dispatch_to_series is very slow

[RunningToTheEdgeOfTheWorld]

Since pandas 0.24.0 I found there is a function in ops.py named dispatch_to_series.
But dispatch_to_series is very slow, that makes my code slower.
when I use pandas==0.23.0, it cost about 100 seconds, but when I use pandas==0.24.* It cost 583 seconds.

This is my profile result.

A and I found pandas become more slower because you call isinstance function too many times
link to history merge : [PERF] use numexpr in dispatch_to_series #22284

[jreback]

show a minimal example with code

[jbrockmendel]

Duplicate of #25883, #24990.

@RunningToTheEdgeOfTheWorld dispatch_to_series was implemented because of a correctness/performance tradeoff, and correctness wins every time. There has been some discussion about how to ameliorate the performance hit, but I'm not aware of any concrete steps. If you have any ideas, PRs are welcome.

[TomAugspurger]

Closing in favor of #24990