2021-01-06 12:01:44 +01:00
|
|
|
An evaluation of the IPFS paper
|
|
|
|
|
===============================
|
|
|
|
|
|
|
|
|
|
2021-01-06 Mark Spanbroek
|
|
|
|
|
|
|
|
|
|
https://ipfs.io/ipfs/QmR7GSQM93Cx5eAg6a6yRzNde1FQv7uL6X1o4k7zrJa3LX/ipfs.draft3.pdf
|
|
|
|
|
|
|
|
|
|
Goal of this evaluation is to find things to adopt or avoid while designing
|
|
|
|
|
Dagger. It is not meant to be a criticism of IPFS.
|
|
|
|
|
|
|
|
|
|
#### Pros:
|
|
|
|
|
|
|
|
|
|
+ IPFS is designed by simplifying, evolving, and connecting proven techniques
|
|
|
|
|
(§3)
|
|
|
|
|
+ Consists of a stack of separately described sub-protocols (§3)
|
|
|
|
|
+ Uses Coral DSHT to favor data that is nearby, reducing latency of lookup
|
|
|
|
|
(§2.1.2)
|
|
|
|
|
+ Uses proof-of-work in S/Kademlia to discourage Sybil attacks (§2.1.3)
|
|
|
|
|
+ Favors self-describing values such as multihash (§3.1) and multiaddr (§3.2.1)
|
|
|
|
|
+ BitSwap protocol for exchanging blocks supports multiple strategies (§3.4.2)
|
|
|
|
|
+ Uses content addressing (§3.5)
|
|
|
|
|
+ The Merkle DAG is simple, yet allows constucting filesystems,
|
|
|
|
|
key-value stores, databases, messaging system, etc.. (§3.5)
|
|
|
|
|
|
|
|
|
|
#### Cons:
|
|
|
|
|
|
|
|
|
|
- Kademlia prefers long-lived nodes, which is not ideal for mobile enviroments
|
|
|
|
|
(§2.1.1)
|
|
|
|
|
- BitSwap falls just short of introducing a currency with micro payments,
|
|
|
|
|
necessitating additional work for nodes to find blocks to barter with (§3.4)
|
|
|
|
|
- Merkle DAGs and Paths (§3.5.1) feel like an unnecessary complication at the
|
|
|
|
|
protocol level; they could have been implemented on top of a system consisting
|
|
|
|
|
solely of content addressable chunks
|
|
|
|
|
- Object pinning (§3.5.3) inevitably leads to centralized gateways to IPFS, such
|
|
|
|
|
as Infura and Pinata
|
|
|
|
|
- There are no self-describing multiformats for encryption and signing (§3.5.5)
|
|
|
|
|
- IPFS uses variable size blocks instead of fixed-size chunks (§3.6)
|
|
|
|
|
- Supporting version control directly in IPFS feels like an unnecessary
|
|
|
|
|
complication (§3.6.5)
|
2021-01-06 12:04:46 +01:00
|
|
|
- Handles mutable state by hacking it into the routing DHT (§3.7.1), instead of
|
|
|
|
|
using a separate mutable naming system like ENS
|
