+1
I'm fine with this. If we agree to return an Iterable, then we should
make sure to either throw if the size of the Iterable > messages.size()
to at the very least LOG.warn("This combiner is likely to be implemented
wrong"). I prefer an exception, since we have no use case for expanding
the set of messages.
Also, I'd like to have something in the javadoc saying something like
"While the number of messages returned can be equal to the same number
of messages that was inputted, the purpose of the combiner is to reduced
the number of messages from the input."
Avery
On 1/13/12 9:34 AM, Claudio Martella wrote:
Ok,
I guess we can vote then about this, what do you think?
Shall we take 72h?
I'm +1 for returning an iterable that can be empty.
I'm +1 for the returned iterable to be<= messages.size()
On Tue, Jan 10, 2012 at 9:48 PM, Sebastian Schelter<s...@apache.org> wrote:
I think we should make the combiner return a list/iterable that can
potentially be empty. However we should assume that the number of
elements returned is smaller than or equal to the number of input
elements (whats the use of a combiner if this is not given?). I also
concur that the code should not depend on the combiner being applied
(similar to the way combiners work in hadoop).
--sebastian
2012/1/10 Jakob Homan<jgho...@gmail.com>:
A composite object would essentially be a wrapper around a list and
introduce the need for all vertices to be ready to extract that list
at all times. For instance, a combiner passed 10 messages may be able
to combine 7 of them but do nothing with the other three, leaving four
messages. If we allow zero or one return elements, the combiner would
have to create a composite object with a list of those four messages,
whereas if we return a list, it just skips that step and returns the
four messages. Additionally, the receiving vertex would have to
handle the possibility of a composite object every time even though
the combiner may or may not have been run during the superstep, or
even included in that job (since combiners are optional to the job
itself). It would be better if one could write a Giraph application
that was completely agnostic of whether or not a combiner was
included.
On Tue, Jan 10, 2012 at 12:00 PM, Claudio Martella
<claudio.marte...@gmail.com> wrote:
I believe the argument of not letting users shoot their foot doesn't
stand :) Once you give them any API they have the power to do anything
wrong, as they already can with Giraph (or anything else for what it
matters), by designing an algorithm wrongly (which would be what it
would turn out to be a wrong combiner). It's definitely true that a
composite object would make the grouping (List<Group>) but I thought
we were talking about simplifying life to users :). I think it would
be more flexible (for the present and for the future) and also more
elegant, but not necessarily a must (although it'd come practically
for free).
Very cool discussion.
On Tue, Jan 10, 2012 at 8:30 PM, Jakob Homan<jgho...@gmail.com> wrote:
Combiners can only modify the messages sent to a single vertex, so they can't
send messages to other vertices.
Yeah, the more I've thought about this, the more problematic it would
be. These new messages may be generated upon arrival at the
destination vertex (since combiners can be run on the receiving vertex
before processing as well). When would they be forwarded to their new
destinations at that point? It would be possible to get into a
feedback loop of messages jumping around before a superstep could ever
actually be done.
That being said, our inability to think of a good application doesn't
mean there won't be one in the future, and it's probably better to be
more flexible than try to impose what appears optimal now. The
benefit of forcing 0 or 1 message from a combiner seems less than the
flexibility of allowing another list of messages (which may or may not
be the same number of elements as the original, less than, or even
more than).
Good discussion (it's making me really think about this)!
Agreed.
On Tue, Jan 10, 2012 at 11:23 AM, Avery Ching<ach...@apache.org> wrote:
The general idea of combiners is to reduce the number of messages sent.
Combiners are purely an optimization and the application should work
correctly without it (since it's never guaranteed to actually be called).
Combiners can only modify the messages sent to a single vertex, so they
can't send messages to other vertices. Any other work (i.e. sending
messages) should be done by the vertex in the compute() method.
While I think that grouping behavior could actually be implemented within a
message object (still reducing the number of messages to 1 or 0) I suppose
that in some simple cases (i.e. grouping), it might be easier by doing it in
the combiner as you both have mentioned? The only thing I suppose I'm
concerned about is letting users do something that is not optimal.
Generally, expanding messages is not what you want your combiner to do.
