Mailing-List: contact dev-help@flink.incubator.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@flink.incubator.apache.org
Received-SPF: pass (nike.apache.org: domain of vasilikikalavri@gmail.com
 designates 209.85.217.171 as permitted sender)
MIME-Version: 1.0
In-Reply-To: 
 <CANC1h_toP-fHcZSL9bRqcwHg75h8QNskO+2XnCyxAEfDmaXwAA@mail.gmail.com>
References: 
 <CAJZ2dcV-Qt=w9bXDU7LXCf+qkscQF+-9_UN7sX7O1S_S9=GUqA@mail.gmail.com>
	<CANC1h_toP-fHcZSL9bRqcwHg75h8QNskO+2XnCyxAEfDmaXwAA@mail.gmail.com>
Date: Fri, 11 Jul 2014 16:55:53 +0200
Message-ID: 
 <CAJZ2dcX8ZpWYFpAdAoOR0zCTTuf7W=brzCEp56D0q9VFq7NvRQ@mail.gmail.com>
Subject: Re: Fixpoint api
From: Vasiliki Kalavri <vasilikikalavri@gmail.com>
To: dev@flink.incubator.apache.org
Content-Type: multipart/alternative; boundary=001a11c260d04ea7d204fdec23d0

--001a11c260d04ea7d204fdec23d0
Content-Type: text/plain; charset=UTF-8

Hey,

thanks for the reply!

My delta iteration doesn't have a convergence criterion (other than the
default empty workset that is).
If I'm not mistaken, one can't define a custom convergence criterion in a
delta iteration.
I think I had opened an issue for this at some point, but it was low
priority and we didn't decide on it.

Even if I do define a custom convergence criterion though, that would mean
that I will need to run at least one delta iteration, even if the maximum
number of iterations has already been reached.

What I am looking for is more something like the following:

int iterationsLeft;

bulkResult = doBulkIterations(parameters, dependencies, stepFunction,
maxIterations); //somehow sets iterationsLeft
if (iterationsLeft > 0 ) {
  doDeltaIterations(bulkResult, dependencies, stepFunction, iterationsLeft);
}

Cheers,
V.


On 11 July 2014 14:25, Stephan Ewen <sewen@apache.org> wrote:

> Hey Vasia!
>
> Nice work, it looks really cool!
>
>
> Concerning the problem of figuring out the number of iterations:
>
>  - In the long run we can save this with the incremental rollout that we
> are working on
>
>  - As a temporary workaround, you could store the number of elapsed
> iterations in a static variable somewhere. You set it from the "open()" or
> "close()" function of one of the UDFs. The convergence criterion for the
> delta iterations can pick it up from there (access the static variable).
>
>
> Concerning the edges with / without weights: Can you use a custom data type
> and have the weight "null" when they are not set/used?
>
> Greetings,
> Stephan
>

--001a11c260d04ea7d204fdec23d0--