Twitter’s backend systems run on a distributed Apache Mesos cluster of decomposed services. There are often cases where data sets will need to resynchronise between instances and data centers.
While I realise that this discrepancy took place over an unusually long period of time, I’ve already seen evidence in the calls you listed earlier in the thread that some instances were reporting different values to others, so I think this is extremely likely to be the background here. You should also take into account that as we continue to clean up spammy accounts or users are suspended, there will occasionally be a sharper drop than unusual in follower or following numbers, and those kind of larger changes may take longer to be reflected on the API than at other times.
