Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Hadoop, mail # user - concurrency


Copy link to this message
-
Re: concurrency
Koert Kuipers 2012-10-12, 17:05
Hey Harsh & Joep,

My main worry was actually the simpler situation in which only new subdirs
are created by loaders. If we for a second focus on this "append-only"
situation, which i admit is only a subset of all cases, even then it is not
entirely clear to me how to go about this. Right now i pass into my query
job (the one that reads) a path like /data/*, which means i do not check if
every subdir has a _SUCCESS flag. If i wanted to add this check then i
would have to do that check myself for every subdir (with another glob
expression i suppose), and then construct a set of paths that go into the
job instead of my simple glob? That set could become very large (thousands
of paths). Is this what you are suggesting Harsh?

Now Joep rightly points out that reality is never that simple. We also have
jobs running like the ones that he described, which compact data in place,
and other fun stuff like that... For us a typical situations are
compactions, but we also do roll-ups (where we replace many subdirs for
hourly updates with a single subdir for say a month). Right now my approach
was to do this stuff once a week in a "reserved" windows in which no
queries are running, but I would love to hear about suggestions for this as
well.

Best (and Groeten Joep!)
Koert
On Fri, Oct 12, 2012 at 12:17 PM, Harsh J <[EMAIL PROTECTED]> wrote:

> Joep,
>
> You're right - I missed in my quick scan that he was actually
> replacing those files there. Sorry for the confusion Koert!
>
> On Fri, Oct 12, 2012 at 9:37 PM, J. Rottinghuis <[EMAIL PROTECTED]>
> wrote:
> > Hi Harsh, Moge Koert,
> >
> > If Koerts problem is similar to what I have been thinking about where we
> > want to consolidate and re-compress older datasets, then the _SUCCESS
> does
> > not really help. _SUCCESS helps to tell if a new dataset is completely
> > written.
> > However, what is needed here is to replace an existing dataset.
> >
> > Naive approach:
> > The new set can be generated in parallel. Old directory moved out of the
> > way (rm and therefore moved to Trash) and then he new directory renamed
> > into place.
> > I think the problem Koert is describing is how to not mess up map-reduce
> > jobs that have already started and may have read some, but not all of the
> > files in the directory. If you're lucky, then you'll try to read a file
> > that is no longer there, but if you're unlucky then you read a new file
> > with the same name and you will never know that you have inconsistent
> > results.
> >
> > Trying to be clever approach:
> > Every query puts a "lock" file with the job-id in the directory they
> read.
> > Only when there are no locks, replace the data-set as describe in the
> naive
> > approach. This will reduce the odds for problems, but is rife with
> > race-conditions. Also, if the data is read-heavy, you may never get to
> > replace the directory. Now you need a write lock to prevent new reads
> from
> > starting.
> >
> > Would hardlinks solve this problem?
> > Simply create a set of (temporary) hardlinks to the files in the
> directory
> > you want to read? Then if the old set is moved out of the way, the
> > hardlinks should still point to them. The reading job reads from the
> > hardlinks and cleans them up when done. If the hardlinks are placed in a
> > directory with the reading job-id then garbage collection should be
> > possible for crashed jobs if normal cleanup fails.
> >
> > Groetjes,
> >
> > Joep
> >
> > On Fri, Oct 12, 2012 at 8:35 AM, Harsh J <[EMAIL PROTECTED]> wrote:
> >
> >> Hey Koert,
> >>
> >> Yes the _SUCCESS (Created on successful commit-end of a job) file
> >> existence may be checked before firing the new job with the chosen
> >> input directory. This is consistent with what Oozie does as well.
> >>
> >> Since the listing of files happens post-submit() call, doing this will
> >> "just work" :)
> >>
> >> On Fri, Oct 12, 2012 at 8:00 PM, Koert Kuipers <[EMAIL PROTECTED]>
> wrote:
> >> > We have a dataset that is heavily partitioned, like this