What tools do you use to build your web applications? Your choice matters a lot because you’ll spend a lot of time working in that environment.
If you make a poor choice, you’ll be stuck with something suboptimal that can slow you down massively and make your project grind to a halt.
If you make a harmonious choice, your tool will aide you on your development journey and let you crank out features to delight your customers.
This topic deserves your consideration because time is one of our most precious resources, and a bad tool choice will eat your time.
Don’t waste your time on tools that are bad fits for your project!
The best way to save your time is to know the landscape of tools. Knowing the landscape will let you choose a tool that will be a great fit for you.
We’re going to explore different modes of doing web development and the tools for each mode so you can get an idea of the styles and decide which one will fit you best.
For my examples, I’m going to assume that you’re building a Django web application. This is a great choice, but I certainly won’t fault you if you make other choices. My use of Django is to illustrate patterns that are common in many web frameworks and simultaneously give you some real code to look at.
You started your amazing project yesterday. You have grand visions of becoming The Hot New Thing™.
This isn’t your first rodeo with Python so you knew to:
Start with a virtual environment.
$ python3 -m venv venv && source venv/bin/activate
Install the web framework.
(venv)$ pip install Django
Generate boilerplate to get started.
(venv)$ django-admin startproject new_hot_thing && cd new_hot_thing (venv)$ ./manage.py startapp awesome # Run more Django stuff
Run the development server.
(venv)$ ./manage.py runserver
This is a pattern that you’ll find in many of the Python web frameworks. Once you make a place for code to live, generate some code from starter templates, and run the code with a local web server, you can make serious headway!
What can we say is good about this setup?
- This is about as minimal as you can get. The low barrier to entry means that web development can be accessible to many people.
- Every web framework worth its salt will explain how to do this. The tutorials want to get developers to success to hook them into joining that ecosystem. This means that you’ll find lots of support when getting this far.
- It feels pretty amazing to see the success page from the development web server after five minutes worth of work (Dopamiiiine!!).
- When you start writing tests, those tests should be fast because they will run directly from your local operating system.
What’s not great with this setup?
- Getting this running on another machine will become a pain.
Since you installed Django directly
pip, it’s a step that you’ll have to remember when running this somewhere else.
This will be a problem when you want to go live and share your creation with the world because running your software from your laptop is not going to go well.
- Your local operating system is not going to match
with the operating system of your live website.
Statistically, you run Windows. Maybe you’re a macOS person like me, but the data shows that Windows still dominates. Most websites run on AWS, GCP, Azure, or on a Virtual Private Server like Digital Ocean or Linode, and most things running on these services are Linux.
The bad news is that the differences between operating systems will bite you. Windows and Linux are different beasts.
I am acutely aware of this fact from writing a cross platform Perl application waaaay back in the day (the filesystem
\differences alone were the bane of my existence).
- The development web server will only take you so far. Those tools are designed for local development and can be light on features. Want to test out HTTPS support with the dev server? Tough luck.
Emerging from primordial web app goo
Weeks or months have ticked by. Version 1.0 of The Hot New Thing is nearly ready, and your site is growing.
Management of this system is nutty because you have four different terminals open to start all the different processes to check your site.
This level of pain is a common step
on the path to growing.
Now is a good time to introduce
and process management tools.
the Ruby community created a tool called
foreman manages running multiple processes in a single command.
Each process is given a label (like
and all logging is funneled
to a single terminal window.
You define the processes to run in a
with something that could look like:
web: ./manage.py runserver worker: celery worker --app new_hot_thing:celeryapp --loglevel info frontend: webpack --watch
This model of working makes it very quick
to start (and stop!) your development system.
Procfile model is also used
as the way to specify your system
for their Platform as a Service.
That’s very convenient if you want to go live quickly
without the hassle of managing cloud infrastructure.
foreman is a great tool if you’re running a Ruby on Rails or Sinatra app because the tool is based on Ruby and you already have a Ruby runtime available.
But we’re working on a Django app!
Thankfully, there is a Python clone of foreman called honcho. honcho gives you the same benefits without needing Ruby.
