This week as I had some downtime after PyCon Canada, I started working on resolving all the issues that I postponed. One of these issues involved applying security updates to my Linode server and rebooting the server. However when I did so… I noticed that the Rookeries site went down. When I logged into the server, I quickly found the problem: Docker refused to start after the kernel updates.
This causes the containerdservice that docker-ce is dependent on to not start. When looking at the logs (using sudo journalctl -xe), you’ll see an error along the lines of:
modprobe: ERROR: ../libkmod/libkmod.c:514 lookup_builtin_file() could not open builtin file '/lib/modules/4.18.16-x86_64-linode118/modules.builtin.bin'
modprobe: FATAL: Module overlay not found in directory /lib/modules/4.18.16-x86_64-linode118)
Thankfully there is a workaround to resolve this problem. From the instructions you need to an override configuration for the containerd service:
At the end of May, I presented a talk at PyCon 2016, on using Docker with Python microservices. You can imagine the rush I felt getting to present on such a popular topic at such a large and important conference as PyCon! While it took me a while to recuperate after PyCon and Portland both of which were amazing. I would definitely do another talk at PyCon given the opportunity. Anyways I hope you enjoy watching the video of the talk! Below the video I also wrote about preparing for the talk, its reception, and a bit of the controversy that it stirred up below the fold. 🙂 (And I apologize for the lateness of this post, its been sitting in my backlog waiting to get finished for a few weeks now. 🙁 )
Microservices and Docker are all the rage for developing scalable systems. But what challenges will you face when developing and deploying Python apps using Docker to production? This talk goes into the real-life lessons learned from creating, deploying and scaling Dockerized Python applications.
About the talk
Preparation for the Talk
PyCon talks definitely take quite a bit of time and effort to prepare. In my case, the talk took 3 major revisions before becoming the talk that I actually presented at PyCon. What started off as a intro to some of the concepts of Docker with some minor Python points, became more of a lessons learned targetted at intermediate to advanced developers. One of the things I wished I had (and I planned to but didn’t pull of) was to mention and thank my team for helping me preparing my talk. So thank you Kevin Qiu, Biniam Bekele, Yele Bonilla, and Gavin D’Mello for all your support, sitting me through three versions of my talk, and all the amazing feedback! I’ll make sure to include a slide with thanks next time.
Overall the reception of the talk was amazing! The talk turned out quite a crowd, in fact filling up most of the room. (I’m not sure of the capacity of the room but I estimate over 300 people attended). I was pretty nervous, but with the exception of a few stumbles, I think I pulled off the talk quite well. I really enjoyed some of the questions that were fielded during the Q&A session, and also privately afterwards. I wish could of answered some of the Docker Machine and Amazon ECS questions better, but I simply have not worked with both technologies long enough to give proper advice.
The most surprising aspect of the talk was the controversy it stirred up. At the end of the Q&A you can hear some comments from a young lady about where I supposedly went horrbily wrong, and how there were tweets flying back and forth about it. I had turned off the notifications on my phone when I got up on stage, to avoid getting distracted. She persisted with telling (or trying to explain) what was wrong in the private gathering afterwards. Unfortunately she did not do a wonderful job of communicating, and I felt it took away time from others to ask their questions. It didn’t help her case that she admitted to being a novice at Docker. Please don’t that as an attendee, there are better ways to disagree and communicate that.
I later approached by a gentleman (thank you whoever you are), who mentioned I should go talk to the OpenShift guys since they had some concerns about my talk. News of the Twitter controversy worried me, because I hated the notion that I had gotten on stage and toled people to go and do the wrong thing. Especially when apparently I’m telling the opposite of what Glyph from Twisted said to do. After a brief chat (and a nice demo about their cool Kubernetes suite) from the OpenShift guys, I found out that Graham Dumpleton, the creator of mod_wsgi and who works on OpenShift had done a live tweeting commentary during my talk, where he disagreed with a few of my points. Long story short, eventually I was able to chat with Graham. He was a great sport and explained his points. Interestingly enough I had also talked with the folks at Docker. And they agreed with the points in my talk, and the logic behind my points. Essentially most of my points were based off the best practises they proposed.
