Index building performance fulltext vs schema

[dsolow]

I'm trying to index a unique string property on ~170M nodes. I've noticed that building a native (B-tree) index takes a very long time (several hours), but building a fulltext (Lucene) index takes around 40 minutes. For my use case I'd prefer the schema index since I don't need full-text search, but I'm leaning towards using the fulltext index because of the slow build schema index time.

Is there any way to improve the index build for B-tree indices? I don't need any kind of transactional integrity for building this index, since it's running right after a bulk import and before any read queries arrive.

Thanks

[dana_canzano]

what version of Neo4j? There were improvements to index population performances in recent patches so you will want to make sure you are on a current release? 3.5.6 for example

[dsolow]

Thanks for your reply. I'm using 3.5.5 -- I'll give 3.5.6 a shot tomorrow

[dsolow]

3.5.6 does seem significantly faster. It looks like it will finish in around 2.5 hours, which isn't great, but I can probably work with it for now

I also saw it mentioned here that the Dorg.neo4j.kernel.impl.index.schema.GenericNativeIndexPopulator.blockBasedPopulation=true setting might speed things up. Do you think that's worth trying?