Directory Service Tracking Issue

[nothingmuch]

See #775 for discussion

Sub bullets indicate spinoffs of parent, which are directly dependent. Otherwise tasks mostly don't depend on each other.
Most of these do depend on #1232 however, as the ultimate goal here is to unify the directory and relay services.

• refactor: introduce payjoin_directory::Service #914
• Introduce metrics Collection via prometheus for payjoin directory #848
  • tower_http::metrics for simple in flight reqs metric, indicates ~ number of active sessions proxied by concurrent GET requests that are blocking on data
• Configure directory with config & clap #927
  • Ensure key rotation is accounted for in server and client implementations #445
  • config reloading? not sure how valuable this would be with acme and automatic key rotation
• Replace redis with file based mailbox storage #915
  • DB should also be a tower::Service, using tower::{timeout,limit::rate}
• tokio-rustls-acme
  • configure rustls for PQ
• Fold ohttp-relay into rust-payjoin monorepo #1215
• merge ohttp-relay & payjoin-directory into a Unified payjoin service #1232
  • deny access to own directory from the relay
  • payjoin-service should use axum to simplify routes, split app into sub-apps that all implement the Service trait (e.g. the OHTTP target resources) (supersedes Remove hyper and migrate to axum. #1101)
  • backwards compatibility routes can be done using the tower_http::follow_redirect middleware instead of being aliased as they are today
• use tokio-listener, --listen-addr instead of --port or something like that
  • that makes it more flexible if systemd does the listening for e.g. a nixos module. no more code complexity for us but more flexibility to override the nixos module... but equally systemd unit could just start it and it could bind on something
• Prometheus metrics with axum middleware in payjoin-service #1284
• handle HTTP 2
• maybe factor out a tower_http OHTTP gateway middleware
• remove boilerplate with tower middleware
  • tower_http::trace, remove boilerplate tracing and only leave domain relevant logging
  • tower_http::limit, remove our custom body size limiter
  • tower::limit::rate, POST reqs should be limited to 25/s to ensure all mailboxes have a TTL of minimum 24hrs given default capacity limit
  • tokio_http::cors, removes tiny bit of boilerplate
• Set up GitHub Actions for Docker image builds using flake.nix #1286
• Shift integration tests to payjoin-service once ready
• Deprecate payjoin-directory binary and library
• We also need to beef up our directory service docs. E.g what hardware resources you need to run one, instructions to run, explaining how to configure the directory.

---

Far future:

• QUIC
• arti_client for Tor for bootstrapping #766 (hidden service for directories & tor bridge for relays), kinda tricky without https://gitlab.torproject.org/tpo/core/arti/-/issues/1932 see https://docs.rs/arti-client/0.33.0/arti_client/index.html#-warnings-
• multihop OHTTP Gateway opt-in makes onion routing over relays possible #582
  • • /{id} and /{authority} paths are distinguishable, ShortIDs are 13 bech32 chars, authorities must have : or .

[DanGould]

Now that ohttp-relay is up to date with 1.85.0, it could be brought into this repository as a monorepo item. However, it depends on ureq for proxy which'd be nice to replace with reqwest if possible and it depends on nginx for tls termination which would really be better to replace with whatever native ACME solution you cook up for the directory.

In either case I don't think merging the repos affects payjoin crate's 1.0 stability now so I'm just noting these things for when the time is right.

[zealsham]

Now that ohttp-relay is up to date with 1.85.0, it could be brought into this repository as a monorepo item. However, it depends on ureq for proxy which'd be nice to replace with reqwest if possible and it depends on nginx for tls termination which would really be better to replace with whatever native ACME solution you cook up for the directory.

In either case I don't think merging the repos affects payjoin crate's 1.0 stability now so I'm just noting these things for when the time is right.

@DanGould The ureq crate is only used for integration test in the ohttp-relay repo . i did a replacement with reqswest on my local branch and all test-cases pass .

[benalleng]

Feel free to open a PR on the ohttp-relay so we can take a look