aiomsg 2019.4.12

Socket-based abstraction for messaging patterns

aiomsg

Pure-Python smart sockets (like ZMQ) for simple microservices architecture

Warning

⚠️ Don’t use this! Use ZeroMQ instead. aiomsg is currently a hobby project, whereas ZeroMQ is a mature messaging library that has been battle-tested for well over a decade!

Warning

⚠️ Right now this is in ALPHA. I’m changing stuff all the time. Don’t depend on this library unless you can handle API breakage between releases. Your semver has no power here, this is calver country. When I’m happy with the API I’ll remove this warning.

Table of Contents

Contents

• aiomsg

  • Table of Contents

• Demo

• Inspiration

• Introduction

• Cookbook

  • Publish from either the bind or connect end

  • Distribute messages to a dynamically-scaled service (multiple instances)

  • Distribute messages from a 2-instance service to a dynamically-scaled one

  • Distribute messages from one dynamically-scaled service to another

  • Two dynamically-scaled services, with a scaled fan-in, fan-out proxy

  • FAQ

    • Why do you spell Søcket like that?

    • I want to talk to the aiomsg Søcket with a different programming language

Demo

Let’s make two microservices; one will send the current time to the other. Here’s the end that binds to a port (a.k.a, the “server”):

import asyncio, time
from aiomsg import Søcket

async def main():
    async with Søcket() as sock:
        await sock.bind('127.0.0.1', 25000):
        while True:
            await s.send(time.ctime().encode())

asyncio.run(main())

Running as a different process, here is the end that does the connecting (a.k.a, the “client”):

import asyncio
from aiomsg import Søcket

async def main():
    async with Søcket() as sock
        await sock.connect('127.0.0.1', 25000):
        async for msg in sock.messages():
            print(msg.decode())

asyncio.run(main())

Note that these are both complete, runnable programs, not fragments.

Looks a lot like conventional socket programming, except that these sockets have a few extra tricks. These are described in more detail further down in rest of this document.

Inspiration

Looks a lot like ZeroMQ yes? no? Well if you don’t know anything about ZeroMQ, that’s fine too. The rest of this document will assume that you don’t know anything about ZeroMQ. aiomsg is heavily influenced by ZeroMQ.

There are some differences; hopefully they make things simpler than zmq. For one thing, aiomsg is pure-python so no compilation step is required, and relies only on the Python standard library (and that won’t change).

Also, we don’t have special kinds of socket pairs like ZeroMQ has. There is only the one Søcket class. The only role distinction you need to make between different socket instances is this: some sockets will bind and others will connect.

This is the leaky part of the API that comes from the underlying BSD socket API. A bind socket will bind to a local interface and port. A connect socket must connect to a bind socket, which can be on the same machine or a remote machine. This is the only complicated bit. You must decide, in a distributed microservices architecture, which sockets must bind and which must connect. A useful heuristic is that the service which is more likely to require horizontal scaling should have the connect sockets. This is because the hostnames to which they will connect (these will be the bind sockets) will be long-lived.

Introduction

What you see above in the demo is pretty much a typical usage of network sockets. So what’s special about aiomsg? These are the high-level features:

1. Messages, not streams:

   Send and receive are message-based, not stream based. Much easier! This does mean that if you want to transmit large amounts of data, you’re going to have have to break them up yourself, send the pieces, and put them back together on the other side.

2. Automatic reconnection

   These sockets automatically reconnect. You don’t have to write special code for it. If the bind end (a.k.a “server”) is restarted, the connecting end will automatically reconnect. This works in either direction. Try it! run the demo code and kill one of the processes. And then start it up again. The connection will get re-established.

3. Many connections on a single “socket”

   The bind end can receive multiple connections, but you do all your .send() and .recv() calls on a single object. (No callback handlers or protocol objects.)

   More impressive is that the connecting end is exactly the same; it can make outgoing connect() calls to multiple peers (bind sockets), and you make all your send() and recv() calls on a single object.

   This will be described in more detail further on in this document.

4. Message distribution patterns

   Receiving messages is pretty simple: new messages just show up (remember that messages from all connected peers come through the same call):

   async with Søcket() as sock
       await sock.bind():
       async for msg in sock.messages():
           print(f"Received: {msg}")

   However, when sending messages you have choices. The choices affect which peers get the message. The options are:

   • Publish: every connected peer is sent a copy of the message

   • Round-robin: each connected peer is sent a unique message; the messages are distributed to each connection in a circular pattern.

