How to combine RetryPolicy.ON_ERROR with a specific RetryPolicy.ON_FAILURE + expression using OR logic?

[xiki-tempula]

Hi Hera maintainers and community,

I'm working on defining an Argo Workflow using Hera for a task that requires GPU resources. I want to implement a sophisticated retry strategy, but I'm unsure how to combine two different conditions with OR logic.

Goal:

I want my task to retry if either of the following conditions is met:

The pod fails with a system/node error (covered by RetryPolicy.ON_ERROR). This should cover cases like the node dying, preemption, or potentially some scheduling issues before the container runs properly.
OR
The container itself fails (covered by RetryPolicy.ON_FAILURE) AND the failure message specifically contains the string "Allocate failed due to no healthy devices present".

What I Know:

I know how to create a RetryStrategy for each case individually using Hera:

Case 1: Retry on any OnError:

from hera.workflows import RetryStrategy, RetryPolicy

retry_on_any_error = RetryStrategy(
    retry_policy=RetryPolicy.ON_ERROR,
)

Case 2: Retry on OnFailure only if message matches:

from hera.workflows import RetryStrategy, RetryPolicy

GPU_ALLOCATION_ERROR_SUBSTRING = "Allocate failed due to no healthy devices present"

retry_on_specific_failure = RetryStrategy(
    retry_policy=RetryPolicy.ON_FAILURE,
    expression=f"contains(lastRetry.message, '{GPU_ALLOCATION_ERROR_SUBSTRING}')",
)

The Challenge:

I need the retry to trigger if the conditions for retry_on_any_error are met OR if the conditions for retry_on_specific_failure are met.

I looked through the RetryStrategy parameters and the Argo documentation, but I don't see an obvious way to express this OR condition between a general policy (OnError) and a specific filtered policy (OnFailure + expression). Setting retry_policy seems to take only one enum value, and the expression seems to filter within the chosen policy, not combine different policies.

Question:

Is it possible to configure a single RetryStrategy in Hera (and Argo) that achieves this combined "OR" logic? If so, how would I define it?

If it's not directly possible with a single strategy, are there recommended patterns or workarounds to achieve this behavior?

Thanks in advance for any help or guidance!