accumulo-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Kristopher Kane <kkane.l...@gmail.com>
Subject Re: Write ahead log growth and reduction
Date Thu, 12 Apr 2012 14:21:49 GMT
Eric, thanks for the help.

This is on 1.3.2, looks like we will be able to go to 1.3.4 very shortly.
>From the below JIRA, it looks like it is fixed in 1.3.6.  I will send this
up as I'm not an approving authority on installed/running versions and will
also recommend the trace password in the site.xml.

Once we upgrade to the minor version, I will report back on whether the WAL
GC runs in that version.

-Kris

On Wed, Apr 11, 2012 at 1:00 PM, Eric Newton <eric.newton@gmail.com> wrote:

> Also you will need to restart.
>
> What version of accumulo are you running?  This problem should have been
> fixed with ACCUMULO-243.
>
> -Eric
>
> On Wed, Apr 11, 2012 at 12:28 PM, Eric Newton <eric.newton@gmail.com>wrote:
>
>> in conf/accumulo-site.xml, add this:
>>
>>     <property>
>>       <name>trace.password</name>
>>       <value>your-root-password-here</value>
>>     </property>
>>
>> -Eric
>>
>> On Wed, Apr 11, 2012 at 12:14 PM, Kristopher Kane <kkane.list@gmail.com>wrote:
>>
>>> I know I haven't answered some of the previous questions to me, but, the
>>> GC talk brought me to the source of the problem.  our WAL GC job is not
>>> successfully running.  The GC logs state a bad username/password, so, will
>>> investigate that.
>>>
>>> Thanks for the help!
>>>
>>> -Kris
>>>
>>> On Wed, Apr 11, 2012 at 11:18 AM, Keith Turner <keith@deenlo.com> wrote:
>>>
>>>> When data is written to the accumulo its written to memory and the
>>>> write ahead logs.  The data in memory is sorted and the data in the
>>>> write ahead logs is written asis (unsorted).  When the data in memory
>>>> is flushed to HDFS, the write ahead logs that also contain that data
>>>> are no longer needed.
>>>>
>>>> On Wed, Apr 11, 2012 at 11:12 AM, Eric Newton <eric.newton@gmail.com>
>>>> wrote:
>>>> > Logs are only put into HDFS during a recovery.
>>>> >
>>>> > Flush removes references to WALs, and the accumulo gc will ask the
>>>> loggers
>>>> > to delete them when there are no references to them.
>>>> >
>>>> > -Eric
>>>> >
>>>> >
>>>> > On Wed, Apr 11, 2012 at 11:02 AM, Kristopher Kane <
>>>> kkane.list@gmail.com>
>>>> > wrote:
>>>> >>
>>>> >>
>>>> >>
>>>> >> On Wed, Apr 11, 2012 at 10:48 AM, Keith Turner <keith@deenlo.com>
>>>> wrote:
>>>> >>>
>>>> >>> How big is the partition?  Are the same number of logger servers
>>>> >>> running as tablet servers?
>>>> >>>
>>>> >>>
>>>> >>> You can scan the metadata table to look for tablets that have
alot
>>>> of
>>>> >>> write-ahead logs. I think the command below will show you how
many
>>>> >>> write-ahead logs each tablet has.  Look for any tablets that
have
>>>> too
>>>> >>> many. I think it should sort the tablets with the most tablets
to
>>>> the
>>>> >>> top, but not positive.
>>>> >>>
>>>> >>>   ./bin/accumulo shell -u root -p secret -e 'scan -t !METADATA
-c
>>>> >>> log' | cut -f 1 -d ' ' | uniq -c  | sort -r -n
>>>> >>>
>>>> >>> I think the following command will show you how many active
log each
>>>> >>> logger has.  This should be even.
>>>> >>>
>>>> >>>   ./bin/accumulo shell -u root -p secret -e 'scan -t !METADATA
-c
>>>> >>> log' | cut -f 2 -d ' ' | cut -d ':' -f 2 | sort | uniq -c
>>>> >>>
>>>> >>> You can use the "flush -p" command in the shell to force data
in
>>>> >>> memory to disk and stop referencing write-ahead logs.  Maybe
execute
>>>> >>> the commands above before and after flushing.
>>>> >>>
>>>> >>> Keith
>>>> >>>
>>>> >>>
>>>> >>
>>>> >>
>>>> >>
>>>> >> Thanks for the replies.  I read about the flush command in the docs
>>>> but
>>>> >> didn't make a connection between "memory" to the write ahead logs.
>>>> Is
>>>> >> that correct?  Flush writes write ahead log data to hdfs?
>>>> >>
>>>> >> Thanks!
>>>> >>
>>>> >> -Kris
>>>> >
>>>> >
>>>>
>>>
>>>
>>
>

Mime
View raw message