An asynchronous CherryPy server based on asyncio

CherryPy is a minimalist web application server written in Python. Hundreds of people have relied on it for more than fourteen years now. Recently, I’ve gained interest in the native asynchronous support Python has gained through the implementation of PEP-3156 and PEP-0492. Basically, Python now supports coroutines natively and in a friendly API interface. In addition, thanks to the built-in asyncio module, you can naturally develop concurrent applications.

CherryPy has always used a multi-threaded engine to support concurrent web applications. This may sound surprising, but this is working very well still in 2016. Yet, I love a puzzle and I was interested in making CherryPy run as a set of coroutines rather than a bunch of threads. So a couple of days ago, I set myself on the task to make it happen.

And so here it is, CherryPy on asyncio!

This a branch on my fork, not an official part of the CherryPy project yet.

Now, you can run code like this:

The only differences are:

  • importing cherrypy.async
  • turning page-handlers into coroutines

That is all.

The idea is that the cherrypy.async patches all the internals of CherryPy for you and turn it into an async-aware server.

Note that the code currently runs only on Python 3.5+ as we use the async/await keywords.

Has it been easy?

Turning a project that was not designed for coroutines is not that complicated thanks to the simple interface provided by async/await. However, anytime an I/O operation is performed, it is necessary to transform a call like:


This mundane line is sometimes is part of a function call, in that case, you have to overwrite and copy/paste the whole function to change that one line. Mind you, this can’t be avoided because you must also re-declare the function as a coroutine anyway by prefixing it with async.

As usual, the difficulty lies in the entanglement of your code. The more you made simple to comprehend, the simpler and faster it will be to change with confidence.

What was changed?

Mostly the HTTP server, the machinery of CherryPy: the internal engine/bus, the dispatcher, the request handling.

If we can find a way to re-organise the existing code, actually few lines would eventually be changed.

Note, I made the decision not to make this server WSGI aware because I find that rather counter-intuitive with an async-based server somehow.

Is it production ready?

Not at all. It hasn’t been really tested (this will require to re-write many tests so that they play along with coroutines).

It is also, for some unknown reason yet, much slower than the multithreaded version. Profiling will need to be performed.

Still, if you are feeling like testing it:

I don’t know if this code will go further than this but, maybe this will interest the community enough so that it moves forward. This would make CherryPy more suitable for HTTP2 and websockets.

Deploying a docker container of a CherryPy application onto a CoreOS cluster

Previously, I presented a simple web application that was distributed into several docker containers. In this article, I will be introducing the CoreOS platform as the backend for clusterizing a CherryPy application.

CoreOS quick overview

CoreOS is a Linux distribution designed to support distributed/clustering scenarios. I will not spend too much time explaining it here as their documentation already provides lots of information. Most specifically, review their architecture use-cases for a good overview of what CoreOS is articulated.

What matters to us in this article is that we can use CoreOS to manage a cluster of nodes that will host our application as docker containers. To achieve this, CoreOS relies on a technologies such as systemd, etcd and fleet at its core.

Each CoreOS instance within the cluster runs a linux kernel which executes systemd to manage processes within that instance. etcd is a distributed key/value store used across the cluster to enable service discovery and configuration synchronization within the cluster. Fleet is used to manage services executed within your cluster. Those services aredescribed in files called unit files.

Roughly speaking, you use a unit-file to describe your service and specify which docker container to execute. Using fleet, you submit and load that service to the cluster before starting/stopping it at will. CoreOS will determine which host it will deploy it on (you can setup constraints that CoreOS will follow). Once loaded onto a node, the node’s systemd takes over to manage the service locally and you can use fleet to query the status of that service from outside.

Setup your environment with Vagrant

Vagrant is a nifty tool to orchestrate small deployment on your development machine. For instance, here is a simple command to create a node with Ubuntu running on it:

Vagrant has a fairly rich command line you can script to generate a final image. However, Vagrant usually provisions virtual machines by following a description found within a simple text file (well actually it’s a ruby module) called a Vagrantfile. This is the path we will be following in this article.

