Data-driven Testing with 'pytest', Part Four: Fungible Infrastructure

My previous post discussed how an update to a testing schema opened the test harness to better capture the rest of the statefulness of the ETL pipeline. This post discusses how the infrastructure setup in the test harness evolved to capture the statefulness of each resource dependency.

SSL: Where it all began

The test harness worked quite well after the prior update, and I was able to migrate the remaining test cases to the new framework and delete the old test harnesses entirely. Then one day, one customer asked for SSL support, because they hosted all of their credentials on their own LDAP server and wanted a secure channel for authentication. A very reasonable ask, as credentials should not be passed as plain-text over network, all our services already supported SSL connections, and I didn’t have SSL support in the tool. I thought it would take a few days.

It ended up taking a few weeks when interspersed with other feature development. The difficulty arose when attempting to set the configuration in a reproducible manner. There’s quite a few knobs to turn, and tuning them incorrectly results in the database being in an erroneous state. Getting the database out of the erroneous state took time too, as I did not back up my configuration files and hence had no automated rollbacks. I ended up throwing the updated feature (with maybe a dozen lines of code changed total) over the wall to a colleague, who verified it and sent it to the customer. After that, I really wanted fungible third-party dependencies.

Difficulties with Existing Infrastructure

In the beginning, I had all my third-party dependencies (such as all the databases) installed on bare metal. This resulted in a not-insignificant amount of friction (in addition to the ones already described):

• One difficulty with code, especially code representing large features or changes that are constantly changing, is the risk of bugs and unintended behavior. When you’re writing in a language that embraces state, some of those bugs relate to statefulness. For example, running the same ingest method a fixed number of times to the same database will eventually result in out-of-memory errors in the database, perhaps because an orphaned process didn’t deallocate memory correctly. So, if you’re debugging a particular ingest path requiring actual writes to the resource, and you didn’t catch your bug in time, other bugs may appear that interfere with proper debugging of your bug. Addressing this issue without sufficient infrastructure support may involve restarting the machine instance, or reinstalling the database from scratch, or simply waiting for a patch to come out, among other possibilities.

• We (of course) support multiple versions of our product, as enterprise customers maintain their own upgrade schedule and licensing agreements and hence cannot upgrade immediately after we release a particular feature. I hadn’t been at the company long enough to work on multiple releases, but we maintained a nightly branch for the next release and as one was around the corner, I needed to reconcile features and test the tool against the upgraded database. The longer the tool exists, the more versions it needs to support. This is of course to say nothing of different operating systems (debian-based, Red Hat-based, SUSE-based, etc.), architectures (x86, POWER, etc.), and other build differences. We bundle our own development environments efficiently and it’s easy enough to get multiple clients and client dependencies up and running, but if the servers are deployed on bare metal, it’s impossible to concurrently maintain those one-to-one client/server connections.

• Sometimes, we need to see the behavior of the tool under different system constraints. For example, how does restricting the amount of memory accessible to the database affect ingest performance? Sometimes the system degrades gracefully. Other times, the system may crash. Crashing on bare metal is not fun, as it may affect other applications on your entire system (like Spotify :unamused:).

The takeaway from all this is to have good enough test infrastructure that can effectively absorb or mitigate these risks to development velocity.

Solution: Containerization

Taking yet another page from our QA team, which had good experiences with Docker and containerizing our database for testing purposes, I decided to use Docker and bash to script the database setup with SSL enabled, and tried to make the infrastructure setup as simple as possible. It was also a great opportunity to learn more about Docker and how containerization worked.

Containerization itself was not too difficult. As our database ships as a monothlic RPM, it was easy to create a Dockerfile based on top of the maintained 'centos' Docker image, docker cp over the RPM to the container, and run yum install -y $RPM to deploy the database. The difficult SSL configuration, which blocked me from finishing the infrastructure for a few more weeks, was resolved with the help of a sales engineer who had already wrote a number of bash scripts to address this problem. It was literally a Hail Mary in terms of discovery and efficacy, and this post wouldn’t be possible without his work :sweat_smile:

Here’s some of the other, more interesting gotchas/hacks I encountered/implemented:

• A single Docker container is intended to run exactly one process; when that process stops, the container should stop as well. As our database and analytics platform was very much not one process, we needed to work around this design feature. The solution I ended up with was calling tail -f /dev/null in a bash entrypoint script I copied into the Docker container and executed using the Docker keyword 'ENTRYPOINT'. This keeps a foreground process consistently running in the container, preventing premature shutdown. Then, you can run docker exec -it $CONTAINER_NAME bash to go into the container and manually check the health of the processes you care about (if you’re not binding volumes between container and host and streaming logs, which is another option).

• When I first worked from home after finishing up the infrastructure work, I couldn’t connect to the Internet from within the Docker container. I called ping 8.8.8.8 (Google DNS) and it simply read “Connection reset by peer” or “Destination Host Unreachable”. With some help from QA, I determined that the problem was not due to machine, operating system, or code quirks. Eventually, I stumbled upon this Stack Overflow post that mentioned setting "--net=host" during docker run, which fixed the problem. This means I can’t publish aliases for the ports used by various database processes in the container, which means I can’t have multiple databases running at the same time. At present, it’s an acceptable limitation; I just script the container setup and teardown synchronously and run the subset of tests for each appropriate configuration.

• SSL verification issues cropped up when SSL certificates did not match between the client and server. There was no way to ignore the errors without code changes, which were unacceptable. At first, since the original scripts to enable SSL generated self-signed SSL certificates, I copied the certificates from the container to the host and the client’s SSL certificate store. Later, I removed the certificate generation step and cached a single set of self-signed SSL certificates on the client and server to ensure they always matched. Definitely a hack (I shouldn’t even see root certificates in production), but for testing that SSL is possible, a good enough workaround for now.

Lessons Learned

I’m still very much a newbie at Docker and containerization (My test infrastructure is one container on one machine and likely will be for a while, so no docker-compose or Kubernetes or anything), but there’s some things I really appreciate about going through this process:

• Virtualization is pretty much the only way I’ve found so far to namespace state and make it completely fungible in a pragmatic manner. Sure, there’s tools to create “virtual environments” for a particular language, such as virtualenv and conda env for Python, the JVM for Java and Java-based languages, etc., but you ever move into multi-service and multi-language environments, these virtual environments are just too constricted in scope. With docker, or other virtualization service, you never have to worry about that unless you approach the limits of docker’s capabilities, which are much further away.

• Virtualization in the manner I did it is fairly cheap, in terms of total resource utilization and developer upfront effort. This makes it very appealing as the rate of return is amplified that much more. I can’t speak to actual distributed systems, or other virtualization runtimes such as nvidia-docker, which likely are more complex. I also can’t speak to development setups using virtualization, such as Vagrant, although I’ve heard good things.

Conclusion

Infrastructure setup was the keystone in a long journey to automating testing for this project. Now in pull requests, I can say to the code reviewer on pull requests to check out this branch, build the distribution, and execute this one script in order to run all the tests, instead of digging all the way down into site-packages to find the tests and run them as a one-step process. Much simpler to understand and get started with, with corresponding benefits to onboarding and test coverage. With the requirements I have on my roadmap, it will be a while until I encounter a problem that can’t be tested with the existing design abstracted implemented so far (e.g. an updated test case schema and test harness, another Docker container, etc).

Thanks for reading this four-part series! I hope you found it as informative as I found it a joy to write.