Claude,
These may help:
I have been thinking about an interface that is more oriented to the
storage than the full DatasetGraph.
StorageRDF breaks down all the operations into those on the default
graph and those on named graphs. For just a graph, simply ignore the
named graph operations.
https://github.com/afs/AFS-Dev/blob/master/src/main/java/projects/dsg2/storage/StorageRDF.java
There is an adapter to the DatasetGraph hierarchy (which is needed for
SPARQL):
https://github.com/afs/AFS-Dev/blob/master/src/main/java/projects/dsg2/DatasetGraphStorage.java
If you want to only use existing classes, DatasetGraphTriplesQuads is
the place to start - used by TIM and TDB - yuo can implement without
needing quads/named graphs. Again, simply ignore (throw
UnsupportedOperationException for the named graph calls).
Going the graph route could lead to rework later on for any kind of
performance issues because find(S,P,O) is so narrow and precludes union
default graph except by brute force. DatasetGraph work with the SPARQL
execution engine.
We still need to discuss how best to approach developing work - it
should not get sucked up by the release cycle.
Andy
On 26/10/16 19:21, Claude Warren wrote:
My plan is to start with a Graph implementation. We expect to write 3
tables: SPO, POS, OPS (I think). Currently we don't have an easy way to
handle find( ANY, ANY, ANY) so I suspect we will just start with permitting
a column scan on Cassandra.
I have not looked at DynamoDB but as I recall there are significant
differences under the hood.
I expect that we will move on to a custom model or query engine to get the
best performance but that is not what we are planning for the first cut.
I am still waiting for management approval to do this at work ....
sometimes it takes longer to get the paperwork done than it does to design
the thing.
Claude
On Mon, Oct 17, 2016 at 6:39 PM, Paul Houle <[email protected]>
wrote:
I like DynamoDB as a target for this sort of thing. There are many
tasks which are small-scale yet critical where it would otherwise be
hard to provide a distributed and reliable database. Put that together
with Lambda, which does the same for computation, and you are cooking
with gas.
I wrote a 1-1 translation of DynamoDB documents to RDF that I use
throughout an application; the code is DynamoDB idiomatic in every way,
just the application reads and writes (a constrained set of) RDF
documents.
Right now I dump the documents from the DynamoDB system into a triple
store when I want a panoptic view, but with a distributed graph like
that would mean being able to run SPARQL queries against DynamoDB
directly.
There are many products in the same family as Cassandra and DynamoDB and
it would be good to think through the math so we can approach them all
in a similar way.
--
Paul Houle
[email protected]
On Mon, Oct 17, 2016, at 12:31 PM, A. Soroka wrote:
Yep,
http://iswc2011.semanticweb.org/fileadmin/iswc/Papers/
Workshops/SSWS/Ladwig-et-all-SSWS2011.pdf
indicates that they are indexing by subject. As someone who has
implemented LDP, that is definitely the approach that makes sense there.
---
A. Soroka
The University of Virginia Library
On Oct 17, 2016, at 12:20 PM, Andy Seaborne <[email protected]> wrote:
IIRC It stores CBDs indexed by subject so it is the "other" model to
Rya. Better for LDP (??).
Andy
On 17/10/16 15:41, A. Soroka wrote:
There's also:
https://github.com/cumulusrdf/cumulusrdf
in a similar vein (RDF over Cassandra). Not sure what kind of
particular uses it expects to support.
---
A. Soroka
The University of Virginia Library
On Oct 17, 2016, at 7:02 AM, Andy Seaborne <[email protected]> wrote:
Hi Claude,
There is certainly interest from me.
What the best thing to do depends on various factors. By putting it
in extras I presume you mean it gets added to the release? That is not the
only way forward.
An important aspect of Apache is "Community over code" - will there
be a community around this code? Is that community the same, or
significant overlap, as the Jena community?
There are various reasons for wanting RDF over a column store -
which use cases are the most important for this work?
They lead to different ways of using Cassandra. For example,
Rya(incubating) uses Accumulo tables as indexes, and partial scans of the
table is streaming. Other systems try to use the columns for properties,
possibly more useful for LDP style than SPARQL.
Andy
On 15/10/16 18:38, Claude Warren wrote:
Howdy,
We have a project at work that is implementing Jena Graph on
Cassandra. I
am wondering if there is enough interest here to accept it as a
contribution. I was thinking that it might fit in the Extras
category.
I can not promise release of the code yet as I have to present it
to our
internal Intellectual Property group first.
Thoughts?
Claude
