ksedgwic/clboss v0.17.1-rc1

on GitHub

Release candidate 1 for 0.17.1. Changes since 0.17.0:

Changed

• The XRebalancer grant now fades out as a side's real forwarded
  volume grows. The credit is assumed earned on
  clboss-xrebalance-grant-weight percent of the peer's capacity
  (new dynamic option, default 25, both sides; it was a whole
  capacity-turn), and real forwards replace it one for one: a side
  with no record reads exactly clboss-xrebalance-grant, at half
  the weight the rate is half grant and half record, and from the
  weight on the side reads its own record alone (#339).

Fixed

• A total peer blackout (lightningd started with --offline, Tor
  down, a firewall) no longer demerits every peer or counts as
  CLBOSS uptime; with clboss-auto-close it could close channels
  as the peers came back. CLBOSS also honours --offline itself:
  it no longer connects to peers, tests channel candidates, or
  opens channels while the flag is set, since an explicit connect
  used to go through and defeat both the flag and the guard (#346,
  #347).
• Channel-candidate selection by popularity now samples with A-ES
  (Efraimidis and Spirakis) instead of an A-Chao variant that
  omitted the reservoir-size factor m: selection was
  arrival-order-dominated, and an overweight item could still be
  evicted by later arrivals. The popularity finder now shares
  Stats::ReservoirSampler with the other finders instead of a
  drifted inline copy (#343).
• Backing up a live data.clboss with sqlite3 .backup or .dump
  could kill CLBOSS, and lightningd with it when CLBOSS is an
  important plugin: a write that found the file read-locked failed
  at commit, and that failure was thrown from a destructor, which
  ends the process. The connection now waits up to ten seconds for
  a lock, as lightningd does for its own database; a commit that
  still fails is reported to the caller as an error, and the sendpay
  hook path logs it and carries on. The README gains a backup note
  (#335).
• The channel creator could open a second channel to a peer we
  already had one with. Three of the four candidate finders do not
  check for an existing channel, the candidate table is swept of
  channeled peers only on the 10-minute listpeers poll, and the
  creator runs about every two minutes without a check of its own,
  so a proposal naming an existing peer could reach the planner in
  the window between sweeps. The creator now drops such proposals
  against a fresh listpeerchannels before any funds are planned.
  The channel-state test the sweep relies on also counted only
  OPENINGD and CHANNELD states, so a dual-funded open in
  progress looked unchanneled; it now counts DUALOPEND states
  too, and no longer reads past the end of the shorter ONCHAIN
  (#332).
• A splice no longer reads as losing the channel. CLN keeps a
  spliced channel open and forwarding in CHANNELD_AWAITING_SPLICE
  until the new funding locks in, which the create/destroy monitor
  took as leaving CHANNELD_NORMAL: it announced a
  ChannelDestruction, and a ChannelCreation minutes later, so
  the peer's complaint history was archived, its channel age reset,
  and the fee monitor's state for it flushed. That state now counts
  as live. The monitor's polling path also never reported a
  destruction: peers that had gone were appended to the creations
  list (#344).
• A channel awaiting splice lock-in (CHANNELD_AWAITING_SPLICE, open
  and forwarding for the hours until the new funding confirms) was
  skipped wherever channels were summed or chosen by
  CHANNELD_NORMAL: the node balance swapper's sendable and
  receivable totals, and the by-balance fee modder. They now admit
  that state. The rebalancer and the active prober keep to
  CHANNELD_NORMAL: the xrebalance plugin accepts only that state and
  rejects a request naming any other channel, and a probe is sized
  from a spendable balance that CLN reports for the old funding until
  lock-in. While a splice-out is pending, CLN reports the balance of
  the old funding until the new one locks in but already admits HTLCs
  against the lower post-splice balance; the node balance swapper and
  the by-balance fee modder now deduct the pending splice-out the same
  way, so a channel awaiting a splice-out is never read as holding
  funds it no longer has (#345).

Removed

• The PeerJudge modules (Boss/Mod/PeerJudge/), an earnings-based
  channel closer that was compiled in but never instantiated. Its
  close path issued a unilateral close after 180 seconds regardless
  of whether the peer was connected, the pattern 0.16.3 removed from
  the complaints desk (#324), so wiring it up as it stood would have
  force-closed offline peers as a matter of course. Nothing in the
  running plugin changes; the code stays in history should an
  earnings-based closer be wanted again (#336).