Now you can start up like this:
(venv)$ honcho start 23:02:53 system | web.1 started (pid=25464) 23:02:53 system | worker.1 started (pid=25465) 23:02:53 system | frontend.1 started (pid=25466) 23:02:53 frontend.1 | some logging 23:02:53 web.1 | some more logging 23:02:54 worker.1 | even more logging
Okay, what’s good in this model?
- You can wrangle complexity into a single tool.
If you’re working with a small team
with different expertise
(like backend Python developers
everyone can start the system
by interfacing with
Frontend devs can hone their magical
webpackincantations, and backend folk can swap
runserverwith a more featureful server like
- This system can still run as fast as your operating system will permit. There are no extra layers since honcho will start native OS processes.
- All your logging in a single place on the terminal is excellent.
You don’t have to jump between terminal tabs
tailmultiple log files.
And the bad aspects?
- This complexity comes at the cost of setup. For instance, if you’re running Celery locally, then maybe that means you need RabbitMQ running. You have to install all the extra tools that your system will need onto your laptop. That can get hairy.
- If you’ve developed automation tools to deploy to your live site, you easily create a rift between what you need for your local setup versus what the cloud servers need.
Inception level 2
If The New Hot Thing gets too big with too many moving parts (or you’re a consultant supporting multiple clients which would cause conflicting machine configurations), running on your local operating system loses its appeal.
You may also be bothered that you’re running a different operating system than what your live site runs.
Now we get virtual.
The only way to stay directly on your local laptop and use the same operating system as your live site would be to switch the operating system that your laptop runs. This isn’t practical for most developers.
Instead, you can run your code in a virtual machine (VM). A virtual machine is a fully functioning operating system that lives within your local OS. The local operating system is called the “host” OS.
To run a virtual machine, you’ll use a hypervisor (e.g., VirtualBox). The hypervisor’s job is to pretend to be a physical machine. By acting as a physical machine to the virtual machine, the virtual machine runs as if it was installed directly on the hardware.
The most popular tool for running a development environment within a VM is with Vagrant. Vagrant is designed to make it seamless when interacting with a hypervisor.
To run your virtual machine, run:
$ vagrant up
Vagrant will download the operating system if it doesn’t have it yet. Then it will start it up and configure it with a user account. This makes it very quick to get into your VM. To run commands on the VM, you connect to it as if it were a remote machine. Vagrant installs some development security certificates which means you can ssh into the VM with:
$ vagrant ssh
Once you’re connected to the VM, it will behave like any other operating system that you’ve worked with before.
While the foreman/honcho setup depended
Vagrant depends on the
This configuration file provides the specifics
for your virtual machine. The file can:
- Instruct Vagrant which operating system image it should base the virtual machine on.
- Configure a mapping from a host directory to a directory on the VM. This is useful to access your project code quickly inside the VM.
- Run “provisioning” scripts to set up the virtual machine to run your web server, background task worker, or anything else that your site must run with.
This version of development adds a very large layer into your process so why would you want to include it?
- Your development environment will run the same operating system as your live site. Any problems you catch and fix on the development site will be problems that you solved for your live site.
- You won’t litter your host operating system with development software that you may not want regularly. This means it will be safer and easier to run multiple projects on your machine without worrying about conflict.
- Fewer differences in the operating system means that your development configuration can be closer to your live site and should be less to maintain.
- Easier to share with a team. Since you have to use automated scripts to get your VM going with all the right dependencies, it is easier to get teammates in a similar configuration.
It isn’t all sunshine and roses.
This mode is slower. Running a virtual machine requires running through the virtual OS and the host OS. You can expect a 10-20% performance overhead for running this way.
Mo' layers, mo' problems. Adding complexity at any point in your stack increases the odds of new problems.
For instance, I’ve found myself fighting with hypervisor drivers when my VM and host OS wouldn’t play nicely together. This is a problem that couldn’t exist with local development because local development doesn’t even use a hypervisor!
Dependent on provisioning tools. I’ve run up against scenarios where the automated scripts for the VM get so complicated that no one wants to touch it anymore. That’s not a good place to be.
You can get extremely far with a virtual machine-centric development mode. I’ve been at businesses with millions of users that exclusively use virtual machines for development. It’s a solid choice for teams for a long time.
What can you do when things start getting big? Like, really big?
