Perhaps there's an event-driven best-effort microservice providing an eventually-consistent tally, and it's missing a few events.
The tweet graph database has a persistent edge count cache but it's not designed to handle posts or deletions any faster than a human can produce them.