If you publish for a living, you have probably done a content audit: open a spreadsheet, list every URL, mark what is stale, what is thin, what nobody owns, and decide what to refresh or retire. It is not glamorous, but it is where the real gains hide. Teams that run containers need almost the same exercise, and the discipline that covers it is called ContainerOps in the Ops Atlas. Not a household word, to be fair; the Atlas itself says it is its own label rather than an agreed industry term. Still, the checklist it implies is easy to borrow, and that is what this post offers.
Grab a coffee and pick one service you care about. Nothing gets repaired in this pass; you are only gathering facts, the way you would before touching a blog archive. For background on what the discipline includes, the Ops Atlas page for ContainerOps lists its scope in a tidy in-and-out table, which is handy to have open in another tab.
Part 1: List everything you publish
An audit starts with an inventory, and for containers the inventory is images. Write down, for one service, the following:
- The image name and where it is stored.
- Which base image it is built from, and when that base was last refreshed.
- Whether production refers to it by tag or by digest.
- Who is named as its owner.
The Atlas opens its first month the same way: inventory first, before policies. That feels right to anyone who has tried to set content rules for a site without knowing how many pages it has. Two earlier how-to guides on publishing an article on Substack and writing a post for the DEV Community follow a similar order of operations, where setting up the account and looking around come before writing anything.
Part 2: Check the freshness dates
Old posts rank poorly and mislead readers; old base images carry known problems forward. The Atlas suggests measuring base-image age and the time it takes for a fix to travel from release to production, offering seven days for critical fixes as an example target. Add two columns to your sheet: “base age in days” and “last rebuilt”. Sort by the oldest. The top of that list is your refresh queue, exactly as it is on a content calendar.
Part 3: Look for orphans
Every audit turns up pages that nobody remembers commissioning. The container equivalent is an image or a running service without an owner. The entry names unowned estates that quietly get older as an anti-pattern, and the cure is dull but effective: attach a team name to every image repository, and review the list on a schedule. If you cannot find an owner, mark it for retirement and see who objects.
Part 4: Check the identifiers
This is the part that sounds technical and is really about not being misled. A tag is a label that can be moved to another build. A digest is a fingerprint of the content itself and stays put. For each service ask: “if I wrote down what is live today, could I prove later that I meant exactly this?” With a floating tag, the honest answer is no. The Atlas’s target is full digest coverage in production, and its maturity descriptions mention reaching more than 80 percent coverage by the sixty-day mark on the way there.
A second question belongs here: how do those digests get updated? A dependency bot can propose pull requests when a base image changes, which stops pinning from turning into neglect. That answers the worry that fixed references mean stale software.
Part 5: Skim the “technical SEO” of each container
Marketers run a technical pass on top of the content pass: speed, crawl errors, redirects. The container version is about how the thing behaves in practice. Tick these off for your sample service:
- Does it start as a non-administrator user, with a read-only root filesystem where possible?
- Are memory and process-count limits set?
- If it listens on a port, is that port published only to the local address unless it must be public?
- Does the main process receive the polite stop request, or does a shell wrapper swallow it?
- Is the log driver configured so files cannot grow without bound?
Number three deserves a note. The Atlas reports that ports published by the container engine can bypass common host firewall front-ends, and that a port mapping written without a host address listens on every network interface. Its advice is to bind to the loopback address for anything internal. Number four matters because the default grace period before a forced kill in Compose is ten seconds, and a forced kill leaves exit code 137 behind. Number five matters because the default json-file logger has no rotation out of the box.
Part 6: Score what you found
Audits are only useful if they turn into priorities. For scoring, the Atlas supplies a ladder with five rungs. At level zero builds happen on laptops, tags float, and containers run as administrator. Level one adds builds in CI and linting of the build file. Level two means digests, a non-root user and limits, with full digest coverage. Level three adds regular measurement against a policy, and level four automates rebuilds and runs untrusted code in sandboxed runtimes. Most small teams will honestly score themselves a zero or a one, and there is no shame in it. The score simply tells you where the next hour of effort pays best.
Part 7: Make a ninety-day plan
Here is a plan echoing the Atlas’s 30/60/90 outline, rewritten for a small marketing-tech team:
- Days 1 to 30: finish the inventory, set log rotation, add a linter such as Hadolint in warning mode so nothing breaks yet.
- Days 31 to 60: switch deploys to digests, switch on the update bot, and add the non-root user and limits to the ten most important services.
- Days 61 to 90: turn the linter from warning to blocking, agree a time limit for shipping fixes, and apply a retention rule to the registry.
Notice that none of this needs a new platform. The Atlas keeps this discipline to single containers and their images, and leaves orchestration, deep scanning and log pipelines to other fields, so the checklist stays small enough to finish.
A few audit habits worth keeping
Keep the spreadsheet. The first pass tells you where you stand, but the second pass, three months later, tells you whether anything improved, and that comparison is what makes people trust the work. Date every row. Resist the urge to fix as you go on the first pass, because you will lose track of the full picture. And when you share findings, lead with two or three items that matter most rather than a forty-row dump, as you would when handing a content audit to a busy editor.
Finally, remember that scanning alone is not an audit outcome. Scanning with no quick rebuild path is summed up in the entry as a list, not a fix. If your sheet ends with a long list of known issues and no route to ship a rebuilt image quickly, put that gap at the top of the plan. Faster rebuilds beat longer reports every time.