   • By peer identity: you can also send to a specific peer by using its identity directly.

   The choice between pub-sub and round-robin must be made when creating the Søcket():

   from aiomsg import Søcket, SendMode
   
   async with Søcket(send_mode=SendMode.PUBLISH) as sock
       await sock.bind():
       async for msg in sock.messages():
           await sock.send(msg)

   This example receives a message from any connected peer, and sends that same message to every connected peer (including the original sender). By changing PUBLISH to ROUNDROBIN, the message distribution pattern changes so that each “sent” message goes to only one connected peer. The next “sent” message will go to a different connected, and so on.

   For identity-based message sending, that’s available any time, regardless of what you choose for the send_mode parameter; for example:

   import asyncio
   from aiomsg import Søcket, SendMode
   
   async def main():
       async with Søcket() as sock1, Søcket(send_mode=SendMode.PUBLISH) as sock2:
           await sock1.bind(port=25000)
           await sock2.bind(port=25001)
           while True:
               peer_id, message = await sock1.recv_identity()
               msg_id, _, data = msg.partition(b"\x00")
               await sock2.send(data)
               await sock1.send(msg_id + b"\x00ok", identity=peer_id)
   
   asyncio.run(main())

   This example shows how you can receive messages on one socket (sock1, which could have thousands of connected peers), and relay those messages to thousands of other peers connected on a different socket (sock2).

   For this example, the send_mode of sock1 doesn’t matter because if identity is specified in the send() call, it’ll ignore send_mode completely.

   Oh, and the example above is a complete, runnable program which is pretty amazing!

5. Built-in heartbeating

   Because ain’t nobody got time to mess around with TCP keepalive settings. The heartbeating is internal and opaque to your application code. You won’t even know it’s happening, unless you enable debug logs. Heartbeats are sent only during periods of inactivity, so they won’t interfere with your application messages.

   In theory, you really shouldn’t need heartbeating because TCP is a very robust protocol; but in practice, various intermediate servers and routers sometimes do silly things to your connection if they think a connection has been idle for too long. So, automatic heartbeating is baked in to let all intermediate hops know you want the connection to stay up, and if the connection goes down, you will know much sooner than the standard TCP keepalive timeout duration (which can be very long!).

   If either a heartbeat or a message isn’t received within a specific timeframe, that connection is destroyed. Whichever peer is making the connect() call will then automatically try to reconnect, as discussed earlier.

6. Built-in reliability choices

   Ah, so what do “reliability choices” mean exactly…?

   It turns out that it’s quite hard to send messages in a reliable way. Or, stated another way, it’s quite hard to avoid dropping messages: one side sends and the other side never gets the message.

   aiomsg already buffers messages when being sent. Consider the following example:

   from aiomsg import Søcket, SendMode
   
   async with Søcket(send_mode=SendMode.PUBLISH) as sock
       await sock.bind():
       while True:
           await sock.send(b'123)
           await asyncio.sleep(1.0)

   This server above will send the bytes b"123" to all connected peers; but what happens if there are no connected peers? In this case the message will be buffered internally until there is at least one connected peer, and when that happens, all buffered messages will immediately be sent. To be clear, you don’t have to do anything extra. This is just the normal behaviour, and it works the same with the ROUNDROBIN send mode.

   Message buffering happens whenever there are no connected peers available to receive a message. Sounds great right? Unfortunately, this is not quite enough to prevent messages from getting lost. It is still easy to have your process killed immediately after sending data into a kernel socket buffer, but right before the bytes actually get transmitted. In other words, your code thinks the message got sent, but it didn’t actually get sent.

   The only real solution for adding robustness is to have peers reply to you saying that they received the message. Then, if you never receive this notification, you should assume that the message might not have been received, and send it again. aiomsg will do this for you (so again there is no work on your part), but you do have to turn it on.

   This option is called the DeliveryGuarantee. The default option, which is just basic message buffering in the absence of any connected peers, is called DeliveryGuarantee.AT_MOST_ONCE. It means, literally, that any “sent” message will received by a connected peer no more than once (of course, it may also be zero, as described above).

