RailsConf 2014 - Rack-AMQP: Ditch HTTP Inside SOA! by Josh Szmajda

JOSH SZMAJDA: All right. You guys ready to
talk about SOA? Woo!

I'm Josh. Hi. How you doing? I'm the CTO

at Optoro. We make software for retailers.
Help them

deal with their stuff at scale. I'm working
on

a project called RACK-AMQP. RACK-AMQP is a
collection of

different projects, including an RC, some
client libraries, and

most notably, Jackalope, which is a new Ruby
web

application server, like Unicorn, Passenger,
Thin, or, or anything

like that. Except that it's special-purpose.
It's for service-oriented

architectures. And it doesn't actually speak
HTTP at all.

So. Why am I doing this crazy thing? Well,

my app drove me to it. My app is

huge. It's really complicated. This is kind
of a

services diagram of my app. Or where we'd
like

to go with it.

There's over three-hundred and fifty models
in one Rails

application. The test suite takes over eight
hours. It's

kind of a headache. You could put it. So

we decided the way to move forward with it

is to break it up into smaller applications
that

work together. And, of course, that means
SOA.

Now, I mean, SOA is really wrapped up in

a lot of pretty awful baggage, when you come

down to it. SOA, you think about enterprise
service

buses. You think about strict interface definitions.
Versions. I

mean it's a lot of complexity that you have

to worry about. Extra load balancers. More
configuration. More

distributed tracing and monitoring. I mean,
it's a, it's

a real pain.

So I, I hate SOA, right. It makes me

feel like this. So, but when you get down

to it, SOA is actually not all of that.

SOA is really just independent actors communicating.
They're working

together to get something useful done. So,
really, SOA

can mitigate complexity by enforcing simplification.
Each service only

has to think about its own little bubble.
It

only has to worry about its little component.
And

it doesn't have to think about anything else
that

goes on in its entire system.

You can let your system architect worry about
what's

going on across the whole thing. All you have

to do is pay attention to your one little

service and you can be pretty happy.

So, when you get down to it again, SOA

is really just independent actors communicating.
Now that communicating

bit is important. How they communicate actually
determines how

your application behaves. And the kinds of
communication you

can use impacts the kind of application you
can

write.

So, the typical communication pattern that
we're most familiar

with as programmers is sort of direct messaging.
I

am an object. I can send a message to

my friend. My friend is gonna give me back

a response. It's very normal. It's what, in
fact,

we use all the time with HTTP. But when

you think about it, what we actually want
to

do is be able to send different kinds of

messaging structures. We want to publish that
something happened

in my system and have people subscribe to
that

event and take some interesting action.

It's all about decoupling. I don't have to
know

about my neighbor anymore. I just need to
know

that I did something, and then somebody else
might

care that, oh, you did something. That's great.
Let

me do something about that.

But again, like I said, let me. We're Rails

developers. We're used to writing web applications.
We speak

HTTP all the time. HTTP allows us to do

direct messaging, so that's what we tend to
do.

I'd say most SOAs out there, when they get

off the ground, they start building web applications
that

speak with each other through HTTP. It's what
we're

familiar with.

But what would we actually like our transport
mechanism

to be? We want it to be, you know,

data structure independent. HTTP specifies
that there be a

body, but it doesn't specify the format of
the

body. We want it to be fast. We want

it to be open, transparent, understandable,
clear. You know,

debuggable. Observable. And flexible.

So, is HTTP all of those things. We also

need, not only direct request response type
of messaging,

like RPC, but we also need pub, subscribe,
and

broadcast. Additionally, we don't need more
load balancers. More

SSL certificates. Distributed logging. Trying
to figure out what's

going on in this service over here versus
that

one over there. It's a nightmare.

And, you know, looking up the various services
in

your system, it's a pain. So, again, I mean,

HTTP, you know, it has, it is blind. It's

fast. It's open, clear, it's well-understood.
It's what we

know. It's really great for RPC. It's a better

alternative.

And that's AMQP. AMQP is blind, also. It doesn't

care what data you're transmitting across
it. It's extremely

fast. It's an open system. It's great for
all

messaging pattern. Not just direct messaging,
but pub/sub, and

queue-based messaging. It's centralized, which
in terms of a

service-oriented architecture is very beneficial,
in that you have

one point to manage, one point to pay attention

to everything at. And the only issue is that

it's not what we're familiar with. It's not
what

we know. You know, again, we're web developers.
We're

not AMQP developers. At least not the most
of

us.

So it's a challenge.

