hadoop-yarn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Vinod Kumar Vavilapalli <vino...@hortonworks.com>
Subject Re: "Reservation" ambiguity
Date Wed, 05 Aug 2015 17:34:10 GMT
It wasn’t confusing to me given we know the internals, but I can see why it would be.

In my mind, what existed for a long time was ‘internal reservation of resources for individual
containers within an application’ and the newer feature is about user-driven reservation
of resources at application / workload granularity. For the newer metrics if/when we add,
we should just make the distinction explicit - for e.g. numUserReservedResources. Agreeing
with others, this doesn’t warrant us to break existing metrics even in the next major release.

Thanks
+Vinod

On Aug 5, 2015, at 10:16 AM, Carlo Curino <ccurino@microsoft.com<mailto:ccurino@microsoft.com>>
wrote:

+1 on keeping the name "reservation" for the user-visible (2).

On top of the external/internal argument that Chris makes (which I completely agree with),
I noticed the following:

While developing (2)  we spoke with lots and lots of folks both in industry and academia,
and the term
"reservation" was very evocative and intuitive. Within seconds people were using it to refer
to the functionality
and easily grasping the idea.  On the other hand, every time I spoke about (1) using the keyword
"reservation",
I had to add a bunch of context, expand, explain, and even then people were naturally drawn
to refer to it
as "hoarding of resources for large containers", or "large container management".

Other alternative names for (1) could be: "hoarded" or "prefecthed" resources.

My 2 cents...

Cheers,
Carlo

-----Original Message-----
From: Karthik Kambatla [mailto:kasha@cloudera.com]
Sent: Wednesday, August 5, 2015 8:20 AM
To: yarn-dev@hadoop.apache.org<mailto:yarn-dev@hadoop.apache.org>
Subject: Re: "Reservation" ambiguity

Inline.

On Tue, Aug 4, 2015 at 6:48 PM, Chris Douglas <cdouglas@apache.org<mailto:cdouglas@apache.org>>
wrote:

How visible are (1) reservations? They're an internal, implementation
detail exposed in metrics only to explain the edge cases they create.
Are users typically aware of them?


This is internal, and I don't think users are aware of the mechanics.
However, they do see metrics for "reserved" resources.



SLA reservations (2) are user-visible, and express the contract with
users/operators symmetrically. While (1) is a concept, renaming (2)
would require user-breaking code changes.


Yes, I don't think we should rename (2).



Unless you're discussing the intersection- the effect of reservations
(1) on a reservation (2)- it's usually clear from context... I'd
rather avoid breaking anyone listening to the metrics in Hadoop-3.


I propose to add new metrics holdMB, holdCores for reservedMB, reseveredCores. Could we deprecate
the older metrics in Hadoop-2 and Hadoop-3, and remove them in Hadoop-4?



Maybe reservations (2) could have been named "sessions", but that
collided with applications that already used it for a similar concept.
-C

On Wed, Jul 29, 2015 at 10:37 AM, Karthik Kambatla
<kasha@cloudera.com<mailto:kasha@cloudera.com>>
wrote:
Hi folks

We use the word "reservation" to mean both (1) reservations on nodes
to avoid starvation of big container asks, and (2) the recent SLA
work. This is confusing both to developers and end-users.

I was wondering if people are open to calling the first one a "hold"
and the second one a "reservation". We can change the terminology in
the code and add new metrics for hold in branch-2 and remove the
metrics for
reserved* in Hadoop-3?

Thoughts?


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