Currently we write offsets for every field. It could take 1, 2 or 4 bytes depending on the data length of the object. Now suppose we have the following class:
Fields are always sorted in alphabetical order, so we will write them as follows [a, b, c, d, e], and their offsets would always be [0, 5, 14, X, X+5]. As you see, instead of writing 5 offsets, it is enough to write only one offset of a field, which follows another variable-length field. The rest offsets could be saved to metadata.