The oplog truncation markers implementation uses an eventually correct algorithm that assumes the server is alive long enough and receiving enough writes for the oplog to fully roll over, at which point all truncations can accurately update size and count information.

If the server is restarted frequently, like what happens in auto-scaling, the server can get stuck in a place where it is not alive long enough to accurately account for size and count changes. This can have the effect of making truncations ineffective and not delete enough data over time.