At the beginning of chunk migration, we do a force refresh of the collection metadata, this will eventually call CatalogCache::_getCollectionRoutingInfoAt, which will then call CatalogCache::getDatabase. getDatabase, however will always try to call ShardRegistry::getShard of what it thinks is the current primary database. If the shard has already been removed, then it error out with ShardNotFound.

To get around this issue, send a flushRouterConfig to the affected shard.