Container operations rarely makes headlines, yet it is where a surprising number of business decisions are quietly made. A tag changes meaning, a registry tightens a rule, a runtime reaches the end of its support window, and a small team has a weekend of work it never planned. This state-of-the-field piece reads the evidence available in the Ops Atlas entry on ContainerOps and asks what it says about where the discipline stands.
First, a caution about vocabulary. The Atlas states that ContainerOps is its own label for single-container work beneath the orchestrator and is not a standard term. It also lists unrelated projects that use the same name, among them an open-source orchestration platform whose default branch last saw a commit in December 2018. When you search the name, expect noise. The full sheet is at the Atlas ContainerOps page.
The timeline that shaped current practice
Several dated milestones explain why container habits look the way they do. The first public demonstration of Docker took place on 15 March 2013, and the Open Container Initiative launched on 22 June 2015, which gave the industry shared specifications for images and runtimes. The image specification stands at version 1.1.1 from March 2025, and the runtime specification at 1.3.0 from November 2025.
A further set of dates concerns plumbing underneath. Kubernetes 1.24 removed dockershim in 2022, which pushed teams toward runtimes such as containerd. More recently, Docker Engine 29 made the containerd image store the default for new installations, and both Docker and Kubernetes 1.35 deprecated the older cgroup v1 resource interface. The Atlas lists cgroup v1 support continuing until May 2029, so there is time, but not unlimited time.
Support windows are the quiet deadline
One data point stands out for planners: the Atlas records containerd 1.7 reaching end of life on 30 September 2026, with the 2.3 line carrying long-term support through 30 April 2028. Anyone reading this after that September date and still on the older line is running an unsupported runtime. The Atlas treats runtime currency as a metric in its own right, alongside base-image age.
That observation points to a broader trend. Operating a container estate has moved from a question of “does it run” to “does it stay current”. The measures the Atlas proposes reflect this: coverage of digest-pinned images, the time from a fix being available to it reaching production, and the age of base images.
What the evidence says about supply changes
Two stories illustrate how external changes land on small teams. The first is an image publisher’s announcement on 16 July 2025 that, from 28 August 2025, free images would no longer be available from their long-standing location. The legacy catalogue would not receive updates, free images would be offered only under the latest tag, and users were told to update pipelines, mirrors and clusters. For anyone who had written that location into dozens of build files, the notice meant an unscheduled project.
The second is the steady pressure of registry limits. The Atlas reports anonymous Docker Hub pulls capped at 100 per six hours, with authenticated personal accounts getting 200 over the same window and paid plans unlimited under fair use. It also builds an illustrative scenario, flagged as hypothetical, in which thirty CI runners generating 180 anonymous pulls an hour hit rate-limit errors and are rescued by authentication plus a pull-through cache.
Security findings that changed defaults
Vulnerabilities in the low-level runtime have been a recurring driver. The Atlas cites a container-breakout flaw published on 31 January 2024 involving leaked file descriptors, fixed in runc 1.1.12, and a later one published on 5 November 2025 whose fixes landed in several release lines. For the second, the advice quoted is to use user namespaces with the host root user left unmapped, and to run rootless where practical. NIST’s container security guide, SP 800-190 from September 2017, remains the standing reference.
Notably, the sheet keeps deep supply-chain topics, such as signing and provenance, on the security side of the line. ContainerOps as defined here stops at the basics that make those higher layers possible: traceable images, fixed references and constrained runtimes.
Where adoption looks uneven
The Atlas offers five maturity levels. At level zero are laptop builds, floating tags and root users. Level two brings digests, non-root execution and limits, with complete digest coverage. Level three adds measurement, with fix-to-production staying inside policy for two quarters, and level four covers automated rebuilds, attestation and sandboxed runtimes.
The sheet does not publish an adoption survey, so we should not claim how many teams sit at each rung. What we can say is that the common anti-patterns it lists remain easy to find in everyday practice: floating tags, patching running containers by hand, scanning without the ability to rebuild quickly, and logs that are never rotated. The Atlas describes a scan without a fast rebuild as “a list, not a fix”.
A closer look at throttling and memory as signals
The Atlas also supplies measures that make trends visible inside one organisation. Two deserve a mention. Out-of-memory kills per thousand container-days shows whether limits are realistic, and the throttle ratio, throttled periods divided by total periods, shows whether processor caps are hurting latency. Its worked example has a container with a two-processor cap that averaged only 1.5 processors yet was throttled in 40 percent of its periods, with six seconds of waiting. Averages hide this kind of behaviour, which is why the Atlas suggests a target below five percent for services where response time matters.
The same sheet adds a subtler runtime wrinkle: some programming environments size their worker pools from the host’s processor count rather than the container’s limit, so a cap can be misjudged unless the runtime is new enough to respect it. Such details rarely appear in trend summaries, but they decide whether a migration goes smoothly.
Reading the trend for a business audience
Taken together, the dated evidence suggests three shifts. Dependencies on external registries are now an explicit risk line. Support windows for runtimes and resource interfaces are shortening the comfortable period between upgrades. And the practices that cost least, such as digests, a non-root user and memory limits, are the ones that reduce exposure most. None of that requires large spending, only attention and ownership.
The logic matches our cash-flow article, finding spending leaks to extend runway, where unnoticed small drains matter more than headline items. A container estate has leaks of its own, in unused images, forgotten services and recurring emergency upgrades. Likewise, the thinking in a blueprint for digital growth planning applies: growth plans hold up better when the operating base underneath them is understood.
What would make the picture clearer
The honest conclusion is that the discipline is well documented and unevenly practised, and that better public data on adoption would help. If you want your own number, the Atlas gives a sensible start: measure the share of production images referenced by digest, track how long fixes take to reach production, and record how many containers are killed for running out of memory per thousand container-days. A quarter of those numbers will tell you more than any industry average.

