This is equivalent to the get_client call I introduced below if getattr(thread_state, "key", False) is True so there is no need to have this additional indirection.

By having only a single call, we're changing precedence slightly s.t. a class providing the __dask_scheduler__ attribute wins but I believe this "more correct" considering it is passed explicitly. I doubt this has any relevance.

i.e. if we're not requesting anything specific but are executing the default we will always try to get a global distributed client before checking the collections default.
The precedence there is

1. Worker client
2. Current client

there is effectively no performance penalty in case there is no distributed worker or client other than raise+except