There are various string-valued informational messages in serverStatus which may or may not be present depending on whether the condition they are informing about is present. If this presence of this field changes between samples this is detected by ftdc as a change in schema, causing a new chunk to be started. If this varies rapidly, e.g. every few seconds, then chunks are very short, causing a large decrease in compression efficiency, and therefore a decrease in ftdc retention. This is especially problematic because ftdc is likely to be important in diagnosing the issue.
Since string-valued fields don't contribute to ftdc, there's no reason that a change in the presence of a string valued field should be considered a schema change.