If your job touches data pipelines, you already depend on infrastructure that somebody else defined in code: warehouses, queues, storage buckets, schedulers. How that layer is built and maintained has its own working name, InfraOps, and the picture looks quite different from a year or two ago. Let’s walk through where things stand as of late 2026, keeping to what a public reference actually documents rather than guessing at hype.
A name that nobody governs
Start with the vocabulary problem. A public glossary-style resource called the Ops Atlas, put out by the xOps School team, states that neither a standards body nor a certification scheme has ever defined InfraOps. It is used loosely by analyst firms, by hiring teams writing job titles, and even by a vendor that shipped a product with the same name in late 2025. The Atlas treats that last use as a collision rather than a definition. Its own editorial choice is to apply the label to the lifecycle of infrastructure expressed as code, from writing and reviewing through applying, testing and eventually decommissioning.
For a data engineer, the practical consequence is simple. When a vacancy says InfraOps, read the responsibilities, not the label.
The tooling split most teams now live with
The biggest change on the tool side is licensing. According to the Atlas, Terraform moved to the Business Source License 1.1 starting with version 1.6.0, a change announced in 2023. In response the OpenTofu fork emerged under the Mozilla Public License 2.0, and by 2025 it had joined the Cloud Native Computing Foundation as a sandbox project. For a team choosing between them, the day-to-day commands look similar, so the decision often rests on licensing comfort and ecosystem support rather than on syntax.
Other options exist as well. The Atlas notes a general-purpose-language approach whose engine is Apache-licensed, a cloud-native template service, and Kubernetes-style controllers that reconcile infrastructure continuously. It also draws a sharp line: running an apply from a CI job after a merge is infrastructure as code with pull-request automation, not the continuous reconciliation that the GitOps standard describes.
What data teams are asking for
Conversations in data communities tend to cluster around four themes, and each has a counterpart in the Atlas.
- Smaller state, smaller surprises. The Atlas quotes Google Cloud’s advice not to put more than 100 resources in a single state, with a few dozen as the ideal. It adds that a large workplace-messaging company reached around 1,400 separate state files by 2022 by splitting them per region and per account.
- Safer secrets. State files can contain sensitive values, which is why ephemeral values, arguments that are never stored and OpenTofu encryption of state are getting attention.
- Tests for modules. Terraform has shipped a native test command since 1.6, with mocking arriving in 1.7, and the Atlas mentions third-party testing libraries alongside static checkers.
- Cost visibility. Cost estimation inside pull requests is becoming a normal gate, not an exotic extra.
Measuring the health of the layer
There is no industry benchmark for how healthy an infrastructure estate is, and the Atlas is candid that its list of metrics is its own proposal. It suggests tracking how much of the in-scope estate is actually managed by code, how often and how long drift lingers, how many changes arrive from outside the pipeline, how many applies fail, how large the worst blast radius is and how old your machine images are. It also warns against vanity numbers such as lines of code, module counts or tag counts, which grow without telling you anything about risk.
Data teams will recognise this instinct. It is the same reasoning behind monitoring freshness and null rates rather than counting dashboards. If you are building that muscle on the pipeline side, our write-ups on catching bad records before they reach production with an observability platform and on writing validation expectations for pipeline data show how a measured approach pays off one layer up.
Gaps that still bother practitioners
Several problems remain unsolved. Failed applies have no transactional rollback, so recovery means preparing a new plan and moving forward. A preview with no changes is silent about anything built by hand. Imports tend to mirror historical accidents, so the Atlas, citing Google’s guidance, recommends avoiding wholesale imports when possible. And continuous integration brings its own traps: a role used only to preview changes can read every secret in the state, and a pull request can run workflow edits made by its own author.
Signals worth tracking into 2027
Three signals are worth tracking. First, whether the OpenTofu and Terraform feature sets continue to converge or begin to diverge. Second, whether state encryption and ephemeral values become standard expectations in code review. Third, whether terms like this one settle into a shared vocabulary or keep fragmenting between jobs, products and analyst categories. None of these needs a rush decision. A team that already has a backend with locking, a pipeline that previews every change and a short list of owners for each state is well placed whichever way they go.
Where to read the source material
If you want the underlying text, including diagrams and a maturity ladder, open the Ops Atlas InfraOps entry and read it with a pencil. It is long, roughly nine and a half thousand words, but it is organised so you can jump to scope, tooling or metrics as needed.
Treat what you read there as one editorial view, not a standard, and form your own list of questions to ask the teams that run the infrastructure under your pipelines.