It’s over 9000!1
Another popular mode of development uses a design that bundles dependencies and a web app into a single unit. This unit is a container.
Containers became popular a few years ago when Docker solidified. Since that time, the software industry is abuzz with discussion about containers.
Why? Why would you reach for containers for your app?
When you use containers, you’re trying to address the mismatch between development and your live site. The industry term for a live website is “production,” so you might hear about this problem as the dev/prod (development/production) parity problem.
We’ve already encountered this problem a couple of times in this article. If an environment has differences between how it was developed and how it’s used live, problems from those differences can affect projects.
Let’s think back to our local development example.
In the most primitive case,
the parts that you install to make your website run
will not be the same
as your live site.
If you’re in that statistical group
of Windows on the desktop
and Linux in the cloud,
you immediately have differences
at the OS
and foundational pieces
like virtual environments
venv not VMs).
What if, for instance, The New Hot Thing is a replacement for Instagram? In that case, you’ll be manipulating images. Image code libraries are tricky beasts with lots of weird dependencies (hi, PIL!). It’s very likely that your local setup would be different than production.
Ready for some bad news? Inter-dependency and environment problems get harder as your site grows and the complexity of your software system increases.
Container technology like Docker tries to fix this problem. And that’s why mature teams with complex systems reach for containers.
Obligatory shipping container analogy
The fundamental analogy that is used for containers is that a container is akin to a shipping container.
A shipping container is successful for transporting goods because it has a standard size and interface that make it easy to work with. If you’ve seen one shipping container, you’ve seen them all.
By using standards around sizes, the shipping industry is able to produce tools that can ship all sorts of goods. A container could hold cars or be full of stuffed teddy bears. In either case, the tools that move the container from dock to ship to dock to truck are the same. This produces massive efficiencies in the transport of goods around the globe.
Docker and friends attempt to do something similar for software. If you can shove all your code and dependencies into a container (i.e., a thing with a common interface), then moving that thing around from one environment to another should be easier.
Containers in practice
That’s the theory. How well do containers work in practice and who should attempt this?
That depends on how much staff you have, how much control you want, and maybe even what language you’re using to build your web app.
After you’re able to produce a container (technically, a container image, but let’s not sweat that detail for now), you can move that container around and plug it into a system that manages containers.
Here’s where the staff size comes in. There are services that exclusively manage containers like AWS Fargate. A service like Fargate trades staff time for money. With Fargate, Amazon will run your containers on their infrastructure without forcing you to manage machines. Considering the cost of developer time, this could be an insanely good deal for you. Aside from cost, one downside of this approach is less control.
If you need more control for your container environment, Kubernetes is probably where you will end up. Kubernetes (often abbreviated as k8s) is a “container orchestration” tool. The tool is designed to run huge distributed infrastructure. This makes perfect sense when you learn that Google is the original source of the project.
When you use Kubernetes, you have to start thinking about your container/project as a service that must be flexible enough to move around to different machines in an ever-changing infrastructure. There is an entire world of tools and services that emerged to support this development model. If that sounds terrifying (and it’s totally reasonable for that to sound terrifying), then your project probably doesn’t need to go there.
The big cloud vendors like Google offer Kubernetes on their infrastructure. It’s extremely challenging to convert a large project to use k8s. You can expect a DevOps team working on it full time to take months to complete such a transformation.
Developing with containers
You know what’s crazy!? We haven’t even talked about how to develop software in a container world!
I think this is a common problem for this subject of containers. It doesn’t take much to get sucked into this whirlwind of operational concerns about managing dev/prod parity. You can forget that someone still has to produce a product to make users happy.
If the dev/prod parity problem is killing your team and you’ve decided to move to containers and bear the cost of that transition, let’s look at what this will mean for developers on a day-to-day basis.
Step 1 of working with containers is to build a container image.
Because Docker is the most popular container technology,
we’ll use that for our Django example.
To build a container,
your project will need a
Dockerfile. At least the industry has some kind of consistency.
Dockerfile is an unusual format
that is a series of commands
which describe how to produce your project’s image.
Here’s a sample.
I wouldn’t rely on it as best practices.
