Giraph
  1. Giraph
  2. GIRAPH-170

Workflow for loading RDF graph data into Giraph

    Details

    • Type: New Feature New Feature
    • Status: Open
    • Priority: Minor Minor
    • Resolution: Unresolved
    • Affects Version/s: None
    • Fix Version/s: None
    • Component/s: None
    • Labels:
      None

      Description

      W3C RDF provides a family of Web standards for exchanging graph-based data. RDF uses sets of simple binary relationships, labeling nodes and links with Web identifiers (URIs). Many public datasets are available as RDF, including the "Linked Data" cloud (see http://richard.cyganiak.de/2007/10/lod/ ). Many such datasets are listed at http://thedatahub.org/

      RDF has several standard exchange syntaxes. The oldest is RDF/XML. A simple line-oriented format is N-Triples. A format aligned with RDF's SPARQL query language is Turtle. Apache Jena and Any23 provide software to handle all these; http://incubator.apache.org/jena/ http://incubator.apache.org/any23/

      This JIRA leaves open the strategy for loading RDF data into Giraph. There are various possibilites, including exploitation of intermediate Hadoop-friendly stores, or pre-processing with e.g. Pig-based tools into a more Giraph-friendly form, or writing custom loaders. Even a HOWTO document or implementor notes here would be an advance on the current state of the art. The BluePrints Graph API (Gremlin etc.) has also been aligned with various RDF datasources.

      Related topics: multigraphs https://issues.apache.org/jira/browse/GIRAPH-141 touches on the issue (since we can't currently easily represent fully general RDF graphs since two nodes might be connected by more than one typed edge). Even without multigraphs it ought to be possible to bring RDF-sourced data
      into Giraph, e.g. perhaps some app is only interested in say the Movies + People subset of a big RDF collection.

      From Avery in email: "a helper VertexInputFormat (and maybe VertexOutputFormat) would certainly [despite GIRAPH-141] still help"

        Issue Links

          Activity

          Dan Brickley created issue -
          Hide
          Dan Brickley added a comment -

          From Paulo in email:

          """ I suspect N-Triples | N-Quads might not be the best option for
          something like Giraph. Something more like an adjacency list might be
          better.

          So, my intuition, is that if you start with RDF in N-Triples format,
          the first step would be a simple MapReduce job to group RDF statements
          by subject (eventually filtering out certain properties):

          Input:

          s1 -p1-> o1
          s1 -p2-> o2
          s1 -p2-> o3
          s2 ...

          Output (adjacency list):

          s1 (p1 o1) (p2 o2) (p2 o3)
          s2 ..."""

          Show
          Dan Brickley added a comment - From Paulo in email: """ I suspect N-Triples | N-Quads might not be the best option for something like Giraph. Something more like an adjacency list might be better. So, my intuition, is that if you start with RDF in N-Triples format, the first step would be a simple MapReduce job to group RDF statements by subject (eventually filtering out certain properties): Input: s1 - p1 -> o1 s1 - p2 -> o2 s1 - p2 -> o3 s2 ... Output (adjacency list): s1 (p1 o1) (p2 o2) (p2 o3) s2 ..."""
          Hide
          Dan Brickley added a comment -

          Another architectural note around RDF:

          RDF is basically simple factual data expressed as sets of binary relationships. In that sense it is a graph directly, already.

          However often RDF describes something that is in a deeper sense also a graph. Common examples include FOAF, where node and edge types (Person, Document, Group, etc.) can express matrix of collaboration, social linkage, etc. Or from DBpedia.org, Freebase etc., we have for example datasets of movies and actors. In the dbpedia case, it's simple enough; a movie node, an actor node, and a typed link between them. Freebase by contrast, reifies the 'starring' relationship into another node, ... so you can represent dates, character name etc. This sort of meta-information (properties of links) is also btw in the BluePrints/Gremlin API.

          One point here is that a 'starring' link pointing from a Movie to an Actor, tells us the same, but in reverse, as what we would have learned from a 'starsIn' link from the Actor to the Movie. For Giraph we may want to consider therefore adding backlinks so each node is equally aware of properties pointing both in, and out.

