There was a discussion over the weekend on the incubator dist-list about
Accumulo and describing what was borrowed from Hadoop-core and Hbase.

http://mail-archives.apache.org/mod_mbox/incubator-general/201109.mbox/%3C8
96611333.83615.1315154356465.javamail.r...@linzimmb04o.imo.intelink.gov%3E


5400 lines: slightly modified versions of Hadoop BCFile and related classes
            (our current file format extends BCFile)
4300 lines: heavily modified versions of MapFile and SequenceFile
            (no longer our default file format, but still included for
backward compatibility)
2000 lines: heavily modified versions of HBase BlockCache and related files
            (Adam didn't count the tests when he said 1500 lines)
1300 lines: heavily modified versions of Hadoop BloomFilters
419 lines: modified Hadoop TeraSortIngest to sort data using Accumulo
325 lines: our Value is an immutable version of Hadoop BytesWritable
142 lines: modified ClassLoader based on commons-jci ReloadingClassLoader







On 9/5/11 5:35 PM, "Joey Echeverria" <[email protected]> wrote:

>The Accumulo implementation of the WAL is a separate set of daemons.
>When you write to the WAL, you send your transactions to three of the
>logging servers. When you do a recovery, I believe one of the three
>servers that has the WAL for the down server copies it to HDFS and
>then a MapReduce job splits the log and re-inserts the recovered data.
>You should have the same survivability that you get with HDFS.
>
>-Joey
>
>On Mon, Sep 5, 2011 at 5:06 PM, Bill <[email protected]> wrote:
>> On 04/09/11 07:43, Mathias Herberts wrote:
>>>
>>> On Sep 4, 2011 1:39 AM, "Bill de hÓra"<[email protected]>  wrote:
>>>>
>>>> On 02/09/11 19:06, Stack wrote:
>>>>>
>>>>> What do folks think?
>>>>
>>>>
>>>> Not putting the log into hdfs seems like a good idea.
>>>
>>> I was somehow thinking the opposite as it makes irrecoverable machine
>>> failures much more problematic. What makes you say it's a good idea?
>>>
>>
>> Allows more control over the write path, specifically sequential I/O and
>> crash recovery. Granted the commit needs to be replicated, but you need
>>that
>> regardless. Thinking a bit more it might not square with the
>>regionserver
>> model anyway, plus the Accumulo proposal mentions a service rather than
>>a
>> local disk. The WAL seems to be hardened up these days anyway making
>>things
>> like https://issues.apache.org/jira/browse/HBASE-4107 more of an edge
>>case..
>>
>> Bill
>>
>
>
>
>-- 
>Joseph Echeverria
>Cloudera, Inc.
>443.305.9434

Reply via email to