GetWorkflowStateAsync "error from internal actor: no such instance exists" about dapr

[philliphoff]

. Specifically, this internal actor code should check for an error result indicating that the workflow wasn't found and then return a nil response rather than an error response.
Alternatively, we could update durabletask-go to export and check for a well-known ErrInstanceNotFound error, and then have the Dapr runtime return this value instead of nil if we determine that a queried workflow doesn't exist.

@cgillum Of the two, I'd vote for the first option, returning nil instead of an error. IMO, "non existence" isn't (necessarily) an error. Or rather, whether it's an error or not should be determined by the calling application and not the object store. Returning an error value, even a very specific "not found" value, will tend to get treated as a generic error on the other side of gRPC boundary. In the case of .NET, that will be an exception, which are tedious to handle as well as impact performance. (I wish the gRPC libraries had the option of treating invocations more more like raw HTTP invocations where the result could be observed before any action is taken.)

[yaron2]

Transferred issue to the runtime repo.

cc @cgillum

[philliphoff]

As best I can tell, it looks like an issue with the workflow component in the Dapr runtime or the Durable Task Go SDK (where it doesn't treat "not found" as a non-error error, and just passes back the original error instead of nil) which then gets passed up to the Durable Task SDK on top of which the Dapr .NET Workflow SDK are built. I'd file an issue in dapr/dapr.

[cgillum]

Is the expectation that no gRPC error is observed or is the expectation that the gRPC error makes it easy to distinguish between "not found" and other errors, for example, by using appropriate status codes (NotFound instead of Unknown)?

[arturotrenard]

well reviewing the existing implementation in the sdk it seems to me that they expect a null in c# and nil in go

[cgillum]

My question was whether you the application developer have an expectation on the ideal behavior. But I assume based on the screenshots you shared that you're okay with the behavior that the SDKs were designed for.

Looking at the durabletask-go code, it seems the gRPC API implementation is consistent with the SDKs, as shown here: https://github.com/microsoft/durabletask-go/blob/60e42f2bb25034de2b20201a3139f5fcc159200e/backend/executor.go#L268-L278. In other words, it expects a nil response from the backend implementation (defined by Dapr in this case) to indicate that an instance doesn't exist, in which case it will return a "doesn't exist" response to the gRPC client, which the SDKs are designed to handle gracefully.

I think the simplest way to fix this is to update the Dapr runtime code here:

dapr/pkg/runtime/wfengine/backends/actors/backend.go

Lines 224 to 230 in d4f06c2

	res, err := abe.actorRuntime.Call(ctx, req)
	elapsed := diag.ElapsedSince(start)
	if err != nil {
	// failed request to GET workflow Information, record count and latency metrics.
	diag.DefaultWorkflowMonitoring.WorkflowOperationEvent(ctx, diag.GetWorkflow, diag.StatusFailed, elapsed)
	return nil, err
	}

. Specifically, this internal actor code should check for an error result indicating that the workflow wasn't found and then return a nil response rather than an error response.

Alternatively, we could update durabletask-go to export and check for a well-known ErrInstanceNotFound error, and then have the Dapr runtime return this value instead of nil if we determine that a queried workflow doesn't exist.

[amzhang]

+1

[arturotrenard]

If correct, that's fine with me

[arturotrenard]

any update guys??