Also, since grouping behavior can be implemented in the message object, it
forces users to avoid shooting themselves in the foot.
Good discussion (it's making me really think about this)!
Avery
On 1/10/12 10:32 AM, Claudio Martella wrote:
Ok, now i see where you're going. I guess that the thing here is that
the combiner would "act" like (on its behalf) D, and to do so
concretely it would probably need some local data related to D (edges
values? vertexvalue?).
I also think that k> n is also possible in principle and we could let
the user decide whether to use this power or not, once/if we agree
that letting the user send k messages in the combiner is useful (and
the grouping behavior shown by the label propagation example should do
so).
On Tue, Jan 10, 2012 at 7:04 PM, Jakob Homan<jgho...@gmail.com> wrote:
Those two messages would have gone to D, been expanded to, say, 4,
which would have then then been sent to, say, M. This would save the
sending of the two to D and send the 4 directly to M. I'm not saying
it's a great example, but it is legal. This is of course assuming
that combiners can generate messages bound for vertices other than the
original destination, which I don't know if that has even been
discussed.
On Tue, Jan 10, 2012 at 9:49 AM, Claudio Martella
<claudio.marte...@gmail.com> wrote:
i'm not sure i understand what you'd save here. if the two messages
were going to be expanded to k messages on the destination worker D,
but you expand them on W, you end up sending k messages instead of 2.
right?
On Tue, Jan 10, 2012 at 6:26 PM, Jakob Homan<jgho...@gmail.com> wrote:
it doesn't have to be expand, k, the number of elements returned by
the combiner, can still be smaller than n,
Right. Grouping would be the most common case. It would be possible
to be great than k, as well. For instance, consider two messages,
both generated on the same worker (W) by two two different vertices,
both bound for another vertex, Z. A combiner on W could get both of
these messages, do some work on them, as it would have knowledge of
both, and generate some arbitrary number of messages bound for other
vertices (thus saving the shuffle/transfer of the original messages).
On Tue, Jan 10, 2012 at 12:08 AM, Claudio Martella
<claudio.marte...@gmail.com> wrote:
it doesn't have to be expand, k, the number of elements returned by
the combiner, can still be smaller than n, the size of the messages
parameter. as a first example, you can imagine your vertex receiving
semantically-different classes/types of messages, and you can imagine
willing to be summarizing them in different messages, i.e. if your
messages come along with labels or just simply by the source vertex,
if required by the algorithm, think of label propagation to have just
an example, or some sort of labeled-pagerank.
On Tue, Jan 10, 2012 at 3:05 AM, Avery Ching<ach...@apache.org>
wrote:
I agree that C&A doesn't require it, however, I can't think of why I
would
want to use a combiner to expand the number of messages. Can you?
Avery
On 1/9/12 3:57 PM, Jakob Homan wrote:
In my opinion that means reducing to a single message or none at
all.
C&A doesn't require this, however. Hadoop's combiner interface, for
instance, doesn't require a single or no value to be returned; it
has
the same interface as a reducer, zero or more values. Would
adapting
the semantics of Giraph's combiner to return a list of messages
(possibly empty) make it more useful?
On Mon, Jan 9, 2012 at 3:21 PM, Claudio Martella
<claudio.marte...@gmail.com> wrote:
Yes, what is you say is completely reasonable, you convinced me :)
On Mon, Jan 9, 2012 at 11:28 PM, Avery Ching<ach...@apache.org>
wrote:
Combiners should be commutative and associative. In my opinion
that
means
reducing to a single message or none at all. Can you think of a
case
when
more than 1 message should be returned from a combiner? I know
that
returning null isn't preferable in general, but I think that
functionality
(returning no messages), is nice to have and isn't a huge amount
of work
on
our side.
Avery
On 1/9/12 12:13 PM, Claudio Martella wrote:
To clarify, I was not discussing the possibility for combine to
return
null. I see why it would be useful, given that combine returns M,
there's no other way to let combiner ask not to send any message,
although i agree with Jakob, I also believe returning null should
be
avoided but only used, roughly, as an init value for a
reference/pointer.
Perhaps, we could, but i'm just thinking out loud here, let
combine()
return Iterable<M>, basicallly letting it define what to combine
to
({0, 1, k } messages). It would be a powerful extension to the
model,
but maybe it's too much.
