The function dict_stats_fetch_from_ps() first looks up the InnoDB persistent statistics tables in a suboptimal way, acquiring and releasing a shared dict_sys.latch multiple times, and then proceeds to acquire an exclusive dict_sys.latch for the duration of the call to que_eval_sql().

I think that this needs to be optimized as follows:

• Try to acquire a shared dict_sys.latch only once in the case that both statistics tables are cached in dict_sys and that a metadata lock (MDL) can be acquired without waiting.
• If one of the statistics tables is not cached or if we have to wait for a MDL, we may fall back to something like the existing logic.
• For the que_eval_sql() call, pass table handles with pars_info_bind_id(). In this way, there should be no need to hold any dict_sys.latch during the execution of the SQL.