[erlang-questions] Twoorl: an open source Twitter clone

Joe Armstrong <>
Sun Jun 1 16:10:22 CEST 2008

It *is* an interesting discussion - not about Twitter - but about architectures.

There seem to be some implicit assumptions here:

Let's suppose that twoorl services are accessed through a *single*
name (www.twooerl.org) or something.

Let's assume we have 20 (for the sake of argument) back-end machines
(could be hundreds or thousands  though)( I guess scalable means that
we can just add/remove backend machines without breaking
things :-)

The first step must be to associate 20 IP addresses with this single
name - some kind of DNS load balancing should be possible. (how do you
do this?????)(is this DNS round robin????)(does this scale to
thousands of machines???)

The user tries to connect to www.twooerl.org and looks up the address
in DNS the result is one of these IP addresses.

The user connects to one of theses IP addresses and and requests data
for "joe" - they are immediately redirected to the machine having
joe's data.

The simplest way to find a machine would be to use the idea in chord.

For a machine with IP we create a tuple

{"", Md5("")} we do this for all machines and
sort my the Md5 sum

so we get a list of twenty machines, say:

   [{"", Md1}, {"", Md2}, ... ]

Which machine is "joe" stored on? - to find this we compute md5("joe")
and find the
first machine in the list whose Md5 sum is greater than md5("joe").

The initial machine would perform this computation and redirect to the
machine with joe's data.

The initial machine can also check the liveness this second machine -
and if it is unresponsive
redirect to a machine containing a replica of joe's data (which could
this be? - imagine the
machines arranged in a circle and redirect to the machine nearest to
180 degrees away from the
original machine)

A problem occurs if the original machine is dead - (ie the one that
DNS said was the address associated
with www.twoo.erl) - if the back-end machines are in pairs that
monitor each other then I guess
a gratuitous arp can fix the problem and reassign the the IP address
of the failing machine to a
new machine (which must now take over the role of the first machine)

In all of this the fundamental problem seems to be that if the server
fails then we have to do a lot
of messing around to keep the server on a fixed IP address.

It would be a zillion times easier if the *client* did all this
messing around. If the client
knew about N server address it could automatically redirect to a
different server if the primary server
failed. if the client knew about N machines and performed the md5(Key)
calculation to find the correct machine then the problem would be very
easy to solve with no messing around in the server.

(( The fact that DNS has multiple addresses make writing a resolver
really easy, if one DNS server
fails just try anothert ))

Now if we're talking "standard browsers" then accessing multiple sites
is painful. Javascript and flash
plugins etc. are very restrictive in the sites that they can open
sockets to (for good reasons)
I think only the originating site is permitted.

If we could persuade the users to install a small proxy one their
machines then all these problems would
go away - a standard browser could talk to a proxy on localhost and
this could talk to the multiple
back ends.

What appears to be a tricky problem in scaling things if we have to
keep the back-end servers on fixed
addresses seems a lot easier if the clients have some limited intelligence.

The next step is (ovf course) to allow all the clients to collectively
behave as if they were as server -
I think therefore that the problem is really one about the base level
of a P2P architecture and not
about a twitter architecture per se.

If we did have a simple proxy component that allowed messaging to
multiple sites then this and many other
problems would be easily soluble.

We might imaging a proxy that that was programmable:

It presents a menu like this:

    {allow, ServiceName, at, Addr1, Addr2, Addr3, .....}

(( example {allow, twitter, at, www.a.b, www.p.q, www.c.d}
    - this means that the proxy can open "twitter" sessions to the
three "trusted" machines in the list

    then the web browser could access the proxy that could talk to the
trusted machines -
the trusted machine should just redirect to other machines, until
finally the desired machines are found. ))

  Would the fact that you have to install a proxy server limit
deployment of a service? - probably.
Also it opens up a new bag of worms (security) - for good reasons
browsers do not allow plugins
to access multiple sites (only the originating sites).

I suppose therefore, that the inner cluster that provides the service
would have a full P2P structure
and that the service would be accessed by DNS round robin with IP
failover to handle the errors.

I suspect that architecture like this are being used in some Erlang
systems (the details might vary)

If anybody would like to post code and go into the details of how to
rig systems for DNS load balancing
(or whatever it's called) and for IP monitoring and fail-over then we
could get to the interesting
part of building the application)

(( the next bit will be to look at the limits of scaling - still
nobody has talked numbers -
   how far can we press a two-tier system - with say 20 name servers
in the front-end that *only* do
   redirects - this bit should be very fast ))

(( By combining twitter with IRC we might make the option of
installing a proxy more attractive -
   The irc "server" for a group G should really me the client that
first created the group G -
    If G drops out the second machine in the group could become the
server and so on.
    Really twitter is like having one IRC grrup per person. many
people can join but only the
    owner can write to it. What's the difference (architectually) .


/Joe Armstrong

On Sun, Jun 1, 2008 at 10:41 AM, David Mitchell <> wrote:
> This is a REALLY interesting discussion, but at this point it's
> becoming obvious that I don't know enough about Twitter...
> Are you suggesting that Twoorl should be architected as follows:
> - when they register, every user gets assigned their own RabbitMQ
> incoming and outgoing queues
> - user adds a message via Web/Yaws interface (I know, this could be
> SMS or something else later...)
> - message goes to that user's RabbitMQ incoming queue
> - a backend reads messages from the user's incoming queue, looks up in
> e.g. a Mnesia table to see who should be receiving messages from that
> user and whether they're connected or not.  If "yes" to both, RabbitMQ
> then forwards the message to each of those users' outgoing queues
> - either the receiving users poll their outgoing queue for the
> forwarded message, or a COMET-type Yaws app springs to life and
> forwards the message to their browser (again, ignoring SMS)
> This seems like a reasonable approach; I'm just curious if that's what
> you're suggesting, or whether you've got something else in mind.
> Great thread, and thanks Yariv for getting this discussion going with Twoorl
> Regards
> Dave M.
> 2008/6/1 Steve <>:
>> On May 31, 5:04 pm, "Yariv Sadan" <> wrote:
>>> ...but it's the only way you can scale this kind of service when N is
>>> big.
>> Hmmm, Yariv, aren't you still thinking about this in the way that Dave
>> Smith pointed to as the heart of the issue? i.e.
>> Dave said: "My understanding is that the reason they have such poor
>> uptime is due to the fact that they modeled the problem as a web-app
>> instead of a messaging system."
>> I'm aware that you are likely a good way away from hitting any
>> scalability problems, but some kind of tiering would seem to be
>> appropriate if twoorl is to be "twitter done right". Yaws at the front
>> end, definitely - but rather /RabbitMQ/ at the back end. I believe
>> that you'd then have the flexibility to distribute/cluster as
>> necessary to scale to the required level (whatever that may be).
>> For sure, Twoorl is a great demo of what can be done with Erlang in an
>> incredibly short time. I'm a relative noob to Erlang, and have learned
>> a great deal from your blog/code/examples.
>> Steve
>> _______________________________________________
>> erlang-questions mailing list
>> http://www.erlang.org/mailman/listinfo/erlang-questions
> _______________________________________________
> erlang-questions mailing list
> http://www.erlang.org/mailman/listinfo/erlang-questions

More information about the erlang-questions mailing list