On Sat, Nov 24, 2012 at 1:42 PM, Jos Backus <[email protected]> wrote:
> On Nov 23, 2012 3:21 AM, "Steve Loughran" <[email protected]> wrote:
>>
>> I've opened a JIRA on the idea of having the hadoop processes optionally
>> create exclusive write-locked files on startup, failing if the lock
>> couldn't be acquired
>>
>> https://issues.apache.org/jira/browse/HADOOP-9086
>>
>> I have to plans to implement this right now -but we all know that the
>> current pid file approach is flawed.
>>
>> -steve
>
> Ugh. Why not use a real process supervisor like daemontools or Upstart? Why
> duplicate this type of code in every application, poorly? The process table
> already tracks this information, without any races or extra files needed.
Here's what I'd like to see happen for Bigtop 0.6.0 (scheduled for the end
of Q1 2013):
#1 we should totally migrate to upstart/systemd for all the platforms that
allow that
#2 we should figure out whether we still care about things like CentOS5
and if we do we should do enough systemd/daemontools backporting
to make it work.
I expect to be very much involved in #1. In fact, looking at the test results
for Quetazal/OpenSUSE12.2 we actually might not have much choice but
to implement this for Bigtop 0.5.0 (more on this next week).
Thoughts?
Thanks,
Roman.
