Right now we ignore records headers when computing collection data size (mongo/db/namespace_details.h):

DiskLoc deletedList[Buckets];
// ofs 168 (8 byte aligned)
struct Stats {
    // datasize and nrecords MUST Be adjacent code assumes!
    long long datasize; // this includes padding, but not record headers
    long long nrecords;
} stats;

However, the value includes padding and also includes the alignment that we use in our storage engine.

When the number of records hits 500,000,000 the headers overhead is 7+ GB, which is perceived as a lost disk space. I suggest to include record headers size into the coll.stats().size metric, or to introduce an additional sizeWithHeaders metric to avoid confusion.