Re: Update on hadoop-0.23

[Arun C Murthy <ac...@hortonworks.com>]

On Sep 30, 2011, at 1:13 PM, Todd Lipcon wrote:

> On Fri, Sep 30, 2011 at 11:44 AM, Roman Shaposhnik <rv...@apache.org> wrote:
>> I apologize if my level of institutional knowledge of these things is
>> lacking, but do you have any
>> benchmarking results between 0.22 and 0.20.2xx? The reason I'm asking
>> is twofold -- I really
>> would like to see an objective numbers qualifying the viability of
>> 0.22 from the performance stand point,
>> but more importantly I would really like to include the benchmarking
>> code into Bigtop.
> 
> 0.22 currently suffers from MAPREDUCE-2266, which, last time I
> benchmarked it, caused a significant slowdown. iirc a terasort ran
> something like twice as slow on my test cluster due to this bug.
> 0.23/MR2 doesn't suffer from this bug.
> 

I don't really know where to start. CHANGES.txt in branch-0.20-security has the full list.

If I remember right, long ago (late 2009)  we benchmarked .21 with gridmix and saw >30% prior to abandoning .21.

Since then 0.20.2xx has had innumerable improvements to JobTracker, TaskTracker etc. etc. 
# JobTracker itself is almost thrice as fast as it used to be in 2009.
# The scheduler is significantly better (>2x locality) and throughput.
# TaskTracker has had innumerable fixes for dist.cache, task launch, shutdown (MR-2266 and lots of other similar fixes). 
# The MR runtime has fixes for latency on innumerable fronts.

Other regressions:
# Security
# Support for multi-tenant clusters.
# Tonnes of operability fixes (jobhistory, task logs i.e. MR-1100) for running MR clusters.

The one redeeming aspect for .22 is the shuffle based on the work we did for winning Terasort/Petasort in 2009 but 0.23 has even more work there with zero-copy with netty (yaay! no more jetty! Thanks to @cdouglas).

> In terms of bugs -- same question. Is there any publicly available
> list of, at least, the critical
> ones that make 0.22 not viable from your point of view?

We marked a lot of them as blockers on .22 and they were discarded by the release master(s). branch-0.20-security/CHANGES.txt is the full list. I really can't spend time enumerating over 4000 commits and > 2000 (?) jiras to that branch at this point.

In my opinion, as someone who has helped develop/run/support very large installs and done this for over 5 1/2 years, a major release with regression on features (security, multi-tenancy) and scalability, performance etc. is distinctly _unviable_.

----

Again, none of this is meant to say you should invest time on fixing them or releasing 0.22 as it stands - just, please, don't label it in a manner which helps build unreasonable expectations among users about it's viability & usability.

thanks,
Arun

Re: Update on hadoop-0.23

[Konstantin Shvachko <sh...@gmail.com>]

I am very glad that the development and testing of 0.23 is going so well.
I see a lot of commits and hundreds of changes going in literally every day.
It is great to see the new technology building!

On the criticism of the 0.22 release.
Arun has a top-down view and I agree a lot of progress have been
achieved with the framework.
My bottom-up view is that you first need a reliable storage layer. If
the file system looses blocks or worse messes up with the image and/or
journals, the performance of the framework is your second problem. I
have said that before. Based on my experience it take time to
stabilize a file system. Anybody seen one that has been stabilized in
less than 2 years?
I do not see the 0.22 release as a wasted effort. And if the progress
with it contributes to the 0.23 rush I am twice as happy.

Thanks,
--Konstantin

On Fri, Sep 30, 2011 at 3:00 PM, Arun C Murthy <ac...@hortonworks.com> wrote:
>
> On Sep 30, 2011, at 1:13 PM, Todd Lipcon wrote:
>
>> On Fri, Sep 30, 2011 at 11:44 AM, Roman Shaposhnik <rv...@apache.org> wrote:
>>> I apologize if my level of institutional knowledge of these things is
>>> lacking, but do you have any
>>> benchmarking results between 0.22 and 0.20.2xx? The reason I'm asking
>>> is twofold -- I really
>>> would like to see an objective numbers qualifying the viability of
>>> 0.22 from the performance stand point,
>>> but more importantly I would really like to include the benchmarking
>>> code into Bigtop.
>>
>> 0.22 currently suffers from MAPREDUCE-2266, which, last time I
>> benchmarked it, caused a significant slowdown. iirc a terasort ran
>> something like twice as slow on my test cluster due to this bug.
>> 0.23/MR2 doesn't suffer from this bug.
>>
>
> I don't really know where to start. CHANGES.txt in branch-0.20-security has the full list.
>
> If I remember right, long ago (late 2009)  we benchmarked .21 with gridmix and saw >30% prior to abandoning .21.
>
> Since then 0.20.2xx has had innumerable improvements to JobTracker, TaskTracker etc. etc.
> # JobTracker itself is almost thrice as fast as it used to be in 2009.
> # The scheduler is significantly better (>2x locality) and throughput.
> # TaskTracker has had innumerable fixes for dist.cache, task launch, shutdown (MR-2266 and lots of other similar fixes).
> # The MR runtime has fixes for latency on innumerable fronts.
>
> Other regressions:
> # Security
> # Support for multi-tenant clusters.
> # Tonnes of operability fixes (jobhistory, task logs i.e. MR-1100) for running MR clusters.
>
> The one redeeming aspect for .22 is the shuffle based on the work we did for winning Terasort/Petasort in 2009 but 0.23 has even more work there with zero-copy with netty (yaay! no more jetty! Thanks to @cdouglas).
>
>> In terms of bugs -- same question. Is there any publicly available
>> list of, at least, the critical
>> ones that make 0.22 not viable from your point of view?
>
> We marked a lot of them as blockers on .22 and they were discarded by the release master(s). branch-0.20-security/CHANGES.txt is the full list. I really can't spend time enumerating over 4000 commits and > 2000 (?) jiras to that branch at this point.
>
> In my opinion, as someone who has helped develop/run/support very large installs and done this for over 5 1/2 years, a major release with regression on features (security, multi-tenancy) and scalability, performance etc. is distinctly _unviable_.
>
> ----
>
> Again, none of this is meant to say you should invest time on fixing them or releasing 0.22 as it stands - just, please, don't label it in a manner which helps build unreasonable expectations among users about it's viability & usability.
>
> thanks,
> Arun
>
>