docker image crashes when RDFS is turned on

[mpagni12]

Current Behavior

I am building a fairly complex repository with about 250 mio triples, by using many LOAD <> INTO GRAPH <> and DELETE {} INSERT {} WHERE {} statements. The code run nicely in GraphDB with RDFS inference turned on. I am attempting to reproduce it in the latest eclipse/rdf4j-workbench docker image, launched as follow:

docker run    -p 8080:8080     -e JAVA_OPTS="-Xms2G -Xmx24G"     -v ~/scratch/rdf4j/data:/var/rdf4j     -v ~/scratch/rdf4j/logs:/usr/local/tomcat/logs eclipse/rdf4j-workbench:latest

The repo build nicely without RDFS inference with the following config:

@prefix config: <tag:rdf4j.org,2023:config/> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix xsd: <http://www.w3.org/2001/XMLSchema#> .

[] a config:Repository;
  rdfs:label "Native store without RDF Schema inferencing";
  config:rep.id "JLW_Native_RDFS";
  config:rep.impl [
      config:rep.type "openrdf:SailRepository";
      config:sail.impl [
          config:delegate [
              config:native.tripleIndexes "spoc,posc,cpos";
              config:sail.defaultQueryEvaluationMode "STANDARD";
              config:sail.iterationCacheSyncThreshold "100000"^^xsd:long;
              config:sail.type "openrdf:NativeStore"
            ] 
          # ; config:sail.type "rdf4j:SchemaCachingRDFSInferencer"
        ]
    ] . 

Now if I uncomment the RDFS inference line, create a new repo and restart the creation process, the repo will build for a while, then the docker image crashes without anything meaningful being reported in the rdf4j logs. I have observed the issue on both OXS (M1) and Linux. I suspect that the parameters of the JVM are not optimal. Any help would be greatly appreciated.

Expected Behavior

no crash

Steps To Reproduce

No response

Version

Latest version: 5.0.3

Are you interested in contributing a solution yourself?

Perhaps?

Anything else?

No response

[hmottestad]

What does "mio" mean?

[JervenBolleman]

mio = million

[hmottestad]

@mpagni12 You could try:

• First insert all your RDFS data, including all classes and properties that are used in your data
• Then load your data

Our reasoner doesn't support deletions very well. Essentially when you delete anything, whatever it is, it needs to delete all inferred data and infer everything again.

Any chance you could share your data with us?

[mpagni12]

Sorry for the French notation

[mpagni12]

• First insert all your RDFS data, including all classes and properties that are used in your data
• Then load your data

Not sure I can achieve this distinction in my older projects. On a more recent project, I have turned inferencing off, and rely on SPARQL updates exclusively. It is much more simple.

Our reasoner doesn't support deletions very well. Essentially when you delete anything, whatever it is, it needs to delete all inferred data and infer everything again.

Thanks for this highlighting clarification. It explains why some innocuous DELETE statement seems to take ages in RDF4J, while they are instantaneous in GraphDB.

Any chance you could share your data with us?

Yes, sooner or later. The older project is ENPKG (paper, resources) and everything is available. Two other projects are on the way. I am preparing a new version of kgsteward that has powered ENPKG. I will advertise it when the code becomes more stable and a user manual is available.

[JervenBolleman]

@hmottestad as @mpagni12 is in my organization I have contacted him and will attempt to debug the symptoms outside of docker with him.

[mpagni12]

Thanks a lot for your supports

[hmottestad]

@JervenBolleman Did you guys figure this out?