          Show
          Dan Brickley added a comment - Another architectural note around RDF: RDF is basically simple factual data expressed as sets of binary relationships. In that sense it is a graph directly, already. However often RDF describes something that is in a deeper sense also a graph. Common examples include FOAF, where node and edge types (Person, Document, Group, etc.) can express matrix of collaboration, social linkage, etc. Or from DBpedia.org, Freebase etc., we have for example datasets of movies and actors. In the dbpedia case, it's simple enough; a movie node, an actor node, and a typed link between them. Freebase by contrast, reifies the 'starring' relationship into another node, ... so you can represent dates, character name etc. This sort of meta-information (properties of links) is also btw in the BluePrints/Gremlin API. One point here is that a 'starring' link pointing from a Movie to an Actor, tells us the same, but in reverse, as what we would have learned from a 'starsIn' link from the Actor to the Movie. For Giraph we may want to consider therefore adding backlinks so each node is equally aware of properties pointing both in, and out.
          Paolo Castagna made changes -
          Field Original Value New Value
          Link This issue relates to GIRAPH-141 [ GIRAPH-141 ]
          Hide
          Paolo Castagna added a comment -

          we may want to consider therefore adding backlinks

          Yep. I'd like to better understand what people currently do if they need incoming and outgoing links for their processing.
          An adjacency list can be constructed listing incoming (a.k.a. backlinks) as well as outgoing links, in one MapReduce job.

          Input:

          s1 p1> o1
          s1 p2> o2
          s1 p2> o3
          s2 p1> s1
          s2 ...

          Output (adjacency list):

          s1 (out: p1 o1) (out: p2 o2) (out: p2 o3) (in: s2 p1)
          s2 ...

          Whether it is better to do it this way or have support from the Giraph APIs avoiding an initial MapReduce job to construct the adjacency list, I do not know yet.

          Show
          Paolo Castagna added a comment - we may want to consider therefore adding backlinks Yep. I'd like to better understand what people currently do if they need incoming and outgoing links for their processing. An adjacency list can be constructed listing incoming (a.k.a. backlinks) as well as outgoing links, in one MapReduce job. Input: s1 p1 > o1 s1 p2 > o2 s1 p2 > o3 s2 p1 > s1 s2 ... Output (adjacency list): s1 (out: p1 o1) (out: p2 o2) (out: p2 o3) (in: s2 p1) s2 ... Whether it is better to do it this way or have support from the Giraph APIs avoiding an initial MapReduce job to construct the adjacency list, I do not know yet.
          Show
          Dan Brickley added a comment - See also http://www.slideshare.net/billonahill/hadoop-summit-2011-using-a-hadoop-data-pipeline-to-build-a-graph-of-users-and-content https://github.com/2nd-metaman/sa-rdf-giraph
          Hide
          Dan Brickley added a comment -

          Paolo (spelled right this time... sorry!), does Pig sound like an appropriate tool for that sort of pre-processing? I thought I'd seen some graph manipulation code around somewhere that might do the ntriples to adjacency list work, but can't find the link. Closest I've found is http://thedatachef.blogspot.com/2011/05/structural-similarity-with-apache-pig.html

          https://github.com/ogrisel/pignlproc also has some code for ntriples parsing from Pig, e.g. https://github.com/ogrisel/pignlproc/blob/master/src/main/java/pignlproc/storage/UriUriNTriplesLoader.java though it doesn't (from quick look) seem to handle literal values.

          Show
          Dan Brickley added a comment - Paolo (spelled right this time... sorry!), does Pig sound like an appropriate tool for that sort of pre-processing? I thought I'd seen some graph manipulation code around somewhere that might do the ntriples to adjacency list work, but can't find the link. Closest I've found is http://thedatachef.blogspot.com/2011/05/structural-similarity-with-apache-pig.html https://github.com/ogrisel/pignlproc also has some code for ntriples parsing from Pig, e.g. https://github.com/ogrisel/pignlproc/blob/master/src/main/java/pignlproc/storage/UriUriNTriplesLoader.java though it doesn't (from quick look) seem to handle literal values.
          Hide
          Dan Brickley added a comment -

          ah https://github.com/ogrisel/pignlproc/blob/master/src/main/java/pignlproc/storage/UriStringLiteralNTriplesStorer.java handles literals; my mistake. Ok, so that'll get ntriples parsed into Pig, ... how to write them out again nicely for Giraph, or ingest directly into Giraph?