   The alternative is to set DeliveryGuarantee.AT_LEAST_ONCE, which enables the internal “retry” feature. It will be possible, under certain conditions, that any given message could be received more than once, depending on timing and situation. This is how the code looks if you enable it:

   from aiomsg import Søcket, SendMode, DeliveryGuarantee
   
   async with Søcket(
           send_mode=SendMode.ROUNDROBIN,
           delivery_guarantee=DeliveryGuarantee.AT_LEAST_ONCE
   ) as sock:
       await sock.bind():
       while True:
           await sock.send(b'123)
           await asyncio.sleep(1.0)

   It’s pretty much exactly the same as before, but we added the AT_LEAST_ONCE option. Note that AT_LEAST_ONCE does not work for the PUBLISH sending mode. (Would it make sense to enable?)

   As a minor point, you should note that when AT_LEAST_ONCE is enabled, it does not mean that every send waits for acknowledgement before the next send. That would incur too much latency. Instead, there is a “reply checker” that runs on a timer, and if a reply hasn’t been received for a particular message in a certain timeframe (5.0 seconds by default), that message will be sent again.

   The connection may have gone down and back up within those 5 seconds, and there may be new messages buffered for sending before the retry send happens. In this case, the retry message will arrive after those buffered messages. This is a long way of saying that the way that message reliability has been implemented can result in messages being received in a different order to what they were sent. In exchange for this, you get a lower overall latency because sending new messages is not waiting on previous messages getting acknowledged.

7. Pure python, doesn’t require a compiler

8. Depends only on the Python standard library

Cookbook

The message distribution patterns are what make aiomsg powerful. It is the way you connect up a whole bunch of microservices that brings the greatest leverage. We’ll go through the different scenarios using a cookbook format.

In the code snippets that follow, you should assumed that each snippet is a complete working program, except that some boilerplate is omitted. This is the basic template:

import asyncio
from aiomsg import Søcket, SendMode, DeliveryGuarantee

<main() function>

asyncio.run(main())

Just substitute in the main() function from the snippets below to make the complete programs.

Publish from either the bind or connect end

The choice of “which peer should bind” is unaffected by the sending mode of the socket.

Compare

# Publisher that binds
async def main():
    async with Søcket(send_mode=SendMode.PUBLISH) as sock:
        await sock.bind():
        while True:
            await sock.send(b'News!')
            await asyncio.sleep(1)

versus

# Publisher that connects
async def main():
    async with Søcket(send_mode=SendMode.PUBLISH) as sock:
        await sock.connect():
        while True:
            await sock.send(b'News!')
            await asyncio.sleep(1)

The same is true for the round-robin sending mode. You will usually choose the bind peer based one which service is least likely to require dynamic scaling. This means that the mental conception of socket peers as either a server or client is not that useful.

Distribute messages to a dynamically-scaled service (multiple instances)

In this recipe, one service needs to send messages to another service that is horizontally scaled.

The trick here is that we don’t want to use bind sockets on horizontally-scaled services, because other peers that need to make a connect call will need to know what hostname to use. Each instance in a horizontally-scaled service has a different IP address, and it becomes difficult to keep the “connect” side up-to-date about which peers are available. This can also change as the horizontally-scaled service increases or decreases the number of instances. (In ZeroMQ documentation, this is described as the Dynamic Discovery Problem).

aiomsg handles this very easily: just make sure that the dynamically-scaled service is making the connect calls:

This is the manually-scaled service (has a specific domain name):

# jobcreator.py -> DNS for "jobcreator.com" should point to this machine.
async def main():
    async with Søcket(send_mode=SendMode.ROUNDROBIN) as sock:
        await sock.bind(hostname="0.0.0.0", port=25001)
        while True:
            await sock.send(b"job")
            await asyncio.sleep(1)

These are the downstream workers (don’t need a domain name):

# worker.py - > can be on any number of machines
async def main():
    async with Søcket() as sock
        await sock.connect(hostname='jobcreator.com', port=25001):
        while True:
            work = await sock.recv()
            <do work>

With this code, after you start up jobcreator.py on the machine to which DNS resolves the domain name “jobcreator.com”, you can start up multiple instances of worker.py on other machines, and work will get distributed among them. You can even change the number of worker instances dynamically, and everything will “just work”, with the main instance distributing work out to all the connected workers in a circular pattern.

This core recipe provides a foundation on which many of the other recipes are built.

Distribute messages from a 2-instance service to a dynamically-scaled one

In this scenario, there are actually two instances of the job-creating service, not one. This would typically be done for reliability, and each instance would be placed in a different availability zones. Each instance will have a different domain name.

It turns out that the required setup follows directly from the previous one: you just add another connect call in the workers.

