[zeromq-dev] Hanging on client "receive" when a responder goes down in extended request-reply pattern (code in C#)

[Lindley French]

It seems to me that this is a good opportunity for credit-based flow
control. The consumers messages to the producer saying "I'm ready for X
more jobs", and the producer keeps track of how many outstanding jobs that
consumer is ready for. If the maximum number of "credit" each consumer is
willing to offer is 1, then this is equivalent to consumers saying "Give me
a job," and the producer replying with one; for larger maximums, efficiency
scales well.

If a consumer goes down, this also bounds the maximum number of lost jobs
you have to worry about a lot more than most timeout-based solutions would.


On Wed, Jan 15, 2014 at 1:00 PM, Chris Whiten <chris.whiten at gmail.com>wrote:

> That was my original thought, and what I had tried before being pointed to
> the pattern I originally described.  I've tried setting the high watermark,
> but this seems to just dictate how many messages the consumer should buffer
> before throwing messages away.  I haven't been able to find any options to
> coax the publisher into blocking when this limit is hit.
>
> Here is my push/pull pattern code that I had tried:
> http://pastebin.com/CJLaY5zk
>
>
> On Wed, Jan 15, 2014 at 12:43 PM, Bruno D. Rodrigues <
> bruno.rodrigues at litux.org> wrote:
>
>> Dumb question, but why not a simple PUSH/PULL pattern and no queues? Each
>> consumer PULLs one message to process. The process pushing data (which may
>> be more than one) will block when the HWM is hit (all consumers are busy,
>> and the “queue” from the producer is full). No need for “hold-on-i’m-busy”
>> messages, as PUSH-PULL does it already by themselves.
>>
>>
>> On Jan 15, 2014, at 16:29, Chris Whiten <chris.whiten at gmail.com> wrote:
>>
>> I'm using zeromq in a competing consumer setup, where one process is
>> pushing data to a set of worker processes to parallelize some data
>> processing.
>>
>> The publisher publishes data much more quickly than the workers can
>> process it, so I want the publisher to block when all of the consumers are
>> too busy.  I'll decide if they're too busy by having them pull data from
>> zeromq and push them to an in-memory queue on each of the workers, and if
>> that in-memory queue is too large we temporarily stop accepting messages
>> from zeromq.
>>
>> When all of the consumers are too busy, I wish for the client that is
>> pushing data to the consumers to block, so we don't lose large chunks of
>> data.  To solve that, I've set up the extended request-reply pattern (as
>> outlined in Figure 16 of the zeromq guide at
>> http://zguide.zeromq.org/page:all).  This seems to work in the happy
>> case, but if a worker goes down the client will never receive a response
>> back from that worker, and it will hang indefinitely.  Have I selected the
>> wrong pattern, or is there an easy way to remedy this problem?
>>
>> My code to test this is at http://pastebin.com/CaZpVayu
>>
>> Thanks for your help
>> _______________________________________________
>> zeromq-dev mailing list
>> zeromq-dev at lists.zeromq.org
>> http://lists.zeromq.org/mailman/listinfo/zeromq-dev
>>
>>
>>
>> _______________________________________________
>> zeromq-dev mailing list
>> zeromq-dev at lists.zeromq.org
>> http://lists.zeromq.org/mailman/listinfo/zeromq-dev
>>
>>
>
> _______________________________________________
> zeromq-dev mailing list
> zeromq-dev at lists.zeromq.org
> http://lists.zeromq.org/mailman/listinfo/zeromq-dev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.zeromq.org/pipermail/zeromq-dev/attachments/20140115/63f5c2b0/attachment.htm>