This is a neat idea, but the unfortunate reality is that every link in an "integration chain" is almost guaranteed to be lossy. The author notes an obvious version of this (what if one node doesn't support a field that the other nodes do) but there are countless other types of loss in each link - structure (i.e. document data types), precision, limits (max field size, max number), other "exotic" datatypes (geography, binary blobs). Even mutability can cause lossiness - some of the fields in these systems are write-once, and even if creates and updates can be propagated, delete propagation is a Wild West in SaaS applications.
If you're old enough to remember what happens when you make a repeated analog copy of something (fax machines, copy machines, VHS cassettes, etc) you know what happens with this kind of data flow topology. We advise our customers to do the /exact/ opposite of this - follow as few links as possible for any piece of data and move towards hub-and-spoke, ideally with the hub as a highly configurable, low loss node - in my world, this is a data warehouse or data lake, but it might also mean an iPaaS or similar.