As far as the size of the messages parameter, I agree with you
that 0
messages gives nothing to combine and it would be somehow
awkward, it
was more a matter of synching it with the other methods getting
the
messages parameter.
Probably, having a more clear javadoc will do the job here.
What do you think?
On Mon, Jan 9, 2012 at 8:42 PM, Jakob Homan<jgho...@gmail.com>
wrote:
I'm not a big fan of returning null as it adds extra complexity
to the
calling code (null checks, or not, since people usually will
forget
them). Avery is correct that combiners are application
specific. Is
it conceivable that one would want to write a combiner that
returned
something for an input of no parameters, ie combining the empty
list
doesn't return the empty list? I imagine for most combiners,
combining a single message would result in that message.
On Mon, Jan 9, 2012 at 11:28 AM, Avery Ching<ach...@apache.org>
wrote:
The javadoc for VertexCombiner#combine() is
/**
* Combines message values for a particular vertex index.
*
* @param vertexIndex Index of the vertex getting these
messages
* @param msgList List of the messages to be combined
* @return Message that is combined from {@link MsgList} or
null if
no
* message it to be sent
* @throws IOException
*/
I think we are somewhat vague on what a combiner can return to
support
various use cases. A combiner should be particular to a
particular
compute() algorithm. I think it should be legal to return null
from
a
combiner, in that case, no message should be sent to that
vertex.
It seems like it would be an overhead to call a combiner when
there
are
0
messages. I can't see a case where that would be useful.
Perhaps we
should
change the javadoc to insure that msgList must contain at least
one
message
to have combine() being called.
Avery
On 1/9/12 5:37 AM, Claudio Martella wrote:
Hi Sebastian,
yes, that was my point, I agree completely with you.
Fixing my test was not the issue, my question was whether we
want to
define explicitly the semantics of this scenario.
Personally, I believe the combiner should be ready to receive
0
messages, as it's the case of BasicVertex::initialize(),
putMessages()
and compute(), and act accordingly.
In the particular example, I believe the SimpleSumCombiner is
bugged.
It's true that the sum of no values is 0, but it's also true
that
the
null return semantics of combine() is more suitable for this
exact
situation.
On Mon, Jan 9, 2012 at 2:21 PM, Sebastian
Schelter<s...@apache.org>
wrote:
I think we currently implicitly assume that there is at least
one
element in the Iterable passed to the combiner. The messaging
code
only
invokes the combiner only if at least one message for the
target
vertex
has been sent.
However, we should not rely on implicit implementation
details but
explicitly specify the semantics of combiners.
--sebastian
On 09.01.2012 13:29, Claudio Martella wrote:
Hello list,
for GIRAPH-45 I'm touching the incoming messages and hit an
interesting problem with the combiner semantics.
currently, my code fails testBspCombiner for the following
reason:
SimpleSumCombiner::compute() returns a value even if there
are no
messages in the iterator (in this case it returns 0) and for
this
reason the vertices get activated at each superstep.
At each superstep, under-the-hood, I pass the combiner for
each
vertex
an Iterable, which can be empty:
public Iterable<M> getMessages(I vertexId) {
Iterable<M> messages =
inMessages.getMessages(vertexId);
if (combiner != null) {
M combinedMsg;
try {
combinedMsg =
combiner.combine(vertexId,
messages);
} catch (IOException e) {
throw new RuntimeException("could not
combine",
e);
}
if (combinedMsg != null) {
List<M> tmp = new
ArrayList<M>(1);
tmp.add(combinedMsg);
messages = tmp;
} else {
messages = new ArrayList<M>(0);
}
}
return messages;
}
the Iterable returned by this methods is passed to
basicVertex.putMessages() right before the compute().
Now, the question is: who's wrong? The combiner code that
returns
a
sum of 0 over no values, or the framework that calls the
combiner
with
0 messages?
--
Claudio Martella
claudio.marte...@gmail.com
--
Claudio Martella
claudio.marte...@gmail.com
--
Claudio Martella
claudio.marte...@gmail.com
--
Claudio Martella
claudio.marte...@gmail.com
