Policies · Mesh* policies
MeshRetry
Retry settings for outbound HTTP, gRPC and TCP traffic.
Behaviour per target
Output: MeshRetry
Copied unchanged unless it has Subset-scenario refs (MeshSubset with service tags, old
_svc_ names), which are rewritten. A deprecated top-level kind is reported.
Legacy Retry field names in to[].default are rewritten. MeshRetry never had them:
2.x drops them silently, so the settings never applied, and 3.0 rejects the write as an
unknown field on the API server and kumactl apply; kubectl apply prunes it with a warning
(UPGRADE).
http.retriableStatusCodes→http.retryOnentries, one per code ("503").http.retriableMethods→http.retryOnentries (GET→HttpMethodGet).tcp.maxConnectAttempts→tcp.maxConnectAttempt. When both are set, the existingmaxConnectAttemptis kept.
The warning says these retries now take effect. A value that is not an HTTP status code
or method is dropped with a warning. spec.targetRef.proxyTypes is reported as for
MeshTimeout.
Output: MeshRetry
The same legacy field rewrite as v2, labels-only refs and control-plane label removal.
Allowed to[].targetRef.kind: Mesh, MeshService, MeshExternalService,
MeshMultiZoneService, MeshHTTPRoute. spec.targetRef.proxyTypes is handled as for
MeshTimeout.
Timeline
| Version | Change |
|---|---|
| 3.0 | Real resources selected by labels only. New meshes no longer get a default MeshRetry. Unknown fields such as retriableStatusCodes are rejected on write. |
Related topics
- Policy targeting — Which targetRef kinds each policy accepts at each position on 2.14 and 3.0, labels-only selection, policy roles, and control-plane-owned labels.