          Show
          Dan Brickley added a comment - ah https://github.com/ogrisel/pignlproc/blob/master/src/main/java/pignlproc/storage/UriStringLiteralNTriplesStorer.java handles literals; my mistake. Ok, so that'll get ntriples parsed into Pig, ... how to write them out again nicely for Giraph, or ingest directly into Giraph?
          Hide
          Paolo Castagna added a comment -

          Pig and Pig Latin can certainly be used to create adjacency lists from RDF in N-Triples|N-Quads format.
          I tend to use more plain MapReduce jobs written in Java, but I found a very old (i.e. it was using Pig version 0.6) example on how one might write an NQuadsStorage which implements LoadFunc and StoreFunc for Pig. I shared it, even if it does not even compile now, just to show how trivial that is.

          It is my intention, in the next few weeks, to create a small library to support people wanting to use Pig, HBase, MapReduce and Giraph to process RDF data.
          For Pig the first (and only?) thing to do is to implement LoadFunc and StoreFunc for RDF data. It seems possible (although not easy) to map the SPARQL algebra to Pig Latin physical operators (and SPARQL property paths to Giraph jobs? ), that would provide a good and scalable batch processing solution for those into SPARQL.
          For HBase, the first step is to store RDF data, even a plain [(G)|S|P|O] solution would do initially.
          For MapReduce, blank nodes can be painful, I have some tricks to share here. Input/output formats and record readers/writers, etc.

          In relation to Giraph, to bring the discussion on topic, until I am proven wrong, I am going for the adjacency list approach as discussed above and do graph processing as other 'usual' Giraph jobs.

          The question: what are the RDF processing use cases which are a good fit for Giraph is still open for me (and I'll find out soon).

          Show
          Paolo Castagna added a comment - Pig and Pig Latin can certainly be used to create adjacency lists from RDF in N-Triples|N-Quads format. I tend to use more plain MapReduce jobs written in Java, but I found a very old (i.e. it was using Pig version 0.6) example on how one might write an NQuadsStorage which implements LoadFunc and StoreFunc for Pig. I shared it, even if it does not even compile now, just to show how trivial that is. It is my intention, in the next few weeks, to create a small library to support people wanting to use Pig, HBase, MapReduce and Giraph to process RDF data. For Pig the first (and only?) thing to do is to implement LoadFunc and StoreFunc for RDF data. It seems possible (although not easy) to map the SPARQL algebra to Pig Latin physical operators (and SPARQL property paths to Giraph jobs? ), that would provide a good and scalable batch processing solution for those into SPARQL. For HBase, the first step is to store RDF data, even a plain [(G)|S|P|O] solution would do initially. For MapReduce, blank nodes can be painful, I have some tricks to share here. Input/output formats and record readers/writers, etc. In relation to Giraph, to bring the discussion on topic, until I am proven wrong, I am going for the adjacency list approach as discussed above and do graph processing as other 'usual' Giraph jobs. The question: what are the RDF processing use cases which are a good fit for Giraph is still open for me (and I'll find out soon).
          Hide
          Benjamin Heitmann added a comment -

          Hello,
          and sorry for being late to contributing to this discussion.

          I am currently using Giraph to implement a graph based recommendation algorithm which uses RDF data from DBPedia. I am not sure if that is enough of a use case for Paolo.

          Generally speaking, statistical analysis of semantic networks should be the most general motivation for using Giraph on RDF. In other words: Since RDF has a native graph database model and RDF processing needs to happen on web scale, Giraph could be a natural fit for processing RDF, if it would support RDF input/ingestion in a native way.

          Regarding the fundamental capabilities required for parsing NTriple files with RDF: The TextInputFormat needs a way to retrieve and alter already created nodes. Currently the assumption for the TextInputFormat class, is that it will get exactly one line for each vertex to create. That one line is assumed to hold all information necessary to create the vertex.
          However, the NTriples format does not work that way, as it can use multiple lines to describe the same subject node.

