The insertion rates were based on the elasticsearch backend, with multiple shards/nodes (using the already flattened image signatures). Don't remember specifics unfortunately. I never tried it with more than 5 lightness buckets, sorry!