Anyways I listed a few of Graham’s points with links to his blog posts (thanks again Graham!), and some of my quick thoughts on each one. A quick disclaimer about some of my points: the advice I gave worked for us in our datacentre, and that it might not work for others in other environments. It should work well, it might not be perfect, but it worked for us, and some of the folks at Mozilla. I gave a disclaimer at my other talk on a Ansible setup for WSGI apps at PyCon Canada, and I thought it was superfluous. But it turns out it is a useful thing to mention, and be explicit.
So the slide that caused a good portion of the controversy was the base image one. There I had provided an example Dockerfile on half the slide and discussed about base images and good Dockerfile practises, with points on the lower half. Now the example was meant as a toy and not necessarily complete. It is difficult, even impossible to present a well formated, perfect Dockerfile in that context. There is only so much room on a slide to fit both an illustrative example and some explanatory points. That is why I included links to some samples, that hopefuly did a better job of it.
Ah yes, the “enfant terrible” of my talk. 🙂 If you want to be controversial in your talk, mentioning something like this will get people’s attention. (Ironically, it was not my desire to stir up a controversy). Graham post a while back why you might want to use virtualenvs in your Dockerized app. It is a longish post, so I’ll give a shortened version. Basically when you base your image off some distro (say Ubuntu, Fedora or what not), there is a good chance of bringing in more Python packages in your system site packages than you expected. e.g. You’re building a Flask app, and the package maintainer included a version of Werkzeug in the base Python install, so now when you pip install Flask as part of your requirements you get the wrong version of Werkzeug.
And that is a valid point (with my example)… except if you use something like the official Python 2.7 base image… which installs just Python. I would argue that you would catch and resolve this issue, if you are auditing your Docker images. (And you should be always doing your due diligence and checking your base and resulting images. ) So yes… you don’t really need virtualenvs, but you can also use them if you are concerned that you might be getting conflicting packages.
Graham was right about the adding volume mapping in the Dockerfile being problematic. You should not define volume mounts in your Dockerfile, since they create extra files with sudo-like permissions on the host (see /var). In your own datacentre that isn’t a problem. A multi-tenant cloud provider like OpenShift, would disallow you to create those files. The documentation argument I provided is not all that useful, since you can document the mountpoints in the README that you would provide with the Docker image.
Base images are hard to get right. And there is a lot of debate whether or not to use tooling instead of base images. Graham says his warpdrive tool will do that sort of a thing. At work we build out our own tooling for building “standard” service Dockerfiles, and that just add another level of abstraction. I prefer base images since it while not ideal, provides less levels of abstractions that can get in the way when you’re debugging your Dockerfile setup. But your mileage may vary here.
So yes, good base images are hard. Try not to build your own unless you find it really useful and you have a great base to work from.
Installing GCC/Build Tools
In an ideal world one ought not have to include GCC, Python dev headers and so on. Yes, one can pip install using wheels, but that doesn’t always work out.
Formatting of the RUN command. This is not one of Graham’s points, but it did come up. Yes, you should format the RUN commands, with a line for each command and using a \ line continuation for readability. My slide didn’t have enough physical space to do so. My Rookeries example does a better job of this.
Running as Root
Graham is right, you should not run containerized apps as root. That is a bad security practise that can lead to an attacker compromising your Docker host via a privileged account on your Docker container. Again a bad example on my part, I should of added a USER command and dropped the VOLUME line, or maybe rethought the use of an example.
UWSGI and the HTTP flag
No, you don’t need it and you should use the UWSGI protocol if you put an NGINX container before your WSGI container. I left the flag in to make sure the example Dockerfile was runnable. My bad on trying to get a good illustrative example, but it wouldn’t be a good idea in production unless you feel comfortable exposing UWSGI to the direct HTTP traffic.
If you’re wondering why I’ve been so quiet these past few weeks, it is because I’ve been busy preparing to go to PyCon US in Portland this year!
I am very excited not only to be attending, but I will be giving a talk at PyCon US this year! I will be talking about Dockerizing Python microservices, and some of the lessons we’ve learned along the way at work. My talk will be on the first day (Monday May 30th) at 3:15-3:45 PM (PST). Videos of the all PyCon talks should be available a few days after the talk.
Finally I will be around in Portland for a few days after the sprints as well. I have never been to Portland, so I want to check out some of the sights around there. Let me know via Twitter or email if you want to meetup with me while I’m there. 🙂