<span id="计算管理错误与恢复"></span>

# Compute administration errors and recovery

The four actions use different handler mappings. The current differences are listed below.

| Actions | Current special cases | Assumptions that do not apply |
| --- | --- | --- |
| List | 400 for parameter or state parsing failures; other errors use the shared writeServiceError mapping. | limit=0 does not use the default of 50. |
| Preflight | 404/NOT_FOUND when the target does not exist; otherwise shared errors. | Skipping draft parsing differs from an error parsing a version. A zero count does not prove that no bindings exist. |
| Delete | 409/DELETE_CONFLICT, 503/DELETE_UNAVAILABLE, otherwise 400/DELETE_FAILED | A failure after acknowledgement does not change the original response to 500. Service errors do not all map to 500. |
| Update | 400 for invalid JSON; shared service error mapping, without separate handling for missing resources, specifications, or dependency failures. | The ordinary update API's 404/NOT_FOUND, duplicate-name 409, and 503/UPDATE_UNAVAILABLE mappings do not apply here. |

Classified shared authorization denials return 403/ErrForbidden; unavailability returns 503/ErrServiceUnavailable. Unclassified shared service errors return 500/ErrServer. Direct compute errors use code/msg/data:null; generic errors may omit data, and messages may be localized. Authentication middleware, deployment configuration, and the full set of error variants still require verification in a real environment. Synthetic JSON is not a substitute for that evidence.

<span id="删除的分阶段恢复"></span>

## Recover deletion by phase

BeginDelete validates and persists a TERMINATING intent first. force does not complete cancellation or unbinding before acceptance. After acknowledgement, identity reads, leases, force cleanup, permission removal, runtime resource deletion, and final persistence can fail independently. Persisted snapshots allow periodic processing to continue. Force cleanup rereads tasks and bindings, computes the complete workflow patch, and then applies it in a transaction. Version parsing can fail before any write. The cross-database and external steps do not form a single atomic transaction.

Repeated requests with the same or a different force value are constrained by the saved intent and parameter hash. On a conflict, inspect the original operation before changing force and resending. A persisted canceled task state does not guarantee that the external process has already stopped. Retain the resource ID, original query parameters, request ID, and timestamp so authorized operators can inspect the background phase. There is no admin-get API, and the list hides terminal states; old source comments do not establish a public endpoint for polling terminal states. Internal status_message snapshots are not a stable, editable API format.

<span id="解码失败后的判断"></span>

## Assess decoding failures

Dependency fields in an HTTP 200 response can fail Go/Python/CLI model decoding. For a write, the server may already have saved the update. Verify through authorized raw HTTP access or an operator before treating a decoding failure as nonexecution. SDK owners should align models, update request options, and add coverage tests. Removing fields from documentation does not resolve the mismatch.
