Platform & DevOps

there is one machine everyone is afraid to reboot, and its configuration exists only in the history of who logged in.

Snowflake server

the container has no stack trace and no error log, just a restart count going up and exit code 137.

OOM kill

the deploy has been sitting at pending for twenty minutes with no error, because no node satisfies every constraint you set.

Unschedulable workload

two nodes ran what the manifest called the same version and behaved differently, because someone had pushed over the tag.

Image tag

somebody fixed it in the console during an incident, and the next apply politely undid the fix.

Infrastructure drift

the same API key is in a pipeline variable, two config maps, a wiki page and someone's laptop, and rotating it breaks three things.

Secret sprawl

nobody knows the names of the machines, and the fix for a sick one is to terminate it rather than to log in.

Cattle not pets

configuration comes from the environment, logs go to stdout, and the same image runs in staging and production unchanged.

Twelve-factor app

the process thinks it has its own machine, and on the host it is one ordinary PID among hundreds.

Container

you ship a tarball of a whole filesystem plus a line saying what to run, and the same bytes go to every environment.

Container image

a vulnerability was reported in a library you have never imported, because it came in the operating system underneath your code.

Base image

the compiler, the test framework and the whole toolchain are used in the build and then thrown away, and the shipped image contains one binary.

Multi-stage build

the deployment names a sha256 hash instead of a version, and there is exactly one set of bytes that could ever satisfy it.

Image digest

the deploy failed at three in the morning because the place you pull images from was rate-limiting you, and it is not your infrastructure.

Container registry

you stopped addressing individual machines and started submitting work to a pool that decides where it lands.

Cluster

a machine joins the pool, advertises how much CPU and memory it has, and starts being given other people's workloads.

Node

two containers share an IP address and a lifetime, and are scheduled onto the same node together or not at all.

Pod

a second container next to your application handles TLS, logging or metrics, and your code was never changed.

Sidecar

you write down what should exist, submit it, and something else works out the steps — including that most of them are already done.

Declarative configuration

you said three replicas, and every time someone deletes one a fourth appears without anybody being asked.

Desired state

the spec says three and the status says one, and the gap between those two numbers is the entire incident.

Observed state

something reads the difference between what you asked for and what exists, takes one step to close it, and then does it all again forever.

Reconciliation loop

you never create the individual containers; you create an object, and a background process creates them to match it.

Controller

you promise the scheduler the workload needs half a core, and that half core is subtracted from the node whether or not you ever use it.

Resource request

the container went one byte over its memory ceiling and the kernel killed it outright, mid-request.

Resource limit

the orchestrator asks whether the process is wedged and restarts the container when it stops answering.

Liveness probeBackend

the instance says "not yet" while the cache warms, so the load balancer keeps traffic off it until it is genuinely able to serve.

Readiness probeBackend

a healthy workload was terminated because the node it happened to be on ran out of disk, and it had done nothing wrong.

Eviction

you mark the machine unschedulable and evict everything on it politely, so the upgrade takes an hour instead of a page.

Node drain

traffic doubled and the replica count went from four to nine on its own, a minute or two after it needed to.

Horizontal autoscaling

a workload could not be placed, so a new machine was bought, booted and joined the cluster — three minutes after you needed it.

Cluster autoscaling

instances come and go constantly and callers never hold a stale address, because nobody is configured with one.

Service discovery

you connect to a name rather than an address, and every workload in the cluster resolves it to whatever is currently healthy.

Cluster DNS

traffic from outside arrives at one front door with the certificate on it, and is routed inwards by hostname and path.

Ingress

the workload can reach only the three services it declared, and its attempt to open a connection anywhere else times out silently.

Network policy

the tool's record of what it built is a third thing, sitting between your file and the real cloud, and all three can disagree.

Infrastructure state

the tool prints exactly what it will create, change and destroy, and you read the destroy list before typing yes.

Plan and apply

a nightly job plans against production and tells you that fourteen things no longer match the code, before the next deploy does.

Drift detection

the team's standard service takes six inputs and produces thirty resources, and nobody writes the thirty by hand any more.

Infrastructure module

a merge triggers a sequence of stages on someone else's machine, and the sequence is now the only way anything reaches production.

Build pipeline

the pipeline produces one file with one version, and everything downstream refers to that file rather than rebuilding.

Build artefact

the exact bytes tested in staging are the bytes that run in production, with only the configuration differing.

Build once, deploy many

it worked in staging and broke in production, because staging had one replica, no mesh and a different database size.

Environment parity

nobody has credentials to deploy; an agent in the cluster watches a repository and makes reality match it.

GitOps

instances are replaced a few at a time, so both versions serve traffic simultaneously for several minutes.

Rolling deployment

error rate crossed the threshold ninety seconds in and the system reverted to the previous version without waking anyone.

Automated rollback

the database password is fetched at start-up from a service that logs who asked, instead of sitting in the repository.

Secret management

the thing calling the API is not a person, and its permissions were granted once in 2022 by someone who has left.

Service account

the service proves what it is to the cloud without holding any credential, because the platform attests to it.

Workload identity

one datacentre lost power and a third of your instances went with it, which is the unit this is designed around.

Availability zone

you run at sixty percent so a zone failure or a traffic spike does not immediately become an outage.

Headroom

the business says four hours, and the restore has never been timed, so nobody knows whether that is true.

Recovery time objective

backups run nightly, so recovering means losing up to a day of writes, and nobody outside the platform team knew that.

Recovery point objective

every service asked for four cores because that is what the template said, and the fleet averages eight percent CPU.

Overprovisioning

you look at a fortnight of real usage and cut the requests in half, and nothing at all happens.

Rightsizing

the unattached disks, the reserved addresses and the load balancer for a service deleted last year are all still billing.

Zombie resource