The current resize policy for the TicketHolder is kGradual, and that's what we use for Execution Control. This means that resizing it works by gradually increasing the number of tickets or taking tickets one by one. This means that if all tickets are taken this operation will block until tickets are released.

I think this is problematic, for instance if the ThroughputProber tries to ProbeDown but all tickets become taken it could block for a long time, never probing up until some tickets are released (this could potentially explain behaviours similar to SERVER-87365 )