The problem is that getting metrics like this is per file and so very network chatty and slow.
If you have a rapidly changing index then it can take a long time to generate the metrics, since we lose out on the benefits of caching the info then.
For this type of data, I think we would prefer stale data to this limitation. Perhaps we should be doing time based caching instead. Fairly trappy to have this work okay on fairly static indexes and then cause a major performance issue as you move towards NRT settings.