Yeah, here-in lies the problem Paolo. We want developers to be able to
extend MutableVertex. At the same time, we don't want users to be able
to call putMessages() and releaseResources() - i.e. shoot themselves in
the foot.
If you have an thoughts on how to handle this, that would be great.
Avery
On 5/25/12 10:52 AM, Paolo Castagna wrote:
Hi Avery,
thanks for your reply.
However, if you expect end users to extend MutableVertex, it should not have
package private methods inherited from BasicVertex and not implemented in
MutableVertex.
Such as:
abstract void putMessages(Iterable<M> messages);
abstract void releaseResources();
I wanted to try to extend MutableVertex because I wanted to see what it would
take to use indexes (with memory mapped files) instead of keeping everything in
RAM. I've not done much progress on that, but I needed to put my stuff in
org.apache.giraph.graph.* because of the two methods above.
Please, correct me if I am wrong.
Cheers,
Paolo
Avery Ching wrote:
Sorry for the delayed response, but feel free to extend MutableVertex if
the other Vertex implementations don't meet your needs. EdgeListVertex
is probably the most commonly used simple case, but something like
IntIntNullIntVertex is very optimized for certain primitives (not
general purposed). Still, it can be nice to same more memory...
Avery
On 5/21/12 10:08 AM, Paolo Castagna wrote:
Benjamin Heitmann wrote:
On 21 May 2012, at 17:15, Paolo Castagna wrote:
A more direct question would be: are Giraph supposed to extend
BasicVertex<I,V,E,M> when they do not find a subclass of
BasicVertex which meets
their needs?
No, they are not. However this is not explicitly documented anywhere.
Or to say it more clearly: If you look in the javadoc of BasicVertex
and if you search the mailing lists,
then you will find that users are discouraged from using/extending
BasicVertex, but you will not find any suggestion of which
Vertex to extend instead. I even asked basically the same question
once, and got very indirect answers.
But that is okay, I figured it out by trial and error ;)
Users are supposed to extend HashMapVertex or EdgeListVertex (both in
org.apache.giraph.graph).
Hi Benjamin,
right, I should have seen those (this is a good sign I should stop for
today
and continue tomorrow morning). I didn't because I was thinking: "I do
not need
my vertexes to be mutable" (since, computing PageRank does not need to
change
the topology of a graph), so I was not focusing my attention on the
MutableVertex hierarchy of classes (my mistake).
Now my question would be: why SimplePageRankVertex extends
LongDoubleFloatDoubleVertex? (But, I'll look at this tomorrow).
If you try to extend them, you will see that you can basically
plug-in any kind of existing class in the<I,V,E,M> signature,
as long as the implement the right interfaces, which are all Writable
(and WritableComparable for I).
Then you just need to add your compute() method, and you are ready to
go.
Both HashMapVertex and EdgeListVertex provide implementations of all
the housekeeping that giraph needs.
Trying to work directly by extending BasicVertex will not work, as a
lot of methods are only accessibly on the same package level.
Yep.
It would probably be a good idea to submit a small javadoc patch
which adds documentation to BasicVertex, that users need to look at
those other two classes.
Yep.
Thanks again for your help and for pointing me in the right direction.
Paolo
