Kubernetes spreads pods where it likes; a few manifest lines and some arithmetic decide the rest

maxSkew bounds a difference. It does not distribute anything. With three zones and four replicas, the tightest legal setting still has to put the fourth replica somewhere, and 2/1/1 is the only shape available ^[6][8][10]. Lose the zone holding two and half the replicas go with it ^[10]. That is the arithmetic inherited by every capacity plan resting on the phrase "we are zone-redundant".

The region detail in Gremlin's own example is the sharper one. Two of the three zones, us-east-1a and us-east-1b, sit in the same region ^[6]. Spreading on `topology.kubernetes.io/zone` treats them as two independent domains, so a perfectly legal 2/1/1 with the pair in us-east-1a puts three of four replicas inside us-east-1 ^[11]. That is the concentration the post opens by warning against ^[7]. A zone key buys zone redundancy and nothing above it.

Then there is `whenUnsatisfiable`. Gremlin sets it to `ScheduleAnyway` so the pod runs even when the constraint cannot be met ^[8], and describes `maxSkew: 1` in the same passage as ensuring that no node carries more than one extra replica ^[9]. Both cannot be true. `ScheduleAnyway` is a scoring preference that favours topologies which reduce skew ^[4]; under real zone pressure, from a drain or a capacity shortfall, the scheduler packs pods into the surviving domains and reports success. `DoNotSchedule`, the default ^[4], is the version that means what it says, and its price is Pending replicas at exactly that moment. Which failure you prefer is the actual decision, and it is buried in one word of YAML.

Worth noticing too that the recommendation counts nodes in prose while keying on zones in the manifest ^[8][9]. A slip in a blog post costs nothing. The same slip written with `topologyKey: kubernetes.io/hostname` gives you tidy node-level balance and no zone guarantee at all.

One structural detail in the field list repays attention: the restriction is one constraint per `topologyKey` and `whenUnsatisfiable` pair ^[5], which leaves both values of `whenUnsatisfiable` available on the same key ^[17]. A hard rule at a loose skew alongside a soft rule at a tight skew is therefore expressible. Most manifests carry one constraint and stop there.

Which leaves the question the material does not answer. The text we have ends at a heading promising how to find pods with missing topology spread constraints ^[16], and at fleet scale that is the whole problem: a cluster where four deployments set the field and forty omit it has no spread policy, it has four opinions. Cluster-level defaults exist ^[2], and they are the only version of this that survives contact with a platform team that does not review every manifest.