Let’s get the code:

From there you can create the cluster as follows:

I am not using directly vagrant to create the cluster because there are a couple of other operations that must be carried to let fleet talk to the CoreOS node properly. Namely:

  • Generate a new cluster id (via
  • Start a ssh agent to handle the node’s SSH identities to connect from the outside
  • Indicate where to locate the node’s ssh service (through a port mapped by Vagrant)
  • Create the cluster (this calls vagrant up internally)

Once completed, you should have a running CoreOS node that you can log into:

To destroy the cluster and terminate the node:

This also takes care of wiping out local resources that we don’t need any longer.

Before moving on, you will need to install the fleet tools.

Run your CherryPy application onto the cluster

If you have destroyed the cluster, re-create it and make sure you can speak to it through fleet as follows:

Bingo! This is the public address we statically set in the Vagrantfile associated to the node.

Let’s ensure we have no registered units yet:

Okay, all is good. Now, let’s push each of our units to the cluster:

As you can see, the unit files have been registered but they are not loaded onto the cluster yet.

Notice the naming convention used for webapp_app@.service, this is due to the fact that this is will not be considered as a service description itself but as a template for a named service. We will see this in a minute. Refer to this extensive DigitalOcean article for more details regarding unit files.

Let’s now load each unit onto the cluster:

Here, we asked fleet to load the service onto an available node. Considering there is a single node, it wasn’t a a difficult decision to make.

At that stage, your service is not started. It simply is attached to a node.

It is not compulsory to explicitely load before starting a service. However, if gives you the opportunity to unload a service if a specific condition occurs (service needs to be amended, the chosen host isn’t valid any longer…).

Now ce can finally start it:

You can see what’s happening:

Or alternatively, you can request the service’s status:

Once the service is ready:

Starting a service from a unit template works the same way except you provide an identifier to the instance:

The reason I chose 1 as the identifier is so that it the container’s name becomes notes1 as expected by the load-balancer container when linking it to the application’s container. As described in the previous article.

Start a second instance of that unit template:

That second instance starts immediatly because the image is already there.

Finally, once both services are marked as “active”, you can start the load-balancer service as well:

At that stage, the complete application is up and running and you can go to http://localhost:7070/ to use it. Port 7070 is mapped to port 8091 by vagrant within our Vagrantfile.

No such thing as a free lunch

As I said earlier, we created a cluster of one node on purpose. Indeed, the way all our containers are able to dynamically know where to locate each other is through the linking mechanism. Though this works very well in simple scenarios like this one, this has a fundamental limit since you cannot link across different hosts. If we had multiple nodes, fleet would try distributing our services accross all of them (unless we decided to constraint this within the unit files) and this would break the links between them obviously. This is why, in this particular example, we create a single node’s cluster.

Docker provides a mechanism named ambassador to address this restriction but we will not review it, instead we will benefit from a flat sub-network topology provided by weave as it seems it follows a more traditional path than the docker’s linking approach. This will be the subject of my next article.

A more concrete example of a complete web application with CherryPy, PostgreSQL and haproxy

In the previous post, I described how to setup a docker image to host your CherryPy application. In this installment, I will present a complete – although simple – web application made of a database, two web application servers and a load-balancer.

Setup a database service

We are going to create a docker image to host our database instance, but because we are lazy and because it has been done already, we will be using an official image of PostgreSQL.

As you can see, we run the official, latest, PostgreSQL image. By setting the POSTGRES_USER and POSTGRES_PASSWORD, we make sure the container creates the according account for us. We also set a name for this container, this will be useful when we link to it from another container as we will see later on.

A word of warning, this image is not necessarily secure. I would advise you to consider this question prior to using it in production.

Now that the server is running, let’s create a database for our application. Run a new container which will execute the psql shell:

We have connected to the server, we then create the “notes” database and connect to it.

How did this work? Well, the magic happens through the –link wedb:postgres we provided to the run command. This tells the new container we are linking to a container named webdb and that we create an alias for it inside that new container. That alias is used by docker to initialize a few environment variables such as:

Notice the POSTGRES_ prefix? This is exactly the alias we gave in the command’s argument. This is the mechanism by which you will link your containers so that they can talk to each other.

Note that there are alternatives, such as weave, that may be a little more complex but probably more powerful. Make sure to check them out at some point.

