It’s been two weeks that I joined Datadog as a Software Engineer working at the Logs team. During this two-week onboarding, not only I participated in a lot of presentations and training, but I also learned a lot about DevOps best practices. That’s why I’m very excited to share them with you.
Datadog gives a lot of materials on Day 1 to help you onboard: the latest version of Macbook Pro 13 and accessories, a State Backpack, a Beats headphone, and much more. As a developer, you will also have a Dell 27” Quad HD monitor and a standing desk. If they don’t fit your need, you can ask IT team for other material, even for your preferred mechanical keyboard found on the internet :) Many meetings are scheduled to help you understand different subjects:
- HR: Datadog history, contacts, tools and resources
- IT: Computer set up, especially the security enforcements
- Security: Security training
- Team: Architecture overview and different big topics
- Code: How to set up your environment
- Health: sign the healthcare plan, the death and disability plan, etc
Each person also has a mentor and a buddy: the mentor is someone inside your team and buddy is someone in another team. Mentor, buddy, teammates and team lead, they helped me a lot to learn, discover, and adapt to this new environment. Things are going fast and I may forget to share some other information for you, but I think what matters most here is: Datadog cares about new employees and gives you the best they can to help you onboard.
In the following sections, I want to share something different: as a new Software Engineer, the good practices I observed over development and operations using the Logs team as an example.
Architecture. Architecture has an important role in the whole team. In the last two weeks, I participated in different presentations about the Logs platform. Despite the focus of each speaker, all of them shared the same overview of the platform: they use the same vocabulary and identify major components in the same way, without ambiguities. As many data-intensive applications, we use Apache Kafka to build our system. It makes the system easy to decouple and each component easy to describe. Typically, when new data arrives, it is stored in a Kafka topic with replicas. After another step, data is modified and stored again in another topic, … and so on. In this way, we reduce the impact of incidents since data is persistent and avoid eventual data loss. There are physical and virtual isolations for different business units and different customers to protect data and avoid incident escalation. Whenever there is an important change in the system, people draft their proposal as “Request For Change (RFC)” and discuss it with others.
Dependency. As a Software Engineer, I am quite surprised by how our dependencies are up-to-date. In the Logs team, we are using the latest version of JUnit 5, AssertJ, … even JDK—we are running on JDK 13 for several days now. The team treats software dependencies as a first-class concept and handles them very with care.
Build. The build is very fast: even though there are 60+ modules and 3M+ lines of code, the build pipeline can finish within less than 30 minutes with all the tests executed. I believe the big secret behind the screen is the multi-steps build, where many steps are being executed in parallel. It is worth mentioning that some plugins are enabled to enforce the Java compiler, such as Google Error-Prone, which catches common Java mistakes as compile-time errors.
Release. At Datadog internal repositories, we don’t use Semantic
Versioning, because release happens so often that semantic
versioning does not fit the need anymore. Instead, the release version is the
combination of the commit id and the Docker image id:
Every pull-request merged into the
prod branch generates a new release
automatically. For example, there are 196 releases over the last two weeks in
the Logs team:
$ git log --oneline --merges --since='2 weeks' | wc -l 196
Deploy. Deployment is started manually but the process is 100% automated. The deployment granularity is fine: engineers have the choice to deploy artifacts per environment and component. For a given component, there are also flags for separating deployment and activation: artifacts can be deployed without being enabled. In terms of deployment, engineers did a great job on system design, so deploying artifacts does not cause any downtime. Also, a lot of technical detail is configurable at runtime so deployment is not even necessary.
Operate. We use Ansible, Terraform, Chef, Kubernetes and other modern frameworks to instantiate, provision and operate our cloud environments. Infrastructure as code (IaC) might be the best summary of this part. In case of an incident, there is a well-defined process. The severity of incidents are categorized is 5 levels from 1 to 5. Level 1 and level 2 are said “outage”, level 3 to 5 are said “regular incident”. Outage requires an incident commander and several responders. On-call members will be notified within minutes, even it is 3 AM in the morning. A lot of tools have been developed to ensure the process goes smoothly. When an incident happened, a Slack channel is automatically created for a live conversation. A postmortem is created for each incident regardless of the severity. Another incident-related meeting also organized to ensure the postmortem action points have been addressed. It is also worth to notice that the on-call members are software engineers because the one who knows the best about the system is the one who created it in the first place.
Monitor. Monitoring is the core business of Datadog. Software engineers monitor our system via our own product. We have dashboards, monitors (alerts), APM, Logs and much more to ensure we observe correctly the health of our production. There are also other products used for different purposes. For example, we also keep track of the errors that happened in production.
In this article, I shared the onboarding process and some good practices I observed these two weeks. For now, what amazed me the most is the ownership of the software: engineers code, test, release, deploy, operate in the same team. There must be a lot of cool things that I didn’t cover in this article. But I plan to write more :) Interested to know more? You can subscribe to my feed, follow me on Twitter or GitHub. There are many open positions at Datadog. If you would like to join, free feel to ping me at “mincong.huang [ AT ] datadoghq.com”. Hope you enjoy this article, see you the next time!
- Semver, “Semantic Versioning 2.0.0”, Semantic Versionning, 2019. https://semver.org
- Wikipedia, “Infrastructure as code”, Wikipedia, 2019. https://en.wikipedia.org/wiki/Infrastructure_as_code
- Error Prone, “Error Prone”, Error Prone, 2019. https://errorprone.info