3

I am planning to use Janusgraph for building graph of different uses our team handles and I see that janus graph has option to use BigTable or Cassandra as storage backend. I am looking for any recommendation on which backend is more optimal/performant ( I am mainly talking about gremlin query performance on 2 hop neighbor of a node ) with JanusGraph.

I understand that performance is pretty subjective and varies based on datasize/graph connectivity and use case so best approach will be to try out myself, which I am planning to do. But has anyone else has done similar performance comparison ? Is there any general recommendation about storage backend here ?

Vishal
  • 635
  • 4
  • 9

1 Answers1

1

You're right in that performance is both:

  • subjective
  • depends largely on data size

I can tell you that I have done this exercise as well. To that end, I think it's important to share this comparison from DB-Engines.com.

In terms of performance, the biggest thing I'd be looking at is how each handles consistency. As a general rule, databases which enforce stronger levels of consistency typically have to sacrifice performance.

  • BigTable == strong-consistent
  • Cassandra == eventually consistent

Other factors worth considering, are the fact that BigTable limits you to Google Cloud (GCP). And if you don't want to lose performance over the network, you'll also need to pay for more (Janus) instances on GCP for data locality.

In terms of raw DB-Engine "score," Cassandra is currently at 114.112, while BigTable is at a paltry 3.582. These scores will change month-to-month, but in general this signifies that Cassandra has a much stronger community around it. Similarly, Cassandra has 18182 questions on this site, while BigTable only has 449. Bottom line, is that it'll be much easier to get support and answers to questions.

Just based on the underlying strength of the community, Cassandra is the better option here.

Having supported JanusGraph on Cassandra for the last few years, I can tell you that overall it's been solid. The difficulties tend to come into play with bulk data loading. But outside of that, things seem to run pretty well.

Aaron
  • 55,518
  • 11
  • 116
  • 132
  • Thanks for response. Can you provide a bit more details on "bulk data loading" problem you mentioned ? Are you talking about issue with high write rate or doing bulk read/query operations ( like OLAP usecase ) ? – Vishal Jun 10 '21 at 19:48
  • @Vishal Specifically, JanusGraph's Spark connector has proven to be problematic for high write throughput. – Aaron Jun 10 '21 at 19:55