the change, which has
> had longer to bake in 0.92/trunk.  If the patch can be verified in enough
> real workloads, then I can support it with suitable messaging in the
> release.
>
> Ultimately I think it would be good to more directly expose a cluster
> "connection" as an entry point to the client APIs.  Then client code could
> share connections or not, however it chose.  The current dependence on
> Configuration and hiding under HTable leaves us in an odd gray area where
> it's not clear if the current behavior is expected/needs to be supported,
> or
> just an implementation detail.  But that's a whole separate discussion that
> may or may not be worth it. :)
>

Re: backporting HBASE-3777 to 0.90

[Ted Yu <yu...@gmail.com>]

Unit test suite passed for patch v4.

Cheers

On Tue, Oct 4, 2011 at 8:44 PM, Ted Yu <yu...@gmail.com> wrote:

> Bright will upload patch v4 which would disable
> TestHCM.testManyNewConnectionsDoesnotOOME (disabled in TRUNK)
>
> Otherwise patch v3 is ready for wider validation effort.
>
> Thanks in advance for people who put this backport in their clusters.
>
>
> On Wed, Sep 28, 2011 at 3:45 PM, Gary Helmling <gh...@gmail.com>wrote:
>
>> On Wed, Sep 28, 2011 at 1:39 PM, Andrew Purtell <ap...@apache.org>
>> wrote:
>>
>> > > I'd switch from -1 to +1 if we can get +1s from people who have tried
>> > > it on clusters with several different real existing apps written by
>> > > several different teams.
>> >
>> >
>> > This makes sense. My +1 was partly an agreement that I'd try it.
>> >
>> >
>> I think I can agree to this as well.  Despite my previous messages, my
>> bigger concern is subtler side effects from scope of the change, which has
>> had longer to bake in 0.92/trunk.  If the patch can be verified in enough
>> real workloads, then I can support it with suitable messaging in the
>> release.
>>
>> Ultimately I think it would be good to more directly expose a cluster
>> "connection" as an entry point to the client APIs.  Then client code could
>> share connections or not, however it chose.  The current dependence on
>> Configuration and hiding under HTable leaves us in an odd gray area where
>> it's not clear if the current behavior is expected/needs to be supported,
>> or
>> just an implementation detail.  But that's a whole separate discussion
>> that
>> may or may not be worth it. :)
>>
>
>