Currently, the streaming pull manager is not aligned with proper API retry behavior.

It should retry on:

• Aborted
• Cancelled
• DeadlineExceeded
• GatewayTimeout
• InternalServerError
• ResourceExhausted
• ServiceUnavailable
• Unknown

And should terminate on:

• InvalidArgument
• NotFound
• PermissionDenied
• Unauthenticated
• Unauthorized
• Any non-API exception (Python is a dynamic language; TypeError, ValueError, generic exception, etc. shouldn't be retried)

Additionally, fatal error should immediately resolve the streaming pull future to the error state so that the caller is informed of the issue.

I will be submitting a fix for this.