Posting this early for feasibility feedback. Main use for this would be https://github.com/ipfs/rainbow/issues/125 and fetching from big providers that currently pay high price for supporting bitswap.
TLDR
Instead of making HTTP retrieval an alternative to bitswap sessions, try a simpler approach that makes both systems improve each other in a backward-compatible way instead.
This is a proposal to enhance Bitswap system with ability to do HTTP retrieval of blocks over HTTP instead of libp2p. The opportunistic HTTP retrieval would happen when a peer announces support for it via /http* multiaddr.
Why
Serving data over HTTP is less expensive for multiple reasons (caching, billing), and if IPFS nodes were able to leverage HTTP retrieval ("bitswap over HTTP") instead of "bitswap over libp2p", that would be net positive for the ecosystem.
we have a swarm that speaks bitswap 1.2.0 already, and peers are able to ask if peer has data over bitswap, without retrieving it
/http* multiaddrs are ignored by existing clients, and shoudl be future-proof thanks to libp2p HTTP spec (spec itself being backward-compatible with plain HTTP trustless gateways being mounted on /ipfs).
Together, we have building blocks to implement and ship an incremental improvement to the public IPFS swarm.
Why doing this in bitswap client?
Evolution over revolution.
Bitswap is how IPFS swarms exchange data. Asynchronous fetch of a block over HTTP is a very simple protocol that maps closely to existing bitswap abstractions
Surgical improvement: we don't have to refactor how content retrieval system works on the high level, limiting the risk of unexpected bugs / regressions, including counter-intuitive ones (example)
Ease of adoption
If we have HTTP retrieval support in bitswap client, and enabled by default in Kubo and Rainbow, that does not require storage providers to invest a lot to benefit from it. All they need to do, is to announce additional multiaddr to rollout HTTP retrieval endpoints in backward-and-future-compatible way.
Reusing bitswap and multiaddrs maximizes interop, enables opportunistic HTTP retrieval everywhere, including private swarms that don't use DHT nor IPNI
Ease of implementation
I don't feel we can afford to rewrite block exchange and routing systems right now, nor that it could happen in reasonable time without risk of regressions.
How
Details TBD, posting this to gather feasibility feedback, broad strokes idea is:
I strongly believe this should be done client-side (client making decision to use HTTP) and not require change in behavior of existing bitswap servers or protocols.
High level mechanics:
When bitswap client receives "i have" response, don't ask for block over libp2p, but instead fetch it over HTTP from URL based on the /http* multiaddr (bitswap over HTTP), and notify just like when it would be fetched over libp2p.
(TBD, feels like good way to minimize risk even further) If HTTP retrieval fromt the peer fails, retry with regular bitswap over libp2p
Initially it would be disabled by default, and we would enable it only in rainbow. Once we are happy with it, we would make it enabled by default in boxo/bitswap, and other GO-based implementations like Kubo would get it as well.
Integration gotchas
Chicken and the egg problem: we don't have bitswap peersids that also announce /http* multiaddrs.
How would ecosystem rollout "bitswap with oportunistic HTTP retrieval"?
DHT: storage providers running IPFS Cluster / Kubo could announce HTTP retrieval capability via opt-in configuration options
Existing AppendAnnounce could be where providers add /http* multiaddr pointing at their trustless gateway, and bitswap clients could pick this up via peer routing. This is one-line config change providers do to signal they prefer clients to use HTTP retrieval
IPNI:
main case is web3.storage, which announces with two different peerids for bitswap and http:
bitswap peer QmQzqxhK82kAmKvARFZSkUVS6fo9sySaiogAnx5EnZ6ZmC listening at /dns4/elastic.dag.house/tcp/443/wss
fake peerid QmUA9D3H7HeCYsirB3KmPSvZh3dNXMZas6Lwgr4fv1HTTp which has multiaddr /dns4/dag.w3s.link/tcp/443/https
Perhaps this is an opportunity to clean this up and remove the need for bogus HTTp peerids? At minimum, we could ask web3.storage to start announcing /dns4/dag.w3s.link/tcp/443/https under bitswap peer QmQzqxhK82kAmKvARFZSkUVS6fo9sySaiogAnx5EnZ6ZmC, making their announcements compatible with this proposal.
TLDR
Instead of making HTTP retrieval an alternative to bitswap sessions, try a simpler approach that makes both systems improve each other in a backward-compatible way instead.
This is a proposal to enhance Bitswap system with ability to do HTTP retrieval of blocks over HTTP instead of libp2p. The opportunistic HTTP retrieval would happen when a peer announces support for it via
/http*multiaddr.Why
Serving data over HTTP is less expensive for multiple reasons (caching, billing), and if IPFS nodes were able to leverage HTTP retrieval ("bitswap over HTTP") instead of "bitswap over libp2p", that would be net positive for the ecosystem.
On the HTTP side:
application/vnd.ipld.rawOn the libp2p side:
/http*multiaddrs are ignored by existing clients, and shoudl be future-proof thanks to libp2p HTTP spec (spec itself being backward-compatible with plain HTTP trustless gateways being mounted on/ipfs).Together, we have building blocks to implement and ship an incremental improvement to the public IPFS swarm.
Why doing this in bitswap client?
Evolution over revolution.
Ease of adoption
Ease of implementation
How
Details TBD, posting this to gather feasibility feedback, broad strokes idea is:
/http*multiaddr (bitswap over HTTP), and notify just like when it would be fetched over libp2p.boxo/bitswap/options.go:boxo/bitswap, and other GO-based implementations like Kubo would get it as well.Integration gotchas
Chicken and the egg problem: we don't have bitswap peersids that also announce
/http*multiaddrs.How would ecosystem rollout "bitswap with oportunistic HTTP retrieval"?
AppendAnnouncecould be where providers add/http*multiaddr pointing at their trustless gateway, and bitswap clients could pick this up via peer routing. This is one-line config change providers do to signal they prefer clients to use HTTP retrievalQmQzqxhK82kAmKvARFZSkUVS6fo9sySaiogAnx5EnZ6ZmClistening at/dns4/elastic.dag.house/tcp/443/wssQmUA9D3H7HeCYsirB3KmPSvZh3dNXMZas6Lwgr4fv1HTTpwhich has multiaddr/dns4/dag.w3s.link/tcp/443/httpsHTTppeerids? At minimum, we could ask web3.storage to start announcing/dns4/dag.w3s.link/tcp/443/httpsunder bitswap peerQmQzqxhK82kAmKvARFZSkUVS6fo9sySaiogAnx5EnZ6ZmC, making their announcements compatible with this proposal.Feedback welcome.