          I already raised this issue on the user mailing list. (However I did not create a Jira issue for it.) This is the fundamental capability which is lacking in Giraph. If this is enabled, parsing NTriples will be easy. The starting points for the email threads in which this was shortly discussed are in [1] and [2].

          AFAIR, Dionysis Logothetis suggested that he may look into adding this capability to giraph. So you might want to contact him directly to check on the progress.

          Now a few details on how I use RDF data for my Giraph job:
          Currently I use a subset of DBPedia, which is roughly 5.5GB unpacked.
          As this DBPedia subset stays static for all my recommendations, it is enough to preprocess it once
          using a quite simple MapReduce job. I basically join all lines on the subject of the triple,
          and then output the following line for each subject:
          SubjectURI NumberOfOutLinks Predicate1 Object1 ... PredicateN ObjectN
          (I call this the RDFAdjacencyCSV

          For my specific algorithm, the direction of the the link in the RDF graph does not play any role,
          so for each input triple, I add it once to the subject entity and once to the object entity.

          The processing job took two days, but it was my first hadoop programm, so it probably was inefficient.
          The output size was 6GB.

          For running my algorithm, my Giraph job first loads the complete DBPedia dataset in memory. While doing this it also loads the user profiles from via DistributedCache.getLocalCacheFiles(conf). This is done in my own custom TextVertexInputFormat class. The profiles are used to prime the graph, i.e. to identify the starting points for the algorithm. I also need to manage which starting points belong to which user profiles.

          Challenges which I will have in the near future:

          • Giraph does not seem to scale very well for my kind of data and processing: Independent of the number of workers, my Giraph job only uses about 30% of a 24 node machine. And I would like to utilise all available processing resources.
          • Integration of RDF reasoning capabilities: I will need to perform subclass reasoning on the DBPedia graph. The most pragmatic solution seems to be, to have an external RDF store with reasoning, and to let the Giraph workers be able to query the RDF store.

          [1] https://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201203.mbox/%3CE5D0BE74-7903-4145-BE10-52CBD6489AC8%40deri.org%3E
          [2] https://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201203.mbox/%3CC6DA4465-B387-474A-B823-84019967DA3E%40deri.org%3E

          Show
          Benjamin Heitmann added a comment - Hello, and sorry for being late to contributing to this discussion. I am currently using Giraph to implement a graph based recommendation algorithm which uses RDF data from DBPedia. I am not sure if that is enough of a use case for Paolo. Generally speaking, statistical analysis of semantic networks should be the most general motivation for using Giraph on RDF. In other words: Since RDF has a native graph database model and RDF processing needs to happen on web scale, Giraph could be a natural fit for processing RDF, if it would support RDF input/ingestion in a native way. Regarding the fundamental capabilities required for parsing NTriple files with RDF: The TextInputFormat needs a way to retrieve and alter already created nodes. Currently the assumption for the TextInputFormat class, is that it will get exactly one line for each vertex to create. That one line is assumed to hold all information necessary to create the vertex. However, the NTriples format does not work that way, as it can use multiple lines to describe the same subject node. I already raised this issue on the user mailing list. (However I did not create a Jira issue for it.) This is the fundamental capability which is lacking in Giraph. If this is enabled, parsing NTriples will be easy. The starting points for the email threads in which this was shortly discussed are in [1] and [2] . AFAIR, Dionysis Logothetis suggested that he may look into adding this capability to giraph. So you might want to contact him directly to check on the progress. Now a few details on how I use RDF data for my Giraph job: Currently I use a subset of DBPedia, which is roughly 5.5GB unpacked. As this DBPedia subset stays static for all my recommendations, it is enough to preprocess it once using a quite simple MapReduce job. I basically join all lines on the subject of the triple, and then output the following line for each subject: SubjectURI NumberOfOutLinks Predicate1 Object1 ... PredicateN ObjectN (I call this the RDFAdjacencyCSV For my specific algorithm, the direction of the the link in the RDF graph does not play any role, so for each input triple, I add it once to the subject entity and once to the object entity. The processing job took two days, but it was my first hadoop programm, so it probably was inefficient. The output size was 6GB. For running my algorithm, my Giraph job first loads the complete DBPedia dataset in memory. While doing this it also loads the user profiles from via DistributedCache.getLocalCacheFiles(conf). This is done in my own custom TextVertexInputFormat class. The profiles are used to prime the graph, i.e. to identify the starting points for the algorithm. I also need to manage which starting points belong to which user profiles. Challenges which I will have in the near future: Giraph does not seem to scale very well for my kind of data and processing: Independent of the number of workers, my Giraph job only uses about 30% of a 24 node machine. And I would like to utilise all available processing resources. Integration of RDF reasoning capabilities: I will need to perform subclass reasoning on the DBPedia graph. The most pragmatic solution seems to be, to have an external RDF store with reasoning, and to let the Giraph workers be able to query the RDF store. [1] https://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201203.mbox/%3CE5D0BE74-7903-4145-BE10-52CBD6489AC8%40deri.org%3E [2] https://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201203.mbox/%3CC6DA4465-B387-474A-B823-84019967DA3E%40deri.org%3E
          Hide
          Benjamin Heitmann added a comment -

