Under the hood, CONN is based on three "pools" of peers:

• ConnDB: a persistent database of addresses to connect to
• ConnHub: a façade API for currently active connections
• ConnStaging: a pool of potential new connections

ConnDB contains metadata on stable servers and peers that have been successfully connectable. ConnHub is the central API that allows us to issue new connections and disconnections, as well as to track the currently active connections. ConnStaging is an in-memory ephemeral storage of new possible connections that the user might want to approve or disapprove.

Then, ConnQuery has access to those three pools, and provides utilities to query, filter, and sort connections across all those pools.

ConnScheduler is an opinionated (⚠️) plugin that utilizes ConnQuery to select peers to connect to, then schedules connections to happen via ConnHub, as well as schedules disconnections if necessary. Being opinionated, CONN provides an easy way of replacing the default scheduler with your own scheduler, see instructions below.

There is also a Gossip Compatibility plugin, implementing all the legacy APIs, so that other SSB plugins that call these APIs will continue to function as normal.

When you install the ssb-plugin, it will actually setup three plugins:

[conn, connScheduler, gossip]

The default scheduler is roughly the same as the legacy ssb-gossip plugin, with some opinions removed and others added. The scheduler has two parts: discovery setup on startup, and periodic connections/disconnections.

Discovery setup:

• Read the SSB log and look for "pub" messages, and remember them
• Listen to a stream of Bluetooth nearby devices, and stage them

Periodic connections/disconnections:

• With (5sec) exponential backoff, try to connect to at most 5 room servers
• With (10sec) exponential backoff, try to connect to at most 2 non-room peers that we have connected successfully before
• With (30sec) exponential backoff, try to connect to at most 2 non-room peers that we have never with connected before
• With (1min) exponential backoff, try to connect to at most 3 non-room peers that have we always failed to connect with
• With (4min) exponential backoff, try to connect to at most 1 non-room peer that seem to run a legacy version of the gossip plugin

In none of the cases above shall we connect to a peer that we block. In addition to the above, the following actions happen automatically every (approximately) 1 second:

• Connect to (at most 3) staged peers we follow
• Disconnect from connected peers that have just been blocked by us
• Disconnect from peers that have been connected with us for more than 30min
• Disconnect from peers that have been pending in "connecting" status for too long
  • "Too long" means 30sec for LAN peers
  • "Too long" means 1min for Bluetooth peers
  • "Too long" means 5min for DHT invite peers
  • For other types of peers, "too long" means 10sec
• Stage non-blocked peers that are in ConnDB marked as autoconnect=false
• Unstage peers that have just been blocked by us
• Unstage LAN peers that haven't been updated in ConnStaging in 10 seconds
• Unstage Bluetooth peers that haven't been updated in ConnStaging in 30 seconds

Database cleanups:

Upon starting the scheduler:

• Remove database entries for any LAN or Bluetooth peers (these are rediscovered just-in-time)

Other events:

• Upon wakeup (from computer 'sleep'), fully reset the ConnHub
• Upon network (interface) changes, fully reset the ConnHub
• Upon a disconnection, try to connect to some peer (section above)

To experiment with your own opinions for establishing connections, you can make your own ConnScheduler, which is just a typical SSB plugin. You can write in the traditional style (like other SSB plugins), or with OOP decorators. The example below uses OOP decorators.

Here is the basic shape of the scheduler:

import {plugin, muxrpc} from 'secret-stack-decorators';

@plugin('1.0.0')
module.exports = class ConnScheduler {
  constructor(ssb, config) {
    // basic setup here
    this.ssb = ssb;
  }

  @muxrpc('sync')
  public start = () => {
    // this is called when the scheduler should begin scheduling connections

    // You have access to CONN core here:
    const query = this.ssb.conn.query();
    this.ssb.conn.stage(addr);
    this.ssb.conn.disconnect(addr);
    // ...
  }

  @muxrpc('sync')
  public stop = () => {
    // this is called when the scheduler should cancel its jobs
  }
}

Note that the name of the plugin must be exactly ConnScheduler (or connScheduler) and it must have the methods start() and stop(), because the CONN core will try to use your scheduler under those names. The rest of the contents of the ConnScheduler class are up to you, you can use private methods, etc.

When you're done building your scheduler, you can export it together with CONN core and the gossip compatibility plugin like this:

var CONN = require('ssb-conn/core')
var Gossip = require('ssb-conn/compat')
var ConnScheduler = require('./my-scheduler')

module.exports = [CONN, ConnScheduler, Gossip]

That array is a valid secret-stack plugin which you can .use() in ssb-server.

The legacy gossip plugin is one of the oldest parts of the SSB stack in Node.js, and it contained several old opinions. It wasn't designed with multiserver in mind, so it made a lot of assumptions that peers have host/port fields. Nowadays with Bluetooth and other unusual modes of connectivity, that assumption breaks down often.

The gossip plugin also did not have the concept of "staging", which is useful for ephemeral connections (LAN or Bluetooth) in spaces that may have many strangers. So the gossip plugin tended to connect as soon as possible to any peer discovered.

Also, since the gossip plugin was a monolith, it had all these concerns (cold persistence, in-memory tracking of current connections, ephemeral peers, scheduling, old and new style addresses) squashed into one file, making it hard and brittle to change the code.

The objectives with CONN were to:

• Untangle the codebase into modular components with single responsibilities
• Standardize the assumption that addresses are always multiserver addresses
• All "pools" (DB, Hub, Staging) are key-value pairs [address, dataObject]
• Make scheduling easily customizable but provide an opinionated default