Nine Parts-Room Habits That Explain ContainerOps

Nine Parts-Room Habits That Explain ContainerOps

Ask anyone who has worked a stores counter at a maintenance hangar what makes a good parts room, and the answers come fast. Everything has a tag. Nothing is installed without paperwork. Old stock is not left on the shelf until nobody remembers what it was for. Years ago a stores clerk I know used to say that a hangar is only as trustworthy as its shelves. I kept thinking of her sentence while reading about ContainerOps, the name the Ops Atlas gives to looking after individual software containers: how they are built, named, stored and run.

The software world has rediscovered, in its own vocabulary, nearly every rule a parts room already follows. Here are nine of them, taken one by one, with the container version beside the hangar version. The source for the container side is the Ops Atlas description of ContainerOps, which also admits that the term is not a settled industry standard, so consider it a useful lens rather than a rulebook.

1. Every part carries a serial number, not just a nickname

On the shelf, a bracket may be called “the long one”, but the paperwork follows its serial. Container images work the same way. A version tag is a nickname somebody can reassign; a digest is a fingerprint worked out from the content itself. The Atlas recommends that production deploys refer to the digest and that release tags, once published, are never moved. Then if an engineer says “this is what ran on the night of the fault”, nobody has to wonder which build that meant.

2. Receive stock from a known supplier, and keep a local copy

Good stores do not depend on a single distant vendor being open at midnight. Container teams run a mirror or pull-through cache close to their build machines. The Atlas gives an illustrative scenario: thirty build runners, anonymous downloads from a public service, and a budget of 100 pulls per six hours. The runners exhaust it, requests start failing with a “too many requests” error, and builds stall. Signing in and fronting the registry with a cache resolved it. The same entry also warns that a cache is not a backup, which any stores clerk would translate as “a second bin is not an archive”.

3. Do not modify installed parts on the spot

A mechanic does not file a new hole into an installed fitting because the paperwork is slow. They replace the part and record it. The Atlas lists patching a running container by opening a shell inside it. The change vanishes at the next restart and nobody recorded it. The preferred route is to rebuild the image from source, push it, and redeploy the new digest, so that what is running always matches what is written down.

4. Shelf life is real, so track the age of what you hold

Sealants, o-rings and certain fluids carry expiry dates, and a parts room that ignores them eventually ships something unusable. Images age as well, mainly through their base layer, the foundation that supplies the operating system files. The Atlas proposes tracking base-image age as one of its measures, alongside how quickly a fix reaches production. It sets an example target of seven days for critical fixes. The point is not the exact number but the habit of having one and checking it.

5. Count what comes in, because it adds up fast

Stock rooms fill up quietly. The Atlas works through registry growth with round numbers. Imagine 40 services, each pushing 25 builds a week for two processor platforms at about 15 mebibytes apiece; that lands near twenty-nine gibibytes in a week. With a retention policy in place it levels off near 185 gibibytes. Without one the shelves keep growing until somebody has to fetch a bigger building. So retention rules and garbage collection are as ordinary here as scrapping time-expired stock.

6. Keep the bins tidy so you can find the part

A shelf crowded with unlabelled bags produces unnecessary purchases and unnecessary delays. A related theme on the software side is the small final image. A multi-stage build compiles the application in one stage and copies only the result into a bare last stage; the entry mentions a static distroless base that weighs about 2 MiB, which shows how lean the end product can be, and it also suggests that bots, not people, keep base references current. Hangar crews would call that a standing order for stock rotation. Less clutter in the bin means less to inspect and less to be wrong. If you enjoy the buying side of this tidiness, our note on common mistakes when buying control system parts covers the same instinct in hardware.

7. Set the limits before the part goes into service

Every component has a rated load. Container teams set equivalent numbers, a memory ceiling and a cap on process count, on every container, because one misbehaving workload should not take down the whole machine. The Atlas is more reserved about processor limits and says to use them cautiously, since short bursts get throttled. It also notes a classic surprise: a program that reads the processor count from inside a container may still see the host’s full count, so a thread pool sized from it is too large.

8. Know who owns each shelf

The quietest failure in any store is an area nobody feels responsible for. One of the Atlas’s listed anti-patterns is an unowned estate that ages without anyone noticing, and its five-level maturity ladder starts with ad hoc builds from a laptop and climbs to automated rebuilds with sandboxed running. Its thirty, sixty and ninety day outline begins with a simple inventory: find out what you have, set sensible log defaults, and start linting build files in warning mode before you start blocking anything.

9. Keep the logbook, and turn the pages

An aircraft logbook is useless if it is a single endless scroll nobody can open, and just as useless if pages are torn out. Container output behaves a lot like that. The Atlas warns that the default json-file log driver has no rotation unless you configure it, so a chatty container can quietly fill a disk. A small setting for maximum size and number of files is all it takes. Where the logs travel afterwards belongs to another discipline, but writing them somewhere finite is squarely a container matter.

What the hangar and the registry share

Notice that none of the nine habits needs clever tools. They need labels that mean something, a habit of recording changes, a regular look at what is getting old, and a named person. People who have cared for aircraft components already hold the instincts; they only need the new nouns. If your own path leans toward the engineering side of operations, a related read is building a career in infrastructure engineering through DevOpsSchool, which approaches these skills from the training angle.

One small exercise for this week

Pick a single service you or a colleague run in containers and walk it through the nine points as if it were a rack of parts. Write down which image it uses, whether the deploy refers to a digest, how old the base is, what limits exist, where the logs go, and who owns it. Nine short answers fit on half a page. The gaps you find will be the first items on a real improvement list, and you will have produced them without buying anything or changing a line of the service.