So the RACK-AMQP project that I'm working
on is

really designed to make AMQP simple. We are
trying

to introduce as few opinions as absolutely
necessary in

order to make this possible. And to give you

a stepping stone into the world of AMQP where

you can start to integrate more advanced concepts
into

the system, without having to relearn your
entire world.

So it also lets us continue to build what

we know. Because it's built on Rack, you know,

we're web-developers. We can continue to use
the web

concepts that we're familiar with. We can
developer locally

on our web machines using the systems that
we're

comfortable with. And then when we deploy,
we simply

deploy it to Jackalope, instead of to Unicorn
or

Passenger or whatever, and it's magically
in this world

of AMQP.

So. How does that work? Well, let's look at

AMQP a little bit more in depth, first. So

AMQP is the advanced messaging queue, message
queueing protocol.

It is an open standard. It's developed by
Oasis,

which is an open standards body. It, again,
like

I said, uses a centralized distribution pattern.
That centralized

distribution pattern allows you to do the
various kinds

of messaging that are really handy for us.
The

pub/sub and the other interesting thing. This
implies that

we have a broker in the center that all

of our services connect to and transmit messages
through

that broker. And that broker can make intelligent
decisions

about how to route those messages.

AMQP messages themselves, they have headers
and bodies, just

like HTTP. There are well-known fields, like
content type

and priority, in this case. And we can actually

leverage the conventions that we understand
from HTTP to

pull AMQP into the world. So AMQP, like I

mentioned, has these headers and bodies. It
has well-known

fields. And HTTP essentially has headers and
bodies and

also well-known fields. So we can essentially
emulate HTTP

over AMQP without too much trouble. And, again,
the

AMQP supports the RPC messaging pattern, which
is what

HTTP is.

So, one of the parts of the project is

the AMQP HTTP RC, which, it simply does as

little as it has to to define the structure.

It's just a documentation on how we're doing
the

mapping. One of the other goals is that we

want to make inner-operability with other
languages and other

platforms really easy. So having an RC will
allow

it, allow us to have just a central document

to refer back to in constructing other servers.

So let's look at HTTP a little bit more

in-depth. How does HTTP work? Well, you know,
when

we send a request, we send a specific request

to a server, to a, you know, specific IP

address and port. And we get back a response.

And, of course, the request has things like
the

HTTP verb, the path, headers, the bodies.
And the

response comes back with a response code.

It's all stuff we're familiar with. So all
we

had to do is map that into AMQP terms,

which is essentially what we've done here.
And this

is just a reference for just a few of

the things that we're doing. Like content
type. There's

already a content type header in AMQP, so
we

just reuse that. Some things didn't make sense.
Like

the protocol for example. In HTTP you've got
HTP

or HTPS or HTTP plus dev or whatever. But

with AMQP, your protocol's already negotiated
when you connect

to the broker. So it doesn't really make sense

for us to worry about that in the messages

we're passing around the system.

The host name becomes, essentially, the routing
key. The

queue target that we're sending to in AMQP.
And

it makes it really simple. So, this looks
like

this, essentially, in code. This is an example
of

how you could write a client that speaks AMQP,

HTTP. You create a call-back queue. We'll
get back

to that in one sec. You publish a message

that looks like an HTTP message to the target

queue. And then you wait for a response.

Now, the callback queue is actually a really
important

thing here. AMQP is a hundred percent queue
based.

It's all asynchronous under the hood. But,
you know,

we want to write a synchronous system. HTTP
is

synchronous. I send a request. I wait for
the

response, or receive the response and move
on. So,

to do that in AMQP, we create a respond

queue that we are going to listen to in

the broker. It's an anonymous queue. There's
a great

convention. We're doing this in AMQP. It's
very easy.

So we create a response queue, we get the

name of that queue, and then as part of

the message we send to our target, we say,

here's the response queue to reply back at.
And

then we simply wait for the response on that

call-back queue to come back. So that lets
us

get the synchrony that we're familiar with,
that we

need, with HTTP style programming, while still
having an

asynchronous system under the hood.

RACK, of course, we're all using RACK today.
Even

if you're not familiar with it. RACK, of course,

is what runs all Ruby web application servers.
Turns

out that RACK actually doesn't care about
HTTP. All

that RACK really defines is that you had defined

a call method that receives an environment,
which is

basically a hash, and it responds back with
a

three-element array. The HTTP response code,
headers, and the

body.

That hash doesn't actually have to be anything
but

a hash. It just has to look like an

HTTP environment. So emulation was actually
pretty simple. So

this is basically how it looks. We subscribe
to

a queue on the server, and then for every

message we receive back for that queue, we
unpack

some things into variables. We create what
looks like

the HTTP environment and we have the request
method

and query string. If you ever wrote CGI way

back in the day, this might look familiar
to

you.

