I use LS with large traffic of 1000+ requests per second. After running for 3-7 days the CPU idle time reduces very noticeably by 20-30% under the same traffic load. Performing graceful restart of LS instantly increases CPU idle time back to where it should be. There are obviously some inefficiencies that over the course of hundreds of millions of requests make LS hog much more CPU than it should. I got into a habit of performing weekly graceful restarts. I hope this can be identified and addressed in future releases.