Looks like a queue worker issue. From eng: "We're moving about 5000 events per second now. Definitely shrinking now, and memory usage is dropping too."

Should be fixed.