Joanthan:
Your patch for 0.90 was submitted when I wrote my first email.
Actually Bright's patch for HBASE-4508 fixes the bug you mentioned by adding
the following:
HBASE_INSTANCES.clear();
I think we should fix this bug instead of circumventing it.
Cheers
On Wed, Oct 19, 2011 at 8:57 PM, Jonathan Hsieh <[email protected]> wrote:
> Ted,
>
> I've posted a version of the 0.90 .META. rebuild that doesn't require
> HBASE-4508/HBASE-3777. The failure I encountered had to do with a cached
> HConnection and a what I think is a bug in deleteAllConnections in the 0.90
> branch. Specifically in 0.90 HConnectionManager.deleteConnection closes a
> specific connection and removes it self from HBASE_INSTANCES, a singleton
> map of instances. The aggregate version,
> HConnectionManager.deleteAllConnections, closes all the connections but does
> *not* remove the HConnections from HBASE_INSTANCES.
>
> When the running a full shutdown and restart of a minicluster, we have the
> same configuration and get a stale closed connection and get failures from
> that.
>
> Jon.
>
>
> On Wed, Oct 19, 2011 at 8:51 PM, Ted Yu <[email protected]> wrote:
>
>> Hi,
>> I have reviewed latest patch for HBASE-4508 and run test suite.
>>
>> Jonathan has made good progress in his offline .META. builder whose 0.90
>> patch is blocked by the integration of HBASE-4508.
>>
>> rocketfuel has been running 0.90 with HBASE-4508 for 5 months. I feel we
>> should integrate HBASE-4508 into 0.90.5 soon.
>>
>> Please provide your comments.
>>
>> Thanks
>>
>
>
>
> --
> // Jonathan Hsieh (shay)
> // Software Engineer, Cloudera
> // [email protected]
>
>
>
