- [x] when we have the whole torrent, there's no point talking to peers that also have the whole torrent and keep reconnecting to them.
- [ ] per-file stats
- [x (partial)] per-peer stats
- [x] use some concurrent hashmap e.g. flurry or dashmap
- [x] tracing instead of logging. Debugging peers: RUST_LOG=[{peer=.*}]=debug
      test-log for tests
- [x] reopen read only is bugged
- [x] initializing/checking
  - [x] blocks the whole process. Need to break it up. On slower devices (rpi) just hangs for a good while
  - [x] checking torrents should be visible right away
- [x] server persistence
  - [x] it would be nice to restart the server and keep the state
- [x] torrent actions
  - [x] pause/unpause
  - [x] remove including from disk
- [ ] DHT
  - [x] bootstrapping is lame
  - [x] many nodes in "Unknown" status, do smth about it
  - [x] for torrents with a few seeds might be cool to re-query DHT once in a while.
  - [x] don't leak memory when deleting torrents (i.e. remove torrent information (seen peers etc) once the torrent is deleted)
  - [x] Routing table - is it balanced properly?
  - [x] Don't query Bad nodes
  - [-] Buckets that have not been changed in 15 minutes should be "refreshed." (per RFC)
    - [x] Did it, but it's flawed: starts repeating the same queries again as neighboring refreshes
          don't know about the other ones, and DHT returns the same nodes again and again.
  - [x] it's sending many requests now way too fast, locks up Mac OS UI annoyingly
  - [x] store peers sent to us with "announce_peer"
  - [x] announced peers should be persisted (partial)
  - [ ] clean up announced peer cache
    - [ ] only send a token to torrents really close to us
  - [x] After the search is exhausted, the client then inserts the peer contact information for itself onto the responding nodes with IDs closest to the infohash of the torrent.
  - [x] Ensure that if we query the "returned" nodes, they are even closer to our request than the responding node id was.

incoming peers:

- [x] error managing peer: expected extended handshake, but got Bitfield(<94 bytes>)
- [x] do not announce when merely listing the torrent

someday:

- [x] cancellation from the client-side for the lib (i.e. stop the torrent manager)
- [x] favicons for Web UI

desktop:

- [x] on first run show options
- [x] allow to change options later (even with a session restart)
- [ ] look at logs - allow writing them to files? Set RUST_LOG through API

persistence:

- [ ] store total uploaded bytes, so that on restart it comes back up

efficiency:

- [ ] once the torrent is completed, we don't need to remember old peers

refactor:

- [x] session persistence: should add torrents even if we haven't resolved it yet
- [x] where are peers stored
- [x] http api pause/unpause etc
- [x] when a live torrent fails writing to disk, it should transition to error state
- [x] something is wrong when unpausing - can't finish. Recalculate needed/have from chunk tracker.
- [x] silence this: WARN torrent{id=0}:external_peer_adder: librqbit::spawn_utils: finished with error: no longer live

- [x] start from error state should be possible from UI
- [x] checking is very slow on raspberry
      checked. nothing much can be done here. Even if raspberry's own libssl.so is used it's still super slow (sha1)
- [ ] .rqbit-session.json file has 0 bytes when disk full. I guess fs::rename does this when disk is full? at least on linux. Couldn't repro on MacOS