Why poll?

[richardschneider]

Every 10 seconds a multicast query is sent. I assume it is done to find new IPFS peers.

Every time a new peer starts, it will send a query and ALSO respond to it with it's info.

Since this is multicast, all other nodes will see its response and add the peer to it's tables.

So polling every 10s is not needed.

[richardschneider]

@whyrusleeping @diasdavid @Stebalien please comment

[whyrusleeping]

@richardschneider are you sure? If thats the case, then we could remove the polling. I would like to see it working in practice without polling before I would be okay disabling it though

[richardschneider]

@whyrusleeping added a test to PR #71

  it('find all peers', function (done) {
    const options = {
      port: 50001 // port must be the same
    }
    const mdnsA = new MulticastDNS(pA, options)
    const mdnsB = new MulticastDNS(pB, options)
    const mdnsC = new MulticastDNS(pC, options)
    const peersA = {}
    const peersB = {}
    const peersC = {}

    // After all the peers have started, each peer should see two other peers.
    function check (receiver, peerInfo) {
      receiver[peerInfo.id.toB58String()] = true
      if (Object.keys(peersA).length === 2 && Object.keys(peersB).length === 2 && Object.keys(peersC).length === 2) {
        parallel([
          (cb) => mdnsA.stop(cb),
          (cb) => mdnsB.stop(cb),
          (cb) => mdnsC.stop(cb)
        ], done)
      }
    }
    mdnsA.on('peer', (peerInfo) => check(peersA, peerInfo))
    mdnsB.on('peer', (peerInfo) => check(peersB, peerInfo))
    mdnsC.on('peer', (peerInfo) => check(peersC, peerInfo))
    series([
      (cb) => mdnsA.start(cb),
      (cb) => setTimeout(cb, 500),
      (cb) => mdnsB.start(cb),
      (cb) => setTimeout(cb, 500),
      (cb) => mdnsC.start(cb)
    ], () => {})
  })

[richardschneider]

ipfs/js-ipfs#1123

[Stebalien]

That sounds correct however, it doesn't handle the case of joining/leaving networks. Ideally, libp2p would expose an API to monitor for new network addresses (I want this so badly) but implementing this properly is a pain (although we should have something that just does this on a timer anyways...).

For now, maybe poll once every 10 minutes, every half hour? That should at least cover the case of "close laptop, move elsewhere, open laptop, discover peers".

[richardschneider]

I really don't like network polling because it doesn't scale well; see comment.

How about, detecting an "open laptop" and then do a "discover peers". Something like sleeptime can be used to detect "open laptop".

[mkg20001]

Btw, sleeptime has a typo that prevents custom intervals from working (timrach/sleeptime#1) and breaks if the application does too much work on it's main thread (edit: with break i mean that it does trigger even if the laptop wasn't asleep) (edit2: could be fixed by using child_process.fork and notifying the main process over ipc)

[richardschneider]

@mkg20001 Thank's for the heads-up.

I assume that 'breaking' (false positives) is dependent on the interval. The lower the interval the more likely a false-positive is triggered. I would expect our interval to be 1 or 2 minutes. Hopefully, our app would not spend that amount of time on the 'main thread'.

[richardschneider]

Also, sleeptime needs a stop method, so that mdns can gracefully shutdown.

[Stebalien]

@richardschneider that's a fair point but, in go at least, have a server profile that disables mdns for this reason. Also, I'd hope that most datacenters turn of multicast (or at least multicaset between multiple customers).

In go-libp2p, we have a InterfaceListenAddresses method that'll return all fully-specified we're listening on, resolving, e.g., 0.0.0.0 to our local machine's actual IP addresses. Is there an equivalent in javascript? That's what I'd use to detect interface changes.

Note: this still won't handle cases of networks joining each other but... that's not exactly a common case that I care about.

[richardschneider]

I think I read in the MDNS RFC that mutlicasting is for the local network and not joined networks, but I could be wrong.

[Stebalien]

I was thinking about network partitions. If my network gets partitioned, I join, and then it gets unpartitioned, I will never advertise myself to the other partition. However, on local networks this is really not that much of an issue.

[mkg20001]

Also, sleeptime needs a stop method, so that mdns can gracefully shutdown.

@richardschneider I have created a rewritten version of sleeptime (https://github.com/mkg20001/sleeptime2) that fixes that issue and the other I mentioned above