aurora-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Maxim Khutornenko <ma...@apache.org>
Subject Re: [PROPOSAL] Revisit task ID format
Date Thu, 28 Jan 2016 04:09:24 GMT
You're right, ignore me. I guess I was mourning the loss of StorageBackfill
too hard :) Obviously, we don't have to force this change over existing
tasks and let them die out naturally. Some user scraping tools may be
broken due to this but we always advised against taking a dependency on our
task ID format.

+1 to the proposal.

On Wed, Jan 27, 2016 at 5:46 PM, Bill Farner <wfarner@apache.org> wrote:

> I don't believe backwards compatibility is an issue here.  This would be an
> alteration to generation of new task IDs.  AFAIK we don't do anything that
> requires comprehension or synthesis of previously-generated task IDs.
>
> On Wed, Jan 27, 2016 at 5:39 PM, Maxim Khutornenko <maxim@apache.org>
> wrote:
>
> > What's the cluster upgrade story going to look like? Since task IDs are
> > used as unique identifiers for Mesos, I expect this change would require
> > rebooting the entire cluster to the new format? I am not really sure how
> > this can be done in a graceful manner without sending the entire cluster
> to
> > LOST on first reconciliation run.
> >
> > Unless we have answers to the above, I am -1 to this proposal. The
> benefits
> > don't seem significant enough to offset the pain of migrating to the new
> > format.
> >
> > On Tue, Jan 26, 2016 at 6:25 PM, Mauricio Garavaglia <
> > mauriciogaravaglia@gmail.com> wrote:
> >
> > > +1 to dropping the timestamp.
> > >
> > > Agree that having the jobkey at hand has been helpful for debugging.
> > >
> > > On Tue, Jan 26, 2016 at 10:39 PM, Zameer Manji <zmanji@apache.org>
> > wrote:
> > >
> > > > +1 to removing the timestamp.
> > > >
> > > > The timestamp has not provided me with any benefit as an operator.
> The
> > > > mangled jobkey and UUID have been very useful in grepping logs and
> > > > diagnosing failing jobs.
> > > >
> > > > On Tue, Jan 26, 2016 at 3:49 PM, Zhitao Li <zhitaoli.cs@gmail.com>
> > > wrote:
> > > >
> > > > > +1 for dropping the time and keeping the mangled jobkey. Unless we
> > are
> > > > sure
> > > > > that all internal logging of Mesos master and agent contains an
> > > > identifier
> > > > > with user some user generated data, changing it to UUID will make
> > adhoc
> > > > > debugging through Mesos logging harder.
> > > > >
> > > > > On Tue, Jan 26, 2016 at 3:17 PM, Erb, Stephan <
> > > > Stephan.Erb@blue-yonder.com
> > > > > >
> > > > > wrote:
> > > > >
> > > > > > +1 for dropping the timestamp
> > > > > >
> > > > > > However, I am not sure regarding the mangled jobkey. It tends
to
> > make
> > > > it
> > > > > > easier to correlate Mesos tasks to Aurora jobs when skimming
log
> > > files,
> > > > > > viewing the Mesos-UI or even when using the Thermos [1]. I guess
> > the
> > > > > > traceability of all of those usecases could be improved, but
that
> > > would
> > > > > > probably additional work.
> > > > > >
> > > > > > [1]
> > > > > >
> > > >
> > https://github.com/apache/aurora/blob/master/docs/images/runningtask.png
> > > > > > ________________________________________
> > > > > > From: Bill Farner <wfarner@apache.org>
> > > > > > Sent: Wednesday, January 27, 2016 12:03 AM
> > > > > > To: dev@aurora.apache.org
> > > > > > Subject: [PROPOSAL] Revisit task ID format
> > > > > >
> > > > > > Context: a task ID is a unique identifier for a task.  Aurora
and
> > > Mesos
> > > > > > both require this uniqueness.  Within mesos, frameworks are
> > required
> > > to
> > > > > > craft their own task IDs as they see fit.
> > > > > >
> > > > > > Our task ID format is currently [1]
> > > > > >
> > > > > > TIMESTAMP-ROLE-ENV-JOBNAME-INSTANCE-UUID
> > > > > >
> > > > > >
> > > > > > for an example:
> > > > > >
> > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
> 1453847837931-vagrant-test-http_example_docker-1-a23f55e2-151f-409e-9cea-76ec79482533
> > > > > >
> > > > > >
> > > > > > In addition to guaranteed uniqueness, this format has the benefit
> > of
> > > > > being
> > > > > > somewhat human-friendly.  Within logs, it is somewhat possible
to
> > > > > partially
> > > > > > recognize a task based solely on the text ID.
> > > > > >
> > > > > > *I would like to propose that we remove the TIMESTAMP- prefix
> from
> > > the
> > > > > task
> > > > > > ID.*  It was originally included so that task IDs would be
> > temporally
> > > > > > sortable for scheduling prioritization.  At present, tasks are
> not
> > > > sorted
> > > > > > using the ID.
> > > > > >
> > > > > > While proposing the above, i think it's also prudent to take
the
> > > > > > opportunity to consider a complete overhaul of the ID contents.
> > *An
> > > > > > alternative approach would be to only use the UUID.*  This has
> the
> > > > > benefit
> > > > > > of decoupling arbitrary user input from the various ways task
IDs
> > are
> > > > > used
> > > > > > (as an example - mesos uses them in file names, limiting length
> and
> > > > > allowed
> > > > > > characters).  Task IDs also become fixed width, which offers
a
> > (very)
> > > > > > marginal memory reduction over the status quo, and makes console
> > line
> > > > > > wrapping more consistent when perusing logs.  Additionally,
it
> > > eschews
> > > > > the
> > > > > > potential problem of users parsing task IDs and coupling to
its
> > > format.
> > > > > >
> > > > > > Any thoughts on this?
> > > > > >
> > > > > >
> > > > > > [1]
> > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
> https://github.com/apache/aurora/blob/master/src/main/java/org/apache/aurora/scheduler/TaskIdGenerator.java
> > > > > >
> > > > >
> > > > >
> > > > >
> > > > > --
> > > > > Cheers,
> > > > >
> > > > > Zhitao Li
> > > > >
> > > > > --
> > > > > Zameer Manji
> > > > >
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message