[
https://issues.apache.org/jira/browse/FLINK-2899?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14980448#comment-14980448
]
Chesnay Schepler commented on FLINK-2899:
-----------------------------------------
Looking at the Graph class, when groupReduceOnEdges (..., TypeInformation<T>
typeInfo) is used, two separate type hints are given to the system. One via
.returns(typeInfo), and one via the ResultTypeQueryable interface in the
ApplyNeighborCoGroupFunction class. When mixing these two, the interface takes
priority as far as i know, so the .returns() call is always ignored.
Actually, the usage of the interface doesn't make a lot of sense to me, because
it only calls TypeExtractor.createTypeInfo(...) to extract the generic typeinfo
from a function, which is the same thing that would happen when the interface
wouldn't be implemented. I don't think the function within the
ApplyNeightborCoGroupFunction holds more information.
Without trying things out myself, I assume that 2 solutions exist: either
always supply the ReturnType with .returns() and scrap the ResultTypeQueryable,
or ditch .returns() and pass the TypeInfo to the function and always use the
interface just like you did in your hack.
> The groupReduceOn* methods which take types as a parameter fail with
> TypeErasure
> --------------------------------------------------------------------------------
>
> Key: FLINK-2899
> URL: https://issues.apache.org/jira/browse/FLINK-2899
> Project: Flink
> Issue Type: Bug
> Components: Gelly
> Affects Versions: 0.10
> Reporter: Andra Lungu
>
> I tried calling groupReduceOnEdges (EdgesFunctionWithVertexValue<K, VV, EV,
> T> edgesFunction, EdgeDirection direction, TypeInformation<T> typeInfo) in
> order to make the vertex-centric version of the Triangle Count library method
> applicable to any kind of key and I got a TypeErasure Exception.
> After doing a bit of debugging (see the hack in
> https://github.com/andralungu/flink/tree/trianglecount-vertexcentric), I saw
> that actually the call to
> TypeExtractor.createTypeInfo(NeighborsFunctionWithVertexValue.class, .... in
> ApplyNeighborCoGroupFunction does not work properly, i.e. it returns null.
> From what I see, the coGroup in groupReduceOnNeighbors tries to infer a type
> before "returns" is called.
> I may be missing something, but that particular feature (groupReduceOn with
> types) is not documented or tested so we would also need some tests for that.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