          In addition,
          I would like to say that Paolos suggestion of providing some ready made code for Pig, HBase and MapReduce for processing RDF sounds like a really great contribution.

          Please keep us update onthe progress of that!

          Show
          Benjamin Heitmann added a comment - In addition, I would like to say that Paolos suggestion of providing some ready made code for Pig, HBase and MapReduce for processing RDF sounds like a really great contribution. Please keep us update onthe progress of that!
          Hide
          Benjamin Heitmann added a comment -

          Regarding GIRAPH-141,
          I don't think that true multigraph support is required for Giraph in order to use RDF data.

          If I have "subject1 predicate1 object1" and "subject1 predicate1 object2", then there will be a total of three vertices with 2 edges, without any conflict. If I have the same triple "subject1 predicate1 object1" two or more times, then the RDF semantics document states that all of these triples refer to the same two vertices and the edge between them in the RDF graph. So there is no need for a multigraph again.

          If we introduce literals into the mix, then we have the same thing as above, if each literal will be presented by its own Giraph vertex.

          I am not sure if I missed anything, but multigraphs dont seem to be the issue here, neither in theory, nor for my already working code.

          An issue which would be more important, is the capability to retrieve and modify an already created node from inside the TextVertexInputFormat class (as explained above).

          Show
          Benjamin Heitmann added a comment - Regarding GIRAPH-141 , I don't think that true multigraph support is required for Giraph in order to use RDF data. If I have "subject1 predicate1 object1" and "subject1 predicate1 object2", then there will be a total of three vertices with 2 edges, without any conflict. If I have the same triple "subject1 predicate1 object1" two or more times, then the RDF semantics document states that all of these triples refer to the same two vertices and the edge between them in the RDF graph. So there is no need for a multigraph again. If we introduce literals into the mix, then we have the same thing as above, if each literal will be presented by its own Giraph vertex. I am not sure if I missed anything, but multigraphs dont seem to be the issue here, neither in theory, nor for my already working code. An issue which would be more important, is the capability to retrieve and modify an already created node from inside the TextVertexInputFormat class (as explained above).
          Hide
          Paolo Castagna added a comment -

          Hi Benjamin

          > I call this the RDFAdjacencyCSV

          We came to the same conclusion. I ended up using Turtle for this, as explained here: http://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201204.mbox/%3C4F84872E.4050101%40googlemail.com%3E

          Turtle isn't splittable in general, but it can be made so simply writing all the RDF statements with the same subject on a single line.

          > I would like to say that Paolos suggestion of providing some ready made code for Pig, HBase and MapReduce for processing RDF sounds like a really great contribution.

          I am not sure what's the best place to put such code, I started with sharing small examples and experiments on GitHub, here: https://github.com/castagna/jena-grande

          > Integration of RDF reasoning capabilities: I will need to perform subclass reasoning on the DBPedia graph.

          See Apache Jena's RIOT infer command or a MapReduce version of it, here:
          https://github.com/castagna/tdbloader4/blob/master/src/main/java/org/apache/jena/tdbloader4/InferDriver.java

          I wonder if Giraph could be used to implement the RETE algorithm (http://en.wikipedia.org/wiki/Rete_algorithm) which is what Jena uses (with in memory RDF Jena models).