Setup our web application service

We are going to run a very basic web application. It will be a form to take notes. The application will display them and you will be able to delete each note. The notes are posted via javascript through a simple REST API. Nothing fancy. Here is a screenshot for you:


By the way, the application uses Yahoo’s Pure.css framework to change from bootstrap.

Simply clone the mercurial repository to fetch the code.

This will download the whole repository but fear not, it’s rather lightweight. You can review the Dockerfile which is rather similar to what was described in my previous post. Notice how we copy the webapp subdirectory onto the image.

We can now create our image from that directory:

As usual, change the tag to whatever suits you.

Let’s now run two containers from that image:

We link those two containers with the container running our database. We can therefore use that knowledge to connect to the database via SQLAlchemy. We also publish the application’s port to two distinct ports on the host. Finally, we name our containers so that can we reference them in the next container we will be creating.

At this stage, you ought to see that your application is running by going either to http://localhost:8080/ or http://localhost:8081/.

Setup a load balancer service

Our last service – microservice should I say – is a simple load-balancer between our two web applications. To support this feature, we will be using haproxy. Well-known, reliable and lean component for such a task.

Tak some time to review the Dockerfile. Notice how we copy the local haproxy.cfg file as the configuration for our load-balancer. Build your image like this:

And now run it to start load balancing between your two web application containers:

In this case, we will be executing the container in the background because we are blocking on haproxy and it won’t lok to the console anyway.

Notice how we link to both web application containers. We set short alias just by pure lazyness. We publish two ports to the host. The 8090 port will be necessary to access the stats page of the haproxy server itself. The 8091 port will be used to access our application.

To understand how we reuse the the aliases, please refer to the the haproxy.cfg configuration. More precisely to those two lines:

We load-balance between our two backend servers and we do not have to know their address at the time when we build the image, but only when the container is started.

That’s about it really. At this stage, you ought to connect to http://localhost:8091/ to see use your application. Each request will be sent to each web application’s instances in turn. You may check the status of your load-balancing by connecting to http://localhost:8090/.

Obviously, this just a basic example. For instance, you could extend it by setting another service to manage your syslog and configure haproxy to send its log to it.

Next time, we will be exploring the world of CoreOS and clustering before moving on to service and resource management via Kubernetes and MesOS.

Create a docker container for your CherryPy application

In the past year, process isolation through the use of containers has exploded and you can find containers for almost anything these days. So why not creating a container to isolate your CherryPy application from the rest of the world?

I will not focus on the right and wrongs in undertaking such a task. This is not the point of this article. On the other hand, this article will guide you through the steps to create a base container image that will support creating per-project images that can be run in containers.

We will be using docker for this since it’s the hottest container technology out there. It doesn’t mean it’s the best, just that it’s the most popular which in turns means there is high demand for it. With that being said, once you have decided containers are a relevant feature to you, I encourage you to have a look at other technologies in that field to draw your own conclusion.

Docker uses various Linux kernel assets to isolate a process from the other running processes. In particular, it uses control groups to constraints the resources used by the process. Docker also makes the most of namespaces which create an access layer to resources such as network, mounted devices, etc.

Basically, when you use docker, you run an instance of an image and we call this a container. An image is mostly a mille-feuille of read-only layers that are eventually unified into one. When an image is run as a container, an extra read-write layer is added by docker so that you can make changes at runtime from within your container. Those changes are lost everytime you stop the running container unless you commit it into a new image.

So how to start up with docker?

Getting started

First of all, you must install docker. I will not spend much time here explaining how to go about it since the docker documentation does it very well already. However, iI encourage you to:

  • install from the docker repository as it’s more up to date usually than official distribution repositories
  • ensure you can run docker commands as a non-root user. This will make your daily usage of docker much easier

At the time of this writing, docker 1.4.1 is the latest version and this article was written using 1.3.3. Verify your version as follow:

Docker command interface

Docker is an application often executed as a daemon. To interact with it you use the command line interface via the docker command. Simply run the following command to see them:

Play a little with docker

Before we move on creating our docker image for a CherryPy application, lets play with docker.