And then we essentially pass that environment
onto our

RACK application, receive the response from
the RACK app

and then publish that back to the response
queue.

So this is what Jackalope is doing for us.

Jackalope is emulating HTTP for our RACK applications.
And

it just works. I'll show you a demo at

the end. You don't have to modify your code

at all. You can just simply deploy to it,

and it's speaking AMQP instead of HTTP. So
that's

it. We're done. We can go on vacation, right?

Well. I didn't talk about how to actually
put

this in production, other than just use Jackalope.
So

you have to choose a broker in AMQP. So

what we're using is RabbitMQ. RabbitMQ is
a well-known

AMQP broker. If you're doing any AMQP you're
probably

using RabbitMQ, most likely. It's used by
giants out

there. I mean, Google, VMWare, NASA. All these
people

are using Rabbit pretty heavily.

It's extremely scalable. It's fault-tolerant.
It's distributable. And it's

secure. It does everything you'd ever want.
And it

also gives you a really great management console.
We

can go in and see what's going on in

your system. Like I mentioned, the distributed
logging we

saw in the, in the last talk, if you

were here for that, trying to get an idea

of what's going on across your system can
be

challenging. Rabbit doesn't tell you everything
you'd want to

know, but it at least gives you an idea

of what queues are being used. Their depths
at

the point. The kind of information that can
help

you get, at least get a handle on it.

Additionally, you're gonna have to think about
how you

talk to the real world. I mean, AMQP is

all well and good behind the scenes, but how

do I continue to interact and serve my clients

that I actually care about, to the point of

getting paid, right.

So this is the architecture we typically use
at

Optoro. We have our Rabbit server in the middle,

and all the various services sitting on Jackalope
that

talk to Rabbit. And then additionally we have
one

or more API services that are continuing to
be

deployed on Unicorn, in our case. They talk
to

the outside world and internally translate
the, the needs

through AMQP and the rest of the system.

One of the things we actually published for
each

of our services is a client gem that kind

of isolates the worry, the worry about exactly
which

transport we're using away from the consumer,
the API

here. So it actually is, is really simple.

And so the way that we do that is

we typically have been using HTTParty for
our HTTP

communication needs. So we wrote, as one of
the

other projects part of RACK-AMQP, is AMQParty,
which is

just a API-compliant version of HTTParty.
So we can

actually just drop in the AMQParty constant
in place

of the HTTParty constant, and you can see
down

below, that transport variable, we typically
configure that at

boot time to be either HTTParty or AMQParty.
And

everything, again, just works.

We change the, the url a little bit, too.

But we do that all part of the, part

of this setup for each server. For each client.

Of course, if you don't want to use HTTParty,

some people like the various other options
that are

out there, we also are publishing RACK-AMQP
client, which

is what H, AMQParty is built on. It, it's

a very simple API. You get a client, you

tell it which Rabbit to connect to. You send

a request to your target queue with your uri

as part of it, with any, you know, your

HTTP method and all that. You get back your

response. It's synchronous. It's simple. It's
actually built on

Bunny, which is a really great AMQP gem out

there, and that also is very easy to use.

Additionally, we're publishing a sample SOA
using Rails. It's

a work-in-progress at the moment. BUt the
userland and

userland_client are mostly built as, you can
see where

things are at the mo- at the current state.

It kind of gives you an idea about how

we think about SOA at Optoro, and how you

might be able to use Jackalope and the RACK-AMQP

project in your own projects.

So userland is a, is Rails service that essentially

publishes a user's concept out into the world.
And

then userland_client is a gem that consumes
the userland

service and interface uses the userland gem
to talk

to the userland service.

And so then, you know, how fast is this?

Well, there's this bench mark I've got. I'll
show

you it in just a sec. The only weird

thing we have to do here is tell AMQParty

which Rabbit to talk to, and you kind of

saw that in the, in the client gem a

little bit, so that's a little bit of setup

we do. Again, like I said, when we make

the decision about which, whether HTTParty
or AMQParty. So

we tell it which Rabbit. And then, inside,
we,

five hundred times, request this JSON using
AMQParty and

then five hundred times request the same thing
using

HTTParty.

So here, I'll show you that in, in use.

Let's see. OK.

Let's make this big. All right. So here I'm

gonna boot the app using Unicorn. So Unicorn
is

listing on port 8080. Wow. That's really big.
Trust

me, it's 8080 there. I'm gonna boot the same

map using Jackalope. The only difference I
have to

do to boot the app is I have to

say with the right queue name, so it's going

to listen to the test dot simple queue.

