LEVIATHAN SYSTEMS

Networking

The Same Short Service Name Can Select a Different Kubernetes Service

Technical contentPublished Updated

No. After an AI client is redeployed in another namespace, an unchanged short Service name can resolve successfully to a different Kubernetes Service. In the standard Linux Pod case examined here, the client's namespace helps complete that name. Compare the expanded DNS question and its answer with the intended Service before treating lookup success as continuity of destination.

A short Service name leaves part of the destination to context

The name ranker does not specify a namespace. Kubernetes namespaces scope Service names, so two Services can share that short name without being the same Service. An unchanged configuration string can therefore mean either a stable destination or a destination local to whichever namespace runs the client. The application's requirement determines which meaning is correct.

The Kubernetes v1.34 DNS documentation describes namespace-based search expansion and cross-namespace lookup. RFC 1034 section 3.1 supplies the underlying distinction: a relative name is completed using local naming context, while an absolute name includes the path to the DNS root. The input string and the completed DNS question are different pieces of evidence.

References: [K1][K2][R1]

NAME-42: two Services, one intended destination

Consider fictional example NAME-42, using the Kubernetes v1.34 documentation's naming model, in one cluster whose cluster domain is cluster.example. The .example domain is reserved for documentation. Service SB is named ranker in namespace blue; Service SG is named ranker in namespace green. They are ordinary, non-headless IPv4 ClusterIP Services with distinct, successfully published cluster IP addresses. Both AI clients are required to select SG in green.

The clients run in Linux Pods without hostNetwork, using ClusterFirst and the standard namespace search context. The successful expanded questions and answers below are stipulated premises. Custom DNS, Windows, aliases and ExternalName Services, record changes, retained cache effects, connection reuse and lookup failures are outside this case. No query sequence, request count or latency is predicted, and no cluster was tested.

The ordinary-Service assumption matters: its address record represents the Service's cluster IP. A headless Service has no allocated cluster IP and uses endpoint addresses in DNS instead. NAME-42 does not apply that different answer model.

References: [K1][K3][E1]

Two successful lookups can disagree with the same requirement

Client CB runs in blue and supplies ranker. Its successful expanded A question is ranker.blue.svc.cluster.example. The stipulated answer is SB's cluster IP. This selects SB in blue, so it fails the requirement to select SG in green. DNS has supplied the answer for the completed question; the mismatch appears when that answer is compared with the required Service.

Client CG runs in green and also supplies ranker. Its successful expanded A question is ranker.green.svc.cluster.example. The answer is SG's distinct cluster IP. This lookup satisfies the Service-selection requirement.

Read CG as the original client placement and CB as a replacement placement after a redeployment into blue. Both placements report successful resolution of the unchanged input ranker. The completed names and selected Services differ. Neither a DNS outage nor an old record is needed to explain the changed destination.

The counterexample defeats a narrow continuity claim: preserving the short name and lookup success did not preserve selection of SG. If the requirement had instead been to select each namespace's local ranker, those same two observations would fit that different requirement.

References: [K1][K2]

Qualifying green changes what the blue client asks for

For the third supplied observation, CB remains in blue but supplies ranker.green. Under NAME-42's search assumptions, its successful expanded A question is ranker.green.svc.cluster.example. The returned cluster IP belongs to SG. That observation satisfies the same requirement that the first CB lookup failed.

The complete absolute DNS name ranker.green.svc.cluster.example. also identifies the same DNS question from either namespace in this case. The final dot denotes the root in DNS textual notation. It makes the completed name explicit in the explanation; it is not an instruction to append a dot to an application URL. An application's accepted hostname syntax and treatment of that hostname need their own assessment.

Thus ranker.green and the absolute name provide two ways to express the green destination within the stated naming model. The former still relies on the supplied search context; the latter spells out the complete DNS name.

References: [K1][R1]

Require evidence of the destination the new placement selected

Suppose a deployment proposal promises that the client will continue using the green ranker after a namespace change. A return that says only 'ranker resolved successfully before and after' cannot establish that promise: both CB and CG can produce that summary. Keeping the same short string is insufficient evidence for a fixed green destination.

A useful return connects the intended Service to the actual observation: the cluster and client placement, the namespace and resolver context, the supplied name, the expanded question and record type, and the answer's correspondence to the named Service. In NAME-42, that chain ends at SB for the first CB lookup and SG for the other two. The comparison then has a definite result against the requirement.

The practical decision is whether the evidence covers the replacement client's selection. A successful lookup from CG in green cannot stand in for CB's lookup from blue. Equally, evidence that CB selects SG using ranker.green does not establish that CB selects SG using ranker. Each result belongs to its own input and context.

Short names remain useful when selecting the local Service is intentional. For a fixed cross-namespace dependency, make the intended namespace explicit in the requirement and assess the naming behavior used by that client.

Keep the Service-selection conclusion attached to its scope

The successful NAME-42 answer identifies the selected Service under the example's premises. It contains no observation of reachable backends, an accepted application request, the model artifact or version served, network permission, or a TLS exchange. Those outcomes require their own evidence after the intended Service has been identified.

Use the related TLS service-name guide for the expected-name comparison with a certificate. If the actual problem instead involves an absent record, use the DNS negative-cache guide; if a later request still uses an earlier address or connection, use the DNS TTL and existing-connection guide. Those explanations address different evidence from the two fresh, successful questions compared here.

Connect a separate equipment installation to its physical scope

If the wider deployment includes equipment mounting, IT cabling or labels, the related Rack-and-Stack Installation service and project inquiry provide a route to discuss that physical work package. Share a nonconfidential outline of equipment, rack positions, cable endpoints, site authorization and required installation records. Kubernetes DNS behavior and application acceptance remain with the appointed platform and application owners.

References: [L1]

Primary sources

  1. [K1] Kubernetes v1.34 — DNS for Services and PodsChecked 2026-10-21. Static v1.34 snapshot. Namespaces of Services, A/AAAA records, Pod DNS Policy/Config and Windows qualifications support the stated Linux search-context case; they do not prescribe a universal query sequence.
  2. [K2] Kubernetes v1.34 — NamespacesChecked 2026-10-21. Static v1.34 snapshot. Namespace scoping permits the same Service name in different namespaces. This does not make those Services interchangeable.
  3. [K3] Kubernetes v1.34 — ServiceChecked 2026-10-21. Static v1.34 snapshot. ClusterIP and Headless Services distinguish the example's allocated Service address from headless endpoint discovery.
  4. [R1] RFC 1034 — Domain Names: Concepts and Facilities, November 1987Checked 2026-10-21. Section 3.1 distinguishes absolute textual names, relative names and search context; application interfaces can interpret supplied strings differently.
  5. [E1] RFC 2606 — Reserved Top Level DNS Names, June 1999Checked 2026-10-21. Section 2 reserves .example for documentation. NAME-42's cluster, Services, clients and results are fictional.
  6. [L1] Leviathan Systems — Rack-and-Stack InstallationChecked 2026-10-21. Defines agreed physical equipment mounting, IT cabling, labels and installation records. Applicable only when a separate physical installation package exists.

Does the wider deployment include physical installation?

Share a nonconfidential outline of the equipment, rack positions, IT cable endpoints, site requirements and installation records to discuss scope and availability.

Discuss a ProjectEmail Project Scope