The initial step is to pull an existing image. Indeed, you will likely not create your own OS image from scratch. Instead, you will use a public base image, available on the docker public registry. During the course of these articles, we will be using a Ubuntu base image. But everything would work the same wth Centos or something else.

Easy right? The various downloads are those of the intermediary images that were generated by the Ubuntu image maintainers. Interestingly, this means you could start your image from any of those images.

Now that you have an image, you may wish to list all of them on your machine:

Notice that the intermediate images are not listed here. To see them:

Note that, in the previous call we didn’t specify any specific version for our docker image. You may wish to do so as follow:

Let’s pull a centos image as well for the fun:

Let’s now run a container and play around with it:

In the previous command, we start a bash command executed within a container using the Centos image tagged 7. We name the container to make it easy to reference it afterwards. This is not compulsory but is quite handy in certain situations. We also tell docker that it can dispose of that container when we exit it. Otherwise, the container will remain.

This is interesting because it shows that, indeed, the container is executed in the host kernel which, in this instance, is my Ubuntu operating system.

Finally below, let’s see the network configuration:

Note that the eth0 interface is attached to the bridge the docker daemon created on the host. The docker security scheme means that, by default, nothing can reached that interface from the outside. However the docker may contact the outside world. Docker has an extensive documentation regarding its networking architecture.

Note that you can see containers statuses as follow:

Exit the container:

Run again the command:

As we can see the container is indeed gone. Let’s now rewind a little and do not tell docker to automatically remove the container when we exit it:

Let’s see if the container is there:

Nope. So what’s different? Well, try again to start a container using that same name:

Ooops. The container is actually still there:

There you go. By default docker ps doesn’t show you the containers in the exit status. You have to remove the container manually using its identifier:

I will not go further with using docker as it’s all you really need to start up with

A word about tags

Technically speaking, versions do not actually exist in docker images. They are in fact tags. A tag is a simple label for an image at a given point.

Images are identified with a hash value. As with IP addresses, you are not expected to recall the hash of the images you wish to use. Docker provides a mechanism to tag images much like you would use domain names instead of IP address.

For instance, 14.10 above is actually a tag, not a version. Obviously, since tags are meant to be meaningful to human beings, it’s quite sensible for Linux distributions to be tagged following the version of the distributions.

You can easily create tags for any images as we will see later on.

Let’s talk about registries

Docker images are hosted and served by a registry. Often as it’s the case in our previous example, the registry used is the public docker registry available at :

Whenever you pull an image from a registry, by default docker pulls from that registry. However, you may query a different registry as follow:

Basically, you provide the address of your registry and a path at which the image can be located. It has a similar form to an URI without the scheme.

Note that, as of docker 1.3.1, if the registry isn’t served over HTTPS, the docker client will refuse to download the image. If you need to pull anyway, you must add the following parameter to the docker daemon when it starts up.

Please refer to the official documentation to learn more about this.

A base Linux Python-ready container

Traditionnaly deploying CherryPy application has been done using a simple approach:

  • Package your application into an archive
  • Copy that archive onto a server
  • Configure a database server
  • Configure a reverse proxy such as nginx
  • Start the Python process(es) to server your CherryPy application

That last operation is usually done by directly calling nohup python &. Alternatively, CherryPy comes with a handy script to run your application in a slightly more convenient fashion:

This runs the Python module mymodule as a daemon using the given configuration file. If the -P flag isn’t provided, the module must be found in PYTHONPATH.

The idea is to create an image that will serve your application using cherryd. Let’s see how to setup an Ubuntu image to run your application.

First we create a user which will not have the root permissions. This is a good attitude to follow:

Next, we install a bunch of libraries that are required to deploy some common Python dependencies:

Then we create a virtual environment and install Python packages into it:

These are common packages I use. Install whichever you require obviously.

As indicated by Tony in the comments, it is probably overkill to create a virtual environment in a container since, the whole point of a container is to isolate your process and its dependencies already. I’m so used to using virtual env that I automatically created one. You may skip these steps.

Those operations were performed as the root user, let’s make the web user those packages owner.

Good. Let’s switch to that user now:

At this stage, we have a base image ready to support a CherryPy application. It might be interesting to tag that paricular container as a new image so that we can use it various contexts.