So there it is running. And then down here

I'm gonna starty the bench mark. And like
I

said, it's gonna five hundred times hit each
service.

What it actually does is hit each service
three

times, just to make sure that it's warm, and

then it can go through and hit it all

the way through. And there we go.

So you can see, it's two point nine seconds

total to do five hundred AMQP requests, and
three

point six seconds total to do five hundred
HTTP

requests. I was actually surprised when I
got this

result. Jackalope is beta. I mean, it's, we're
using

it at Optoro, but it's not really been battle-tested.

But it's faster, at the moment. So why is

it faster?

Well, maybe I haven't written all the code
yet.

But more likely it, more likely it's because
of

this concept in TCP called the slow-start
phenomonon. And

also, your whole TCP negotiation. So each
one of

these five hundred HTTP requests, close the
connection, open

the connection, send the message, receive
the response. When

you open the TCP connection you go through
a

whole sinaq phase, which I didn't put slides
in

for and maybe should have. But it's.

I highly recommend checking out Eliot Gregorich's
talks, by

the way, about how HTTP works, how TCP works.

But basically, you go through protocol negotiation
for any

TCP connection, and then there's a feature
of TCP

called the slow-start feature, which essentially
allows you to

negotiate your connection speed. And that's
really great when

you're talking over the internet, and you're
not sure

of the latency or the availability of routers
to

be decent along the way. But when you're doing

internal applications, it just gets in your
way. It

just slows you down.

So AMQP, we hold a persistent connection to
the

AMQP broker. We do one sinaq cycle, and then

we get our maximum connection speed and we
stay

there. And then we can multiplex information
over that

single connection and it works really well.
Now, you

could write that with HTTP if you did a

keep-alive connection, held over that connection
and talk to

the same thing back and forth.

But, the bottom line is that most of us

don't do that when we write HTTP. We're using

something like HTTParty or, I always forget
the tiforce

or tif- whatever it is. Sorry.

Most of the time, you know, we drop the

connection, you know, in between requests.
We don't hold

it open. It's just not something we think
about.

So using AMQParty or any of the libraries
that

we're published, it holds that open and it
just

makes it more efficient overall.

So I hope that's the primary explanation of
why

it's so much faster. So I just wanted to

mention a few references, a few things that
were

really inspriring when we kind of went down
this

road. You know, this is really a departure
from

what we're used to in Rails. And the Architecture:

The Lost Years talk and Matt Wynne's original
Hexegonal

Rails talks were all about departing from
what we're

used to, I think. So those are really interesting.

If you haven't checked those out, definitely
do.

Additionally, Martin Fowler posted a recent
article that was

awesome, called Micro Services. Definitely,
definitely read this if

you are looking into service-oriented architectures.
It lets you

not think about that heavy-weight bloat that
we think

about with SOA.

It's really a revisitation of the idea of,
just,

you have independent services that are collaborating
to get

something done in a very simple way. And that's

really, again, also influenced the direction
we're taking here.

We want to introduce as few specific choices
as

possible to just kind of give us a simple

transport to, to build on top of. Additionally,
the

Ruby AMQP gem has some awesome docs about
how

AMQP works, how to use it with Ruby. And

RabbitMQ also publishes some really great
documentation on what

AMQP is, how it works. How to use it

in the organization in your applications.

And there's also this really great article
I found

about how HTTP works, in general. It's, I
think

a piece of school material from a university
in

Singapore. Definitely check it out. It's got
really great

diagrams.

Thanks. Again, I'm Josh. I'm at jszmajda on
the

Twitters. I am the CTO at Optoro. We have

blinq dot com. We have retailers with their
returned

goods and excess goods, help them figure out
what

it is, what the current value is and move

it quickly.

Additionally, I am the host of the Ruby hangout.

Just a quick blurb for the Ruby hangout, it's

a online meetup. It's really handy when you
can

go face-to-face and talk to people if you
have

a local meet up. If you don't have a

local meet up or if you're too busy or

if you can't make it for whatever reason,
the

Ruby hangout is an online meet up to, essentially,

give you that local meet up feel, as much

as we can online. So check that out. It's

a lot of fun. I'm one of the co-organizers

of the DC Ruby Users group. And I want

to give a quick shoutout to Jonathan. Thanks
Jon.

He's on my team, helped me get Jackalope over

the, over the finish line. So super helpful
on

that.

All the code's on GitHub. Github dot com slash

rack dash amqp slash rack dash amqp is where

the RC lives and has bookmarks to the other

parts of the project. And I would love to

take your questions, because that talk was
way short.

And I've got, actually, a whole bunch of stuffed

Jackalopes up here for good questions. So
please feel

free to come grab one.