Baby

Bushbaby Automated Bamboo Yields

Sync for Bamboo stores using the Bushbaby protocol

Connection tuning

The per-connection idle timeout is governed by four settings, all read from Application config (config :baby, ...) at connection startup, and overridable per-connection:

• :outrate — milliseconds between outbox ticks. Default: a random prime near 75 (jittered so connections don't tick in lockstep).
• :handshake_spins — idle budget, in outbox intervals, before a valid HELLO has been received. A connection that never proves itself is dropped quickly so an anonymous flood cannot pin the listener's connection slots. Default: a random prime near 375 (~30s at the default outrate).
• :max_spins — idle budget, in outbox intervals, once the initial replication sync has completed. Default: a random prime near 1200.
• :bootstrap_spins — idle budget while the initial sync is still in progress, giving the peer time to compute its WANT list before being dropped. Default: a random prime near 3000.

The idle counter resets on any send or receive, so the budgets measure consecutive silence rather than wall-clock time on a busy link. A connection that exceeds its budget is dropped and re-established on the next cryout.

To estimate wall-clock time, multiply the budget by :outrate — e.g. a default max_spins of ~1193 at a default outrate of ~80ms is roughly 95 seconds, and a default bootstrap_spins of ~2969 is roughly 4 minutes.

config :baby,
  outrate: 100,
  handshake_spins: 375,
  max_spins: 1200,
  bootstrap_spins: 5000

Wire buffer

Inbound bytes that have not yet formed a complete protocol frame are buffered in the connection's wire state. A peer that dribbles undecodable bytes could otherwise grow this buffer without bound, so it is capped:

• :wire_cap — maximum buffered bytes per connection before the connection is dropped. Read from Application config (config :baby, wire_cap: ...) at connection startup, overridable per-connection. Default: 32MB.

The cap sits comfortably above the largest legitimate frame (a single message may carry an entire log), so it should only trip for garbage; note the protocol spec guidance on authoring log entries that fit within peers' limits.

Listener

The ranch listener that accepts peer connections has a deliberate ceiling on concurrent connections, so a flood of anonymous connections cannot exhaust file descriptors or slot-bound the legitimate peer:

• :max_connections — maximum concurrent connections accepted per listener. Read from Application config (config :baby, max_connections: ...) at startup, overridable per-clump. Default: 256 (ranch's own default is 1024).

config :baby, max_connections: 256

A clump whose configured port is already in use fails the whole application start with {:error, reason} (typically :eaddrinuse) rather than silently running without that listener. Listeners that did bind before the failure -- and any mDNS announcements they made -- are rolled back, so retrying the same configuration succeeds once the port is freed.

Local network discovery (mDNS)

Clumps can advertise themselves to peers on the local network and can find those peers without statically configured addresses. This is built on mdns_lite, which is started automatically when any clump needs it. Services are advertised as standard _bushbaby._tcp DNS-SD records whose TXT payload carries the announcing clump's clump_id, so they are visible to any mDNS tooling (dns-sd, Avahi, etc.), not just other Baby nodes.

Per-clump configuration (config :baby, clumps: [...]):

• announce — true advertises this clump's listener via mDNS; a keyword list of Baby.Mdns.announce/3 options (e.g. instance:) customizes the advertisement. Default: false.
• Within a cryouts entry, mdns: — true (or keyword options) turns that cryout into a "meta cryout": instead of dialing a fixed {host, port}, the monitor periodically browses the local network and connects to every discovered peer of the same clump it isn't already talking to. A period may be given either at the cryout's top level or inside the mdns options; unconfigured meta cryouts cycle about once a minute (a random prime number of seconds near 60, so many peers do not browse in lockstep), while fixed-host cryouts default to {17, :minute}.

config :baby,
  clumps: [
    [
      id: "Quagga",
      controlling_identity: my_identity,
      port: 8483,
      announce: true,
      cryouts: [[mdns: [period: {5, :minute}]]]
    ]
  ]

Because discovery re-runs on every meta-cryout cycle, clump-mates that come and go are picked up (and re-dialed) automatically, exactly as fixed-host cryouts re-establish dropped connections.

Each clump runs its own monitor and its own dynamic supervisor for the connections its cryouts spawn. Clumps are therefore independent -- a failing clump does not disturb its siblings -- and tearing a monitor down (on a clump's stop or crash) takes its spawned connections with it.