We saw a customer case where hazard pointer array maximum length peaked at 2.2 million. The application was using a modest number of session and cursor handles.

The excessive usage appears to be correlated with checkpoints (see below image).

We should review the hazard pointer acquisition/release code associated with creating checkpoints, and see if there is any way it could acquire many hazard pointers at the same time. Interesting the number of hazard pointers does seem to decrease - so this doesn't seem to be a leak.