> It did --- see the Gather node.
But "workers launched: 1"...
To my opinion, such a dramatic difference cannot be explained with avoiding
parallelism, the query was just stuck in a very inefficient plan (even
though almost all source data is read from cache).
Additionally, I think author can try CREATE STATISTICS on the bunch of
columns used in join. Very low rows estimate for this join may come from
multiplying selectivities for each column assuming they are independent.
Vlad