We take the docker container and we commit it as a new image. We then tag the new created image to make it easy to reuse it later on.

Let’s see if it worked. Exit the container and start a new container from the new image.

Well. We are ready to play now.

Run a CherryPy application in a docker container

For the purpose of this article, here is our simple application:

Two important points:

  • You must make sure CherryPy listens on the eth0 interface so just make it listen on all the container interfaces. Otherwise, the CherryPy will listen only on which won’t be reachable from outside the container.
  • Do not start the CherryPy engine yourself, this is done by the cherryd command. You must simply ensure the application is mounted so that CherryPy can serve it.

Save this piece of code into your container under the module name: This could be any name, really. The module will be located in /home/web.

You can manually test the module:

The second line tells us the IPv4 address of this container. Next point your browser to the following URL: http://localhost:9090/

“What is this magic?” I hear you say!

If you look at the command we use to start the container, we provide this bit: -p 9090:8080. This tells docker to map port 9090 on the host to port 8080 on the container alllowing for your application to be reached from the outside.

And voilà!

Make the process a little more developer friendly

In the previous section, we saved the application’s code into the container itself. During development, this may not be practical. One approach is to use a volume to share a directory between your host (where you work) and the container.

You can then work on your application and the container will see those changes immediatly.

Automate things a bit

The previous steps have shown in details how to setup an image to run a CherryPy application. Docker provides a simple interface to automate the whole process: Dockerfile.

A Dockerfile is a simple text file containing all the steps to create an image and more. Let’s see it first hand:

Create a directory and save the content above into a file named Dockerfile. Create a subdirectory called webapp and store your module into it.

Now, build the image as follow:

Use whatever tag suits you. Then, you can run a container like this:

That’s it! A docker container running your CherryPy application.

In the next articles, I will explore various options to use docker in a web application context. Follow ups will also include an introduction to weave and coreos to clusterize your CherryPy application.

In the meantime, do enjoy.

CherryPy documentation new start

Early on this year, a discussion emerged on the CherryPy mailing-list about the project. Most people said they loved the project but had struggled with its documentation. Though rich and extensive, it was felt it left down the project somehow by being not designed in a way that was attractive to new comers. I took upon myself to rewrite it from scratch following some ideas exchanged on the mailing-list.

The general expressed wish was to make it friendliers to people starting with the framework whilst making easy to look for common tasks and patterns. This suited me well as I wanted to carry the work I started on the various recipes I keep on BitBucket.

Eventually, I quickly wrote a set of tutorials to guide people through the general layout of a CherryPy application. Then I developed upon the recipes idea by going through many of the most recurrent questions we have on the mailing-list. Finally, I wrote an extensive section regarding the core features of the framework: plugins, tools, the bus, the dispatchers, etc. Those features are seldom used to their best even though they provide a very powerful backbone to design your application in clean way.

The documentation is now online and seems to have been well-received. It will need to be completed but I believe they already make the project much more appealing and fun to work with.

ws4py – WebSocket client and server library for Python

Recently I released ws4py, a package that provides client and server WebSocket support for Python 2.6 and 2.7.

Let’s first have a quick overview of what ws4py offers for now:

  • WebSocket specification draft-10 of the current specification.
  • A threaded client. This gives a simple client that doesn’t require an external dependency.
  • A Tornado client. This client is based on Tornado 2.0 which is quite a popular way of running asynchronous networking code these days. Tornado provides its own server implementation so I didn’t include mine in ws4py.
  • A CherryPy extension so that you can integrate WebSocket from within your CherryPy 3.2.1 server.
  • A gevent server based on the popular gevent library. This is courtesy of Jeff Lindsay.
  • Based on Jeff’s work, a pure WSGI middleware as well (available in the current master branch only until the next release).
  • ws4py runs on Android devices thanks to the SL4A package

Hopefully more client and servers will be added along the way as well as Python 3.x support. The former should be rather simple to add due to the way I designed ws4py.