The manually-scaled service is as before, but you start on instance of jobcreator.py on machine “a.jobcreator.com”, and start another on machine “b.jobcreator.com”. Obviously, it is DNS that is configured to point to the correct IP addresses of those machines (or you could use IP addresses too, if these are internal services).

# jobcreator.py -> Configure DNS to point to these instances
async def main():
    async with Søcket(send_mode=SendMode.ROUNDROBIN) as sock:
        await sock.bind(hostname="0.0.0.0", port=25001)
        while True:
            await sock.send(b"job")
            await asyncio.sleep(1)

As before, the downstream workers, but this time each worker makes multiple connect() calls; one to each job creator’s domain name:

# worker.py - > can be on any number of machines
async def main():
    async with Søcket() as sock:
        await sock.connect(hostname='a.jobcreator.com', port=25001)
        await sock.connect(hostname='b.jobcreator.com', port=25001)
        while True:
            work = await sock.recv()
            <do work>

aiomsg will return work from the sock.recv() call above as it comes in from either job creation service. And as before, the number of worker instances can be dynamically scaled, up or down, and all the connection and reconnection logic will be handled internally.

Distribute messages from one dynamically-scaled service to another

If both services need to be dynamically-scaled, and can have varying numbers of instances at any time, we can no longer rely on having one end do the socket bind to a dedicated domain name. We really would like each to make connect() calls, as we’ve seen in previous examples.

How to solve it?

The answer is to create an intermediate proxy service that has two bind sockets, with long-lived domain names. This is what will allow the other two dynamically-scaled services to have a dynamic number of instances.

Here is the new job creator, whose name we change to dynamiccreator.py to reflect that it is now dynamically scalable:

# dynamiccreator.py -> can be on any number of machines
async def main():
    async with Søcket(send_mode=SendMode.ROUNDROBIN) as sock:
        await sock.connect(hostname="proxy.jobcreator.com", port=25001)
        while True:
            await sock.send(b"job")
            await asyncio.sleep(1)

Note that our job creator above is now making a connect() call to proxy.jobcreator.com:25001 rather than binding to a local port. Let’s see what it’s connecting to. Here is the intermediate proxy service, which needs a dedicated domain name, and two ports allocated for each of the bind sockets.

# proxy.py -> Set up DNS to point "proxy.jobcreator.com" to this instance
async def main():
    async with Søcket() as sock1, \
            Søcket(send_mode=SendMode.ROUNDROBIN) as sock2:
        await sock1.bind(hostname="0.0.0.0", port=25001)
        await sock2.bind(hostname="0.0.0.0", port=25002)
        while True:
            work = await sock1.recv()
            await sock2.send(work)

Note that sock1 is bound to port 25001; this is what our job creator is connecting to. The other socket, sock2, is bound to port 25002, and this is the one that our workers will be making their connect() calls to. Hopefully it’s clear in the code that work is being received from sock1 and being sent onto sock2. This is pretty much a feature complete proxy service, and with only minor additions for error-handling can be used for real work.

For completeness, here are the downstream workers:

# worker.py - > can be on any number of machines
async def main():
    async with Søcket() as sock:
        await sock.connect(hostname='proxy.jobcreator.com', port=25002)
        while True:
            work = await sock.recv()
            <do work>

Note that the workers are connecting to port 25002, as expected.

You might be wondering: isn’t this just moving our performance problem to a different place? If the proxy service is not scalable, then surely that becomes the “weakest link” in our system architecture?

This is a pretty typical reaction, but there are a couple of reasons why it might not be as bad as you think:

1. The proxy service is doing very, very little work. Thus, we expect it to suffer from performance problems only at a much higher scale compared to our other two services which are likely to be doing more CPU-bound work (in real code, not my simple examples above).

2. We could compile only the proxy service into faster low-level code using any number of tools such as Cython, C, C++, Rust, D and so on, in order to improve its performance, if necessary (this would require implementing the aiomsg protocols in that other language though). This allows us to retain the benefits of using a dynamic language like Python in the dynamically scaled services where much greater business logic is captured (these can be then be horizontally scaled quite easily to handle performance issues if necessary).

3. Performance is not the only reason services are dynamically scaled. It is always a good idea, even in low-throughput services, to have multiple instances of a service running in different availability zones. Outages do happen, yes, even in your favourite cloud provider’s systems.

