ldbc / lsqb Goto Github PK

A dir structure with the scripts in the directories of their respective DBs could be better. For example, neo/load.sh, ddb/load.sh, etc.

Instead of loading everything from under data/social-network-preprocessed

... just vertex/edge tables.

Add a larger test data set (0.003) which better highlights issues due to it exhibiting more corner cases.

Data: The .nt file created by scripts/convert-to-ntriples.sh could work well as Dgraph's input.

The queries need to be rewritten in DQL.

The Virtuoso implementation produces the correct results but it could use some improvements:

• improved load times
• optimizations for certain queries (e.g. Q2, Q3)
• a Python driver for easier result collection

It currently returns with no results on SF100. Need to investigate whether this is a crash or not.

SFexample works but SF0.1 segfaults.

pre-load, load, and post-load should be possible to invoke through just one script init-and-load.sh.

$subject

The data set makes use of multiple labels (e.g. Message = Comment | Post). This is currently blocked on a missing feature in RedisGraph.

RedisGraph is going to include this feature in April (probably earlier in preview releases): RedisGraph/RedisGraph#13

This requires minor changes in the loading script for the following labels ("supertypes"):

• Message
• Place
• Organisation

Q3 can be expressed using LEFT JOIN + WHERE ... IS NULL.
Track duckdb/duckdb#1348 for further instructions.

See RedisGraph/redisgraph-bulk-loader#67

Paper: https://arxiv.org/abs/2001.09078
Repository: https://github.com/karmaresearch/trident/

https://github.com/gh-rdf3x/gh-rdf3x
Build with:

make CXXFLAGS='-std=c++03'

I tried adding indices on the Postgres implementation but this only made the queries slower. A (much) more thorough investigation is needed.

Fails to load the most recently generated .graph files. No idea why. See issue at RapidsAtHKUST/RapidMatch#2

Queries seem to run well past time the timeout.

Many of the benchmarked systems support intra-query parallelism (DuckDB, Umbra), so it would be interesting to benchmark them using different #vCPU cores.

See non-repeatable nodes/edges.

Currently, if the query crashes, the host process crashes. The "client" should run in a separate process or we should take more care for handling the exceptions.

Neo4j 4.3 should be out in the coming weeks

For isomorphic and homomorphic matching.

• Add implementation
• Dockerize it: https://hub.docker.com/r/materialize/materialized
• Add inserts/deletes using Datagen

https://stackoverflow.com/questions/25775266/how-to-keep-docker-container-running-after-starting-services

Our current largest data set is SF100. This limitation happens mainly because the snb-export-only-ids-merged-fk.sql and snb-export-only-ids-projected-fk.sql scripts in the data converter repository crash with DuckDB on SF1000. We could try another database for conversion or wait for the cause of the crash (which might be simple memory exhaustion but might be something else) to be resolved :).

The FK columns have clunky names such as hasTag_Tag instead of (e.g.) TagId. It would be worth revising these.

i.e. Q5

Add Virtuoso configuration file and put it in place in the pre-load.sh script.