The main idea is to make a distinction between the bytes provider and the bytes processing. The former is essentially reading and writing bytes from the connected socket. The latter is the function of making something out of the received bytes based on the WebSocket specification. In most implementations I have seen so far, both are rather heavily intertwined making it difficult to use a different bytes provider.

ws4py tries a different path by relying on a great feature of Python: the possibility to send data back to a generator. For instance, the frame parsing yields the quantity of bytes each time it needs more and the caller feeds back the generator those bytes once they are received. In fact, the caller of a frame parser is a stream object which acts the same way. The caller of that stream object is in fact the bytes provider (a client or a server). The stream is in charge of aggregating frames into a WebSocket message. Thanks to that design, both the frame and stream objects are totally unaware of the bytes provider and can be easily adapted in various contexts (gevent, tornado, CherryPy, etc.).

On my TODO list for ws4py:

  • Upgrade to a more recent version of the specification
  • Python 3.x implementation
  • Better documentation, read, write documentation.
  • Better performances on very large WebSocket messages

Acceptance testing a CherryPy application with Robot Framework

I recently received the Python Testing Cookbook authored by Greg L. Turnquist and was happy to read about recipes on acceptance testing using Robot Framework. We’ve been using this tool at work for a few weeks now with great results. Greg shows how to test a web application using the Selenium Library extension for Robot Framework and I thought it’d be fun to demonstrate how to test a CherryPy application following his recipe. So here we go.

First some requirements:

Let’s define a simple CherryPy application, which displays a input text where to type a message. When the submit button is pressed, the message is sent to the server and returned as-is. Well it’s an echo message really.

Save the code above in a module named

Next, we create an extension to Robot Framework that will manage CherryPy. Save the following in a module It’s important to respect that name since Robot Framework expects the module and its class to match in names.

Note that we start and stop the CherryPy server during the test itself, meaning you don’t need to start it separately. Pure awesomeness.

Finally let’s write a straightforward acceptance test to validate the overall workflow of echoing a message using our little application.

Save the test above into a file named testmyapp.txt. You can finally run the test as follow:

This will start CherryPy, Selenium’s proxy server and Firefox within which the test case will be run. Easy, elegant and powerful.

Hosting a Django application on a CherryPy server

Recently at work I’ve had the requirement to host a Django application in a CherryPy server. I first looked for various projects I knew were doing just that. Unfortunately, after trying them I was rather disapointed. Their approach is to provide a command similar to the famous Django runserver‘s one but I’ve found it to be more complex than necessary. So I wrote my own module that performs those operations by staying much closer to how CherryPy does work, most specifically by using the process bus coming with CherryPy.

I’m sharing a stripped down version of the module I wrote which shows how one could host a Django application in a CherryPy server. Hopefully this might help some of you.

You can find the code along side a minimal Django application showing how this works here (BSD licence). I used Django 1.3 to generate a default project but the code above works well with older version of Django.

Edit 16/03/2012: Thanks to Damien Tougas, I’ve wrapped up a better recipe for hosting a Django application into a CherryPy application server.

WebSocket for CherryPy 3.2

Just a quick note about the first draft of support for WebSocket in CherryPy. You can find the code here.

Note that this is still work in progress but does work against Chrome and the pywebsocket echo client. It supports draft-76 of the specification only and I’m waiting for the working-group to settle a bit more before making any further modification.

The updated code has started integrating draft-06 as well but this is a work in progress.

Running CherryPy on Android with SL4A

CherryPy runs on Android thanks to the SL4A project. So if you feel like running Python and your own web server on your Android device, well you can just do so. You’ve probably not heard something that awesome since the pizza delivery guy rung the door.

How to get on about it? Well that’s the surprise, CherryPy in itself doesn’t need to be patched. Granted I haven’t tried all the various tools provided by CherryPy but the server and the dispatching works just fine.

First, you need get the CherryPy source code, build and copy the resulting cherrypy package into the SL4A scripts directory.

Once you’ve plugged your phone to your machine through USB, run the next commands:

Just change the path to match your environment. That’s it.

Now you can copy your own script, let’s assume you use something like below:

As you can see we must disable the multiprocessing logging since the multiprocessing package isn’t included with SL4A.

Save that script on your computer as for example. Copy that file into the scripts directory of SL4A.

Unplug your phone and go to the SL4A application. Click on the script, it should start fine. Then from your browser, go to http://phone_IP:8080/ and tada! You can also go to the /location path to get the geoloc of your phone.