You can use it more to give you an idea
for what goes into a
FROM ubuntu:xenial RUN apt-get update && apt-get install -y \ python-pip ADD ./requirements.txt /srv/new_hot_thing/requirements.txt RUN pip install -r /srv/new_hot_thing/requirements.txt WORKDIR /srv/new_hot_thing ADD . /srv/new_hot_thing EXPOSE 8000 CMD ["uwsgi", "--ini", "uwsgi.ini"]
This file starts from a base image
installs some Python packages,
and tells Docker what command to run
to execute the container (i.e.
From the directory where your
Dockerfile is located,
you build the image.
$ docker build -t new-hot-thing:1234 .
This image does nothing by itself. It’s an artifact. You can think of it as a blueprint for a process that could run in the future. When you run a container image, it starts a container instance. This is the living process that will actually execute code. Docker has a run command to vivify an image.
$ docker run new-hot-thing:1234
In a container universe, the developer lifecycle looks like:
- Write some code.
- Build an image.
- Run the image as a container instance.
Running the instance may be the only way
that you can run your unit tests.
Since the container is the place where dependencies are stored,
your workflow will start to revolve around Docker.
This could mean wrapper scripts to interact
with the code.
For instance, PyPI is built with Warehouse.
The project uses a Makefile
to run common Docker operations
Developing with multiple containers (without k8s)
Observant readers might have noticed that this container only runs the web server. What happened to Celery, Redis, and webpack? If you guessed “make more container images,” then you get a prize! Here’s an internet high five! 🎉✋🎉
To get back to the same level of productivity as honcho for local development, you need to run multiple containers in tandem. For those starting out with Docker, Docker Compose is your tool. Docker Compose gives you a tool set to link Docker images together to produce a system on your local host.
If you want to use Docker Compose,
you’ll need another YAML file.
This one is called
We can use Warehouse again
since that project has an extensive
The Docker Compose file breaks down the system
These services are very similar to what we saw
Warehouse uses some of the same names
as our example like
Through this configuration file,
we can build out all the containers we require
to run our system.
Here’s a sample service (trimmed a bit for clarity):
worker: build: context: . args: DEVEL: "yes" command: hupper -m celery -A warehouse worker -B -l info volumes: - ./warehouse:/opt/warehouse/src/warehouse:z env_file: dev/environment links: - db - redis
To get the show going:
$ docker-compose up
Developing with multiple containers (with k8s)
Nearly every DevOps person that I speak with about Kubernetes will get almost giddy about it. This strange glee does not carry over to most developers I know. I attribute a large part of that feeling in developers to a lack of knowledge about how to develop well in this world.
We’ve seen the massive amount of complexity that containers introduce. Kubernetes layers on even more. How does a developer survive this?
Because Kubernetes is a different kind of animal with its own lingo like Pods, Namespaces, and Nodes, running as a developer in this world is different.
In contrast to Docker Compose, Kubernetes has a more fluid topology than Compose. At the core, containers run inside a cluster. The cluster is a set of nodes (a.k.a. machines) that act together and are orchestrated by k8s software in a dynamically changing configuration.
We work in Kubernetes by taking a container definition and putting it into a Pod. A container defines the application level needs like language runtime and dependencies, while a Pod defines the cluster level needs like required memory and CPU usage. The Pod tells Kubernetes how to allocate the application to some amount of nodes in the cluster.
If you want to develop in Kubernetes, you need a cluster. As your team gets started, you can probably use minikube for each developer. minikube is a 1 node cluster that contains all the core k8s components inside a single virtual machine.
minikube is Kubernetes-in-a-box.
In the cluster’s default state,
it won’t have any of your application’s services
We can deploy these services
or any other cluster
or some other tool like Helm.
Since we’ve descended down the rabbit hole
for quite a while,
let’s gloss over the details
around this level of provisioning.
I want to skip the subject of provisioning tools because it’s a huge topic that can’t fit into this (already long) article. In any of the development modes that we’ve looked at, provisioning tools are needed to automate some or all portions of deployment. You can do local or VM development with Ansible, Puppet, or Chef as your tool. Or you can work with containers and find tools like Helm. In any modality, you’ll need tooling to help. We can look more at provisioning tools and their benefits in a future article.
