flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Niklas Semmler <nsemm...@inet.tu-berlin.de>
Subject Re: How do network transmissions in Flink work?
Date Mon, 20 Jul 2015 17:00:46 GMT
Hello Ufuk,

thank you very much for the answer. You helped me to bring a great deal 
of context into the problem :).

I have one final question: What is a good indicator that the transfer of 
data contained in a single ResultPartition is finished? Is there any? Or 
can the amount of retrieved data be retrieved from the consumer?

So far the only indicator I could come up with is the release of the 
ResultPartition or the state change to "FINISHED" of the task. However, 
as far as I understand the assigned resources will only be released 
after all ResultPartitions are transferred and the task is finished, so 
that seems to be a rather impractical indicator for the end of the data 

Sorry for the late reply.


On 13.07.2015 15:04, Ufuk Celebi wrote:
> Hey Niklas,
> there is also this Wiki entry: https://cwiki.apache.org/confluence/display/FLINK/Data+exchange+between+tasks
> On 09 Jul 2015, at 21:32, Niklas Semmler <nsemmler@inet.tu-berlin.de> wrote:
>> 1. What does the number of ResultSubpartition instances in the ResultPartition correspond
to? Is one assigned to each consuming task? If so, how can I find for each ResultSubpartition
the corresponding Task, Slot or similar? If not how is decided which piece of the data is
routed to which consuming task?
> Yes, for each consuming task. The wiring depends on the DistributionPattern and the parallelism
of the producing and consuming operator. You can look into the ExecutionGraph to see how the
wiring works (see connect* methods in ExecutionVertex class). Each subpartition corresponds
to an ExecutionEdge, which connects two ExecutionVertex instances, which is an abstraction
for tasks at runtime. This is essentially also where the routing is set.
> Currently there is no way to get from the subpartition to the corresponding task. You
would have to look into the places where the instances are created and pass the reference.
The RuntimeEnvironmenet or Task class create these instances when a new task is submitted
to a task manager.
>> 2. What defines the number of Buffer instances per ResultSubpartition? Does one Buffer
correspond to exactly one serialized Record? Is a Record the single output of an operator,
are there multiple records per operator, or
>> does it differ depending on the operator?
> The number of produced buffers depends on the data the corresponding operator/user function
produces. Each produced record is serialized into a buffer. It can span multiple buffers depending
on the record size.
> There can be zero or more records per produced partition. (There will always be at least
a single buffer containing an end-of-partition event per partition though.)
>> 3. Or are the Buffers defined in a completely different manner? In that case, could
you give me a pointer to understand how Buffer instances are used?
> The buffers is a wrapper for a MemorySegment with a reference to a buffer pool, which
owns the buffer. Buffers are recycled after they have been consumed (e.g. after being written
to the TCP channel or by the user code).
> Feel free to ask further questions or give feedback if you encounter anything you find
weird. :-)
> – Ufuk

PhD Student / Research Assistant
INET, TU Berlin
Room 4.029
Marchstr 23
10587 Berlin
Tel: +49 30 314 78752

View raw message