SDK LRO Poller: Rest the result.HttpResponse.Body before returning

[magodo]

The longRunningOperationPoller.Poll can return different poller errors, including:

• pollers.PollingFailedError
• pollers.PollingCancelledError

These two errors contain a HttpResponse field and a Message field. As a consumer of this error type, one will likely read the body from the HttpResponse, e.g., Azure Private Endpoint can return the following LRO responses:

{
  "status": "Failed",
  "error": {
    "code": "RetryableError",
    "message": "A retryable error occurred.",
    "details": [
      {
        "code": "ReferencedResourceNotProvisioned",
        "message": "Cannot proceed with operation because resource /subscriptions/xxxx/resourceGroups/acctest-mgd-petestt4/providers/Microsoft.Network/virtualNetworks/virtnetnamex/subnets/ssub3 used by resource /subscriptions/xxxx/resourceGroups/acctest-mgd-petestt4/providers/Microsoft.Network/networkInterfaces/example-endpoint-in-ssub3.nic.b1ecd3e7-fbcb-4e40-960b-5e7937dd8b66 is not in Succeeded state. Resource is in Updating state and the last operation that updated/is updating the resource is PutSubnetOperation."
      }
    ]
  }
}

{
  "status": "Failed",
  "error": {
    "code": "StorageAccountOperationInProgress",
    "message": "Call to Microsoft.Storage/storageAccounts failed. Error message: An operation is currently performing on this storage account that requires exclusive access.",
    "details": []
  }
}

Both are retryable, though it is not a good idea to embed this retry logic in the SDK itself. It'd be better to do it per resource type. Therefore, accessing the latest response of the LRO, and inspect the error code shall be done.

Currently, the HttpResponse is initially reset right after directly read in this function:

go-azure-sdk/sdk/client/resourcemanager/poller_lro.go

Lines 130 to 138 in 774530e

	var respBody []byte
	respBody, err = io.ReadAll(result.HttpResponse.Body)
	if err != nil {
	err = fmt.Errorf("parsing response body: %+v", err)
	return
	}
	result.HttpResponse.Body.Close()
	result.HttpResponse.Body = io.NopCloser(bytes.NewReader(respBody))

While it is later consumed indirectly by the call below:

go-azure-sdk/sdk/client/resourcemanager/poller_lro.go

Lines 206 to 209 in 774530e

	lroError, parseError := parseErrorFromApiResponse(*result.HttpResponse.Response)
	if parseError != nil {
	return nil, parseError
	}

This PR resets the response back right after this function call.

Aside: It might be tempted to use the LatestResponse() method from the poller to get the latest response body. Unfortunately, the current implementation of PollUntillDone will set it to nil on error:

go-azure-sdk/sdk/client/pollers/poller.go

Lines 208 to 210 in 774530e

	if p.latestError != nil {
	p.latestResponse = nil
	}

Aside2: The CreateOrUpdateThenPoll method currently generated is not wrapping the poll error:

go-azure-sdk/resource-manager/web/2023-12-01/domains/method_createorupdate.go

Lines 70 to 72 in 774530e

	if err := result.Poller.PollUntilDone(ctx); err != nil {
	return fmt.Errorf("polling after CreateOrUpdate: %+v", err)
	}

This makes users have no way to cast the error to any of the poller errors. The only workaround is to split the call to CreateOrUpdate + PollUntillDone.

Relating to hashicorp/terraform-provider-azurerm#21293