Problem statement
The operational scripts target pods by hand-maintained name lists. Every list is wrong within a day: something is renamed, something is added, something moves environment, and the restart script skips it or hits the wrong one. The namespace fleet-ops runs eight labeled workloads. Answer seven operational questions about them without naming a single pod.
About this exercise
A selector is a query over labels. It is how everything in Kubernetes finds a set of objects it did not write down in advance - a Service finds its endpoints this way, a Deployment finds its pods this way, and so can you. A selector describes the set by what its members are, so it stays correct when the membership changes; a name list describes yesterday's membership and starts rotting immediately.
There are two families of requirement, and both can be attached to a listing:
- Equality-based - a key must equal a value, or must not equal it. Several requirements can be given at once, and they are joined by AND: an object must satisfy every one of them to match.
- Set-based - a key's value must be in a set of values, or notin it; or the key must merely exist on the object, or must not exist on it.
The subtle one is what happens to an object that does not carry the key at all:
- exists matches only objects that carry the key.
- notin matches objects whose value is outside the set - and also every object with no such key, because such an object has no value inside the set either.
So env notin (prod) and env exists are different questions with different answers, and neither is the complement of the other. That distinction is why an audit written with the wrong one quietly skips the unlabeled objects, which are exactly the ones worth finding.
An empty result is a legitimate answer to a legitimate query, not a failure.
kubectl get --help documents how a selector is attached to a listing.
How this project is graded
The grader runs each query against the live cluster itself and compares the seeded matches against your answer file, name for name. An answer that happens to be right because you typed the names out is still counted right here - but the seeded fleet is arranged so that the names do not tell you the answer, and the next exercises will not let you get away with it.
What you will build
Seven answer files in answers/, listed in answers/questions.md. Every file holds the matching pod names, one per line, sorted ascending, and nothing else - no headers, no pod/ prefix, no counts:
tier-web.txt the web tier prod-workers.txt workers running in prod batch-or-web.txt everything whose tier is batch or web not-prod.txt everything whose env is not prod env-set.txt everything that carries an env label at all env-unset.txt everything that carries no env label none.txt batch workloads running in prod
The labels in use are app, tier (web, worker, batch) and env (prod, staging, dev). Not every pod carries all three.
Step 1 - equality, alone and combined
Answer the first two: one requirement, then two requirements that must both hold.
equality_selector_returns_its_matches and combined_requirements_intersect should pass
Step 2 - set membership
Answer the third with a single requirement over a set of values, not two queries stitched together.
set_based_in_returns_the_whole_set should pass
Step 3 - the key that is not there
Answer the next three. Two of them look like opposites and are not: one of them must include the pods that carry no env label at all, and one of them must exclude them.
notin_includes_objects_missing_the_key, exists_returns_only_objects_carrying_the_key and does_not_exist_returns_the_unlabeled_objects should pass
Step 4 - the empty answer
The last query matches nothing. answers/none.txt must exist and be empty - zero bytes, no error text, no "none".
empty_match_is_an_empty_answer_file should pass
When every test passes
You have:
- Sets described by what their members are, not by the names they had yesterday
- Requirements combined into one query instead of filtered by hand afterwards
- The difference between "not that value" and "has the key" straight, and the unlabeled objects landing where they belong
- An empty result treated as an answer
You do not have yet:
- Anywhere to keep the metadata that is not a selector key - ticket numbers, owners, build JSON. Labels are the wrong home for it, and annotations are the next exercise.
Verify
Click Test after every step.