One of the unfortunate side effects of working with containers is the constant need to make container images to run the code. This issue is so problematic that Microsoft and Google created tools to work around this limitation.
Each tool takes a slightly different approach, but I think I could sum up their primary function as: run a process that will watch code files for changes and build and deploy to a Kubernetes cluster when changes are detected.
For Draft, the process to run is:
$ draft up
For Skaffold, the equivalent command is:
$ skaffold dev
Both tools will do roughly the same thing.
- Watch for code changes.
- Build a new container image.
- Push the image to a container repository if you’re working on a remote cluster (a minikube setup will skip this slow step).
- Deploy the container image to your development cluster with whatever provisioning tools that you’ve configured.
Check out the features of each tool for yourself to decide if one is a better fit than the other. From my own analysis, I think Skaffold is a stronger offering because of its support for multiple provisioners and having more flexibility around how you build your container.
That was a lot to say to cover the containers mode of development. Now we can get an idea of the good and the bad.
What makes containers good?
- Containers give you confidence that what you make in development will work nearly identically as what you put on your live site.
- The number of moving parts in an individual container is fewer than in a large monolithic system. This can make individual containers easier to reason about.
- Containers push system design toward a service-oriented architecture. We might debate if this is a good architecture, but it does provide clear boundaries between different parts of the system which is a boon for very large teams.
Containers have definite downsides.
- Using containers adds more layers and increases the complexity of the system a lot.
- Building container images takes time and introduces some friction in the development flow.
- Containers require a lot of effort to make a development environment. As the number of containers/services increases, the effort required also increases because of coordination and communication costs.
In this article, we’ve looked at three modes of development.
- Local development - Starting and working with your project directly on your host operating system
- Virtual machine development - Running your project and all its dependencies inside of a virtual machine (guest operating system) on your computer
- Container development - Building container images that integrate with industrial strength cluster management tools like Kubernetes
Bottom line pros and cons
If you’re racing to the end of this article and are looking for the “answer,” then I’m afraid you’re out of luck. Each mode of development is right depending on your project or team’s context.
When should you consider a particular mode of development?
For local development,
- Pro: It’s great when speed is paramount or if you’re prototyping or exploring a problem space.
- Con: The differences between your development environment and your live site’s environment may cause issues.
For virtual machine development,
- Pro: Virtual machines replicate most of the qualities of your live site so you’ll get a lot of stability from using them.
- Con: Running an entire operating system on your computer is slower and the extra layer of operating systems can create trouble.
For container development,
- Pro: Containers will help large teams scale up their large systems using patterns established by the big dogs like Google.
- Con: The complexity of container management will make most developers feel like mortals who don’t really understand how their code works inside the system.
Here’s my unsolicited advice for choosing a development mode:
Stick with local development for as long as you can tolerate its warts.
Early or small projects need to move quickly to determine if a project is worthwhile. If you’re starting a business or trying to prove out the viability of an idea, the last thing you want is all the decisions associated with writing code in a container-centric world.
Virtual machines will also slow you down as you have to think about bridging between the virtual operating system and your actual computer.
If you really need speed, use a Platform as a Service (PaaS) like Heroku. Considering the cost of your time and the benefit of getting to market quickly, this option may be far cheaper than you think.
Don’t let complexity kill your project prematurely!
If your project or idea has legs, you’ll know when to move up to a more complex workflow. Unless you’re moving straight to a service-oriented architecture (and why would you do that?), I believe that virtual machines are a better choice than containers.
A successful project is likely a project that has a certain amount of complexity. The complexity that was manageable on a local machine setup can start to overwhelm your team. A virtual machine is able to wrap up that complexity. This will smooth out the differences between developer environments so you can collaborate with other developers easier.
Finally, if your system ever accelerates to huge growth, you can move to containers.
Remember: Your Hot New Thing is not Google. You likely do not (and will not) have Google-sized problems. The benefit is that you don’t need Google-sized solutions.
I hope you found this article helpful to understand different modes for developing software.
If you found this useful, please share it with others on Twitter or whatever social media you enjoy so they can benefit too.
I’d love to read your thoughts or try to answer your questions. You can follow me on Twitter or tweet at me at mblayman.