          Show
          Paolo Castagna added a comment - Hi Benjamin > I call this the RDFAdjacencyCSV We came to the same conclusion. I ended up using Turtle for this, as explained here: http://mail-archives.apache.org/mod_mbox/incubator-giraph-user/201204.mbox/%3C4F84872E.4050101%40googlemail.com%3E Turtle isn't splittable in general, but it can be made so simply writing all the RDF statements with the same subject on a single line. > I would like to say that Paolos suggestion of providing some ready made code for Pig, HBase and MapReduce for processing RDF sounds like a really great contribution. I am not sure what's the best place to put such code, I started with sharing small examples and experiments on GitHub, here: https://github.com/castagna/jena-grande > Integration of RDF reasoning capabilities: I will need to perform subclass reasoning on the DBPedia graph. See Apache Jena's RIOT infer command or a MapReduce version of it, here: https://github.com/castagna/tdbloader4/blob/master/src/main/java/org/apache/jena/tdbloader4/InferDriver.java I wonder if Giraph could be used to implement the RETE algorithm ( http://en.wikipedia.org/wiki/Rete_algorithm ) which is what Jena uses (with in memory RDF Jena models).
          Hide
          Sebastian Schelter added a comment -

          Independent of the number of workers, my Giraph job only uses about 30% of a 24 node machine. And I would like to utilise all available processing resources.

          It surprises me, that you don't get a higher load. If you configure your cluster to use one worker/map instance per core you should get a much higher CPU load. Could it be that either the cluster is too powerful for your graph or that your algorithm doesn't work on the whole graph all the time?

          Show
          Sebastian Schelter added a comment - Independent of the number of workers, my Giraph job only uses about 30% of a 24 node machine. And I would like to utilise all available processing resources. It surprises me, that you don't get a higher load. If you configure your cluster to use one worker/map instance per core you should get a much higher CPU load. Could it be that either the cluster is too powerful for your graph or that your algorithm doesn't work on the whole graph all the time?
          Hide
          Claudio Martella added a comment -

          Difficult to say without further investigation (and without a proper definition of 30%) but it could also be that you're stuck in I/O (messaging, checkpointing).

          Show
          Claudio Martella added a comment - Difficult to say without further investigation (and without a proper definition of 30%) but it could also be that you're stuck in I/O (messaging, checkpointing).
          Hide
          Benjamin Heitmann added a comment -

          Thanks guys for your comments!

          Paolo: I will take a look at Jena RIOT for inferencing.

          Sebastian: I did not know that it is possible to assign one mapper to each core in Hadoop, I will try that for sure. Also, my algorithm does only use a part of the graph when it runs. So that might be the easiest explanation for the observed behavior.

          Claudio: Thanks for the suggestion, I will further investigate the issue, and provide an update when I know whats going on.

          Show
          Benjamin Heitmann added a comment - Thanks guys for your comments! Paolo: I will take a look at Jena RIOT for inferencing. Sebastian: I did not know that it is possible to assign one mapper to each core in Hadoop, I will try that for sure. Also, my algorithm does only use a part of the graph when it runs. So that might be the easiest explanation for the observed behavior. Claudio: Thanks for the suggestion, I will further investigate the issue, and provide an update when I know whats going on.
          Hide
          Benjamin Heitmann added a comment -

          In addition, just for the record:

          There is actually a Jira issue about enabling the TextInputFormat class to modify a vertex which it has already created. (That was pointed out in an email on the list.)

          That jira issue is here: https://issues.apache.org/jira/browse/GIRAPH-155
          "Allow creation of graph by adding edges that span multiple workers"

          Show
          Benjamin Heitmann added a comment - In addition, just for the record: There is actually a Jira issue about enabling the TextInputFormat class to modify a vertex which it has already created. (That was pointed out in an email on the list.) That jira issue is here: https://issues.apache.org/jira/browse/GIRAPH-155 "Allow creation of graph by adding edges that span multiple workers"
          Paolo Castagna made changes -
          Link This issue relates to GIRAPH-155 [ GIRAPH-155 ]

            People

            • Assignee:
              Unassigned
              Reporter:
              Dan Brickley
            • Votes:
              0 Vote for this issue
              Watchers:
              8 Start watching this issue

              Dates

              • Created:
                Updated:

                Development