Inspired by this, can anyone explain why distributed protocols more often opt for centralized consensus algorithms like Raft, instead of decentralized schemes like Chord or Kademlia? In all cases, the underlying data structure is a shared key/value store. Intuitively, the p2p approach feels more robust, since each node only needs to worry about itself, and every node is the same. So why add the coordinator node? Is it still the right choice in $current_year, even after so many hours of development invested into strong p2p consensus protocols like libp2p powering Ethereum (currently ~11k nodes btw, not actually that big – and many operated by small number of entities)?
Obviously you don't need consensus protocols if you are not trying to build consensus...
It's like asking why we need filesystems and network cards if all we're trying to do is show lights on screens. Obviously not all patterns of lights are as easy to produce on computers.
I did a bit more research last night, and discovered that Bitfinex actually does something like this internally (anyone know if this is up to date?) [0] — they built a service discovery mesh by storing arbitrary data on a DHT implementing BEP44 (using webtorrent/bittorrent-dht [1]).
This seems pretty cool to me, and IMO any modern distributed system should consider running decentralized protocols to benefit from their robustness properties. Deploying a node to a decentralized protocol requires no coordination or orchestration, aside from it simply joining the network. Scaling a service is as simple as joining a node to the network and announcing its availability as an implementation of that service.
At first glance, this looks like a competitive advantage, because it decouples the operational and maintenance costs of the network from its size.
So I'm wondering if there is a consistent tradeoff in exchange for this robustness — are decentralized applications more complex to implement but simpler to operate? Is latency of decentralized protocols (e.g. average number of hops to lookup item in a DHT) untenably higher than that of distributed protocols (e.g. one hop once to get instructions from coordinator, then one hop to lookup item in distributed KV)? Does a central coordinator eliminate some kind of principle agent problem, resulting in e.g. a more balanced usage of the hashing keyspace?
Decentralization emerged because distributed solutions fail in untrusted environments — but this doesn't mean that decentralized solutions fail in trusted environments. So why not consider more decentralized protocols to scale internal systems?