If you are interested in the code behind this, I wrote an overview last month on the functionality and links to the different code that backs the improvements they talk about: http://hydronitrogen.com/spark-220-cost-based-optimizer-expl...

There's a fair amount of overlap, but where the databricks article explains the techniques with charts and high level explanations, I go over the code instead.

On this topic, I really like the Join Order Benchmark paper: http://www.vldb.org/pvldb/vol9/p204-leis.pdf

It basically shows that most cost-based optimizers are pretty bad at cardinality estimation, which compounds when queries use more joins.

Still catching up with postgres which added multivariate column statistics in 9.6 :)

Not that this isn't a great development in itself...

What's cool about these statistics-based approaches is that you mostly don't even need fully up-to-date statistics, just overall decent stats, unless you have an insane amount of churn. Meaning - you can get query speedup without insertion overhead: you choose to take that overhead any time you want using ANALYZE.

Very neat stuff from the databricks team!

For non latency sensitive queries you can also run dynamic sampling.