4. A separate proxy service as shown above isolates a really complex problem and removes it from your business logic code. It might not be easy to appreciate how significant that is. As your dev team is rapidly iterating on business features, and redeploying new versions several times a day, the proxy service is unchanging, and doesn’t require redeployment. In this sense, it plays a similar role to more traditional messaging systems like RabbitMQ and ActiveMQ.

5. We can still run multiple instances of our proxy service using an earlier technique, as we’ll see in the next recipe.

Two dynamically-scaled services, with a scaled fan-in, fan-out proxy

This scenario is exactly like the previous one, except that we’re nervous about having only a single proxy service, since it is a single point of failure. Instead, we’re going to have 3 instances of the proxy service running in parallel.

Let’s jump straight into code. The proxy code itself is actually unchanged from before. We just need to run more copies of it on different machines. Each machine will have a different domain name.

# proxy.py -> unchanged from the previous recipe
async def main():
    async with Søcket() as sock1, \
            Søcket(send_mode=SendMode.ROUNDROBIN) as sock2:
        await sock1.bind(hostname="0.0.0.0", port=25001)
        await sock2.bind(hostname="0.0.0.0", port=25002)
        while True:
            work = await sock1.recv()
            await sock2.send(work)

For the other two dynamically scaled services, we need to tell them all the domain names to connect to. We could set that up in an environment variable:

$ export PROXY_HOSTNAMES="px1.jobcreator.com;px2.jobcreator.com;px3.jobcreator.com"

Then, it’s really easy to modify our services to make use of that. First, the dynamically-scaled job creator:

# dynamiccreator.py -> can be on any number of machines
async def main():
    async with Søcket(send_mode=SendMode.ROUNDROBIN) as sock:
        for proxy in os.environ['PROXY_HOSTNAMES'].split(";"):
            await sock.connect(hostname=proxy, port=25001)
        while True:
            await sock.send(b"job")
            await asyncio.sleep(1)

And the change for the worker code is identical (making sure the correct port is being used, 25002):

# worker.py - > can be on any number of machines
async def main():
    async with Søcket() as sock:
        for proxy in os.environ['PROXY_HOSTNAMES'].split(";"):
            await sock.connect(hostname=proxy, port=25002)
        while True:
            work = await sock.recv()
            <do work>

Three proxies, each running in a different availability zone, should be adequate for most common scenarios.

TODO: more scenarios involving identity (like ROUTER-DEALER)

FAQ

Why do you spell Søcket like that?

The slashed O is used in homage to ØMQ, a truly wonderful library that changed my thinking around what socket programming could be like. Why would you use HTTP between backend systems when you could use this!

I want to talk to the aiomsg Søcket with a different programming language

WARNING: This section is extremely provisional. I haven’t fully nailed down the protocol yet.

To make a clone of the Søcket in another language is probably a lot of work, but it’s actually not necessary to implement everything.

You can talk to aiomsg sockets quite easily by implementing the simple protocol described below. It would be just like regular socket programming in your programming language. You just have to follow a few simple rules for the communication protocol.

These are the rules:

1. Every payload in either direction shall be length-prefixed:

   message = [4-bytes big endian int32] [payload]

2. Immediately after successfully opening a TCP connection, before doing anything else with your socket, you shall:

   • Send your identity, as a 16 byte unique identifier (a 16 byte UUID is perfect). Note that Rule 1 still applies, so this would look like

     identity_message = b'\x00\x00\x00\x10' + [16 bytes]

     (because the length, 16, is 0x10 in hex)

   • Receive the other peer’s identity (16 bytes). Remember Rule 1.

3. You shall periodically send a heartbeat message b"aiomsg-heartbeat". Every 5 seconds is good. If you receive such messages you can ignore them. If you don’t receive one (or an actual data message) within 15 seconds of the previous receipt, the connection is probably dead and you should kill it and/or reconnect. could either ignore it, or reply with exactly the same. Note that Rule 1 still applies, and because the length of this message is also 16 bytes, the message is ironically similar to the identity message:

   heartbeat_message = b'\x00\x00\x00\x10' + b'aiomsg-heartbeat'

After you’ve satisfied these rules, from that point on every message sent or received is a Rule 1 message, i.e., length prefixed with 4 bytes for the length of the payload that follows.

If you want to run a bind socket, and receive multiple connections from different aiomsg sockets, then the above rules apply to each separate connection.

That’s it!

TODO: Discuss the protocol for AT_LEAST_ONCE mode, which is a bit messy at the moment.