As discussed earlier, generally speaking the rsync algorithm is ineffective for compressed data, unless the new data is only (at least predominantly) appended to the existing content. The --rsync option for gzip is not widely used; if zsync succeeded widely then --rsync might become widespread, but that is a distant prospect.
So zsync could work just for uncompressed and --rsync files, but this would limit its use, given that so much existing content is distributed compressed. There is no fundamental reason why we cannot work on compressed files, but we have to look inside, at the uncompressed data. If we calculate the block checksums on the uncompressed data stream, store these checksums on the server, and apply the rolling checksum approach on the uncompressed data on the client side also, then the basic algorithm is effective.
Having looked at the checksums for the uncompressed data, the normal rsync algorithm tells us which blocks (from the uncompressed stream) we have, and which are needed. Next we must get the remaining blocks from the server. Unfortunately, HTTP Range headers do not allow us to select a given spot in a compressed data stream, nor would it be desirable from the point of view of the server to implement such a feature. So we must have a mechanism for retrieving blocks out of the compressed data stream.