Mailing-List: contact hdfs-dev-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Reply-To: hdfs-dev@hadoop.apache.org
Received-SPF: pass (athena.apache.org: domain of andrew.wang@cloudera.com
 designates 209.85.212.181 as permitted sender)
MIME-Version: 1.0
In-Reply-To: 
 <CAFiYk=oZLeta83U7JPAikak6B0p4h5Xx4U3+BiCTybskB3jj8w@mail.gmail.com>
References: 
 <CAFiYk=pEt1PzmBYqjNxWZJHHZ-f0H75kyFK78Jkn6cCsYjdQGA@mail.gmail.com>
 <CAEfeEE-cX_p9nYhCZsx_p6fAYjzmEOr7r6RgRbALX2JVWLR8dQ@mail.gmail.com>
 <CAGB5D2ZoZLjCJSLUA4VYMroXu8_004BKBwD9xpLaGVAJo8gM1w@mail.gmail.com>
 <CAFiYk=pwJYp3Hj4o+tU46GZ0EE6CM4=-fwsuxYbaS34c08+HHQ@mail.gmail.com>
 <CAGB5D2aBe28V2t9tuAaktv8szCQ8q+Fs_+rLrkh6JfeSZPpYHA@mail.gmail.com>
 <CAFiYk=p1xYXz8v2d+HpmXPtZ_42kF=nJcKqSp1UpTp+shyzDHQ@mail.gmail.com>
 <CAFiYk=rJcLD+Uu2qu9twFRUJomVK55CqbesyuyNwLF0p88FMYw@mail.gmail.com>
 <CA+qbEUMdSawLhaD7fjSGYOXNzrLKejVHy8j7rZf9RmE672O0Hw@mail.gmail.com>
 <CADdVdVHTFboNwK_vYiBAa1LmDOBYTZMxy9rk4D9OtooOvdhwjg@mail.gmail.com>
 <CA+qbEUPPejYqDMagGJMYx98ncnAMNXzSdue=cWPw5zYV9h7OHQ@mail.gmail.com>
 <CAFiYk=qG5CVo8rV-ioAPJ1aqYg_uzJGyMUr5TrE9-TZ8KEwndA@mail.gmail.com>
 <CA+qbEUPGOkmK6mJcdOGqX4JveNTRv0d7AYOK+udUR9ruzE_Xgw@mail.gmail.com>
 <CAFiYk=oZLeta83U7JPAikak6B0p4h5Xx4U3+BiCTybskB3jj8w@mail.gmail.com>
From: Andrew Wang <andrew.wang@cloudera.com>
Date: Thu, 25 Sep 2014 17:54:26 -0700
Message-ID: 
 <CAGB5D2brUa1Br=rw_eATZ+gkgHV49R-=6R8imrDLheJS7TaqwA@mail.gmail.com>
Subject: Re: [VOTE] Merge HDFS-6581 to trunk - Writing to replicas in memory.
To: "hdfs-dev@hadoop.apache.org" <hdfs-dev@hadoop.apache.org>
Content-Type: multipart/alternative; boundary=047d7bd91e840353920503ed5db9

--047d7bd91e840353920503ed5db9
Content-Type: text/plain; charset=UTF-8

-0.

The goal of this feature is to improve write performance. I thought it was
thus reasonable for me to ask for write benchmarks. However, this request
was not received warmly. For a performance oriented feature, I expected
that benchmarks would have been run both to guide the design and evaluate
the implementation. Doing so only after a merge vote is called, and only
after repeated requests, is unorthodox to say the least.

My questions about plans to further improve performance were also not
received warmly. Based on the microbenchmark results, we're still not at a
significant fraction of memory bandwidth. What I've been looking for are
plans to get us there, so we could evaluate how those plans might interact
with this feature. I know Arpit had prototyped short-circuit writes a while
ago, and SCW has the potential to take us well past 39% gains. However,
apparently there are no future plans in this direction.

So, -0. I think this merge vote was called prematurely, and some of the
merge-related interactions have had a tinge of hostility. However, I also
agree with Suresh that merging this will likely not affect the stability of
existing features, and ultimately, I respect all of you as developers and
clearly you all want this in now. I've always valued our collaboration on
the project and I hope can continue fruitfully into the future, I just also
hope we can do it with the intensity dialed down.

Best,
Andrew

--047d7bd91e840353920503ed5db9--