Replies: 6 comments 10 replies
-
|
+1 on multiple labels for nodes and edges/relationships could be pretty handy e.g. for
|
Beta Was this translation helpful? Give feedback.
-
|
Hi All, Unfortunately, we don't have good ideas here, and this is not in our roadmap. I have always found the multi-label decision of the original property graphs not well-thought. If we were to support multiple labels, I would want to think about how to support a general type hierarchy. However, any system that supports general type hierarchies is losing in performance and we are trying not to sacrifice that. I have not seen a good approach (going back to object-oriented dbmss of 90s) that supported type hierarchies in a performant way. To set the expectations correctly though, even if we support that, it won't happen quickly. I think encoding the labels somehow as a List[String/INT] property is one manual approach of course but I understand that's far from ideal. If you have other manual approaches, it would be good to share with others in the community. @mnsRG: When you say "recommended way to load RDF data", I assume you mean Kuzu's RDFGraphs feature. |
Beta Was this translation helpful? Give feedback.
-
|
@semihsalihoglu-uw yup I am referring to this way to load RDF. The current approach seems as a nice workaround to load 'any' rdf dataset (regardless of its implicit schema), but when it is important to express the structure of the data to the schema itself, then another way is needed. But this goes a bit off the multi-label topic (probably an alternative for loading RDF would be an intermediate step to read the implicit schema, write it to kuzudb and then ingest the data - plus find a way to deal with updates) |
Beta Was this translation helpful? Give feedback.
-
|
OK, RDFGraphs is one specific way of mapping RDF triples, so yes if you want the schema to expressed as labels, this won't work. I am not sure if we as a system can support multiple mappings from RDF or how easy that would be. We have to accumulate some feedback and suggestion on RDFGraphs before we make any decisions on this. |
Beta Was this translation helpful? Give feedback.
-
|
maybe of interest, this is a topic not that uncommon, as one might think. |
Beta Was this translation helpful? Give feedback.
-
|
just another idea on this, terminusdb (different graph db type: git like full history/immutable, no cypher) also support multiple labels in a somehow different way and is even with this feature-set faster than neo4j in several queries...
Creating hierarchies at https://terminusdb.com/blog/graph-databases-terminusdb-vs-neo4j/ |
Beta Was this translation helpful? Give feedback.
-
|
thanks for bring up the details! My point is only, that |
Beta Was this translation helpful? Give feedback.
-
|
Yes, this is the type of class hierarchy that makes more sense to me. But I am not aware of a principled system implementation that does this well. Systems can be faster or slower than each other for a variety of reasons, so I would be suspicious that terminusdb is faster for any reasons related to this class hierarchy implementation. I would be curious to learn of a principled design for supporting such general class hierarchies. Not saying there isn't but having perused oo dbms and gdbms literature for years, I'm unaware of one. If you have a many layered class hierarchy and you ask join queries with types that are in the higher levels of the hierarchy, you are forcing the system to consider many combinations of lower classes, which are always hard to evaluate. |
Beta Was this translation helpful? Give feedback.
-
|
TerminusDB is fast mainly because it's implemented in Rust (Neo4j is all Java), and it also uses recursive operations on data structures really well in Rust. They've long been fans of the Prolog language and wanted to implement a solution that allows for efficient traversal of class hierarchies, hence their emphasis on RDF and OWL. |
Beta Was this translation helpful? Give feedback.
-
|
What I'm curious about is the overlap between property graph models and RDF out in the literature or the wild. There aren't too many examples of tools I find that are trying to do what Kùzu is - blending the benefits of both worlds. That being said, there definitely are a lot of use cases in the real world for multilabel graphs and I agree with @hpvd's point there :). |
Beta Was this translation helpful? Give feedback.
-
|
Regarding terminus db, they shared this imho interesting internal: |
Beta Was this translation helpful? Give feedback.
-
|
what do you think of converting this to an issue, type feature request? |
Beta Was this translation helpful? Give feedback.
-
|
new related issue created: |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for opening the issue. Let's keep it around and see if there is anything clever we can come up with at some point. |
Beta Was this translation helpful? Give feedback.
-
|
if you are interested on this, plz comment /ad +1 in #3117 |
Beta Was this translation helpful? Give feedback.
-
Hello, I was wondering if there is any plan to support multiple labels per Node. I am mainly dealing with RDF (and RDF-star) data and many nodes come with inferred types (eg a 'Desk' is also a 'Furniture'). Therefore, I would like to include all of them (not necessarily by keeping their hierarchies - flat is fine, which is how Neo4j handles it more or less). Then, when I ask for all the 'Furniture' in my database, I also get the 'Desk' nodes.
Btw, I am not using the recommended way to load RDF data, because it is important that the schema of the database reflects its actual structure, as this provides the context for an LLM that translates natural language to cypher queries.
Thanks!
Beta Was this translation helpful? Give feedback.
All reactions