Premature evaluation of "publish" if a task has "retry" with "continue-on"

Consider the following example:

wf:
  tasks:
    task1:
      retry:
        continue-on: <% global(global_var) = null %>
        count: 10
        delay: 1
      on-success:
        publish:
          branch:
            global_var_doubled: <% $.global_var + $.global_var %>

    task2:
      wait-before: 3
      on-success:
        publish:
          global:
            global_var: foobar

The problem is that the "publish" clause of "task1" gets evaluated after the first iteration of retry when "global_var" is not yet available in the global context, so it can't find the variable and fails with the error:

 File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/engine/task_handler.py", line 120, in _on_action_complete
    task.on_action_complete(action_ex)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/.tox/py37/lib/python3.7/site-packages/osprofiler/profiler.py", line 160, in wrapper
    result = f(*args, **kwargs)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/engine/tasks.py", line 499, in on_action_complete
    self.complete(state, state_info)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/.tox/py37/lib/python3.7/site-packages/osprofiler/profiler.py", line 160, in wrapper
    result = f(*args, **kwargs)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/engine/tasks.py", line 283, in complete
    data_flow.publish_variables(self.task_ex, self.task_spec)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/workflow/data_flow.py", line 220, in publish_variables
    task_ex.published = expr.evaluate_recursively(branch_vars, expr_ctx)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/expressions/__init__.py", line 100, in evaluate_recursively
    data[key] = _evaluate_item(data[key], context)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/expressions/__init__.py", line 79, in _evaluate_item
    return evaluate(item, context)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/expressions/__init__.py", line 71, in evaluate
    return evaluator.evaluate(expression, context)
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/expressions/yaql_expression.py", line 228, in evaluate
    data_context
  File "/Users/renat/workspace/openstack/mistral_group/mistral/mistral/expressions/yaql_expression.py", line 182, in evaluate
    ", data=%s]" % (expression, str(e), data_context)
mistral.exceptions.YaqlEvaluationException: Can not evaluate YAQL expression [expression=$.global_var + $.global_var, error='global_var', data={'__task_execution': {'id': '668067b7-e015-409d-ba8b-acb38244e482', 'name': 'task1'}, , '__env': {}, '__execution': {'id': '610d252f-574e-44e7-a14f-9c5a109d5ed7'}, }]
] (execution_id=610d252f-574e-44e7-a14f-9c5a109d5ed7)

If we move "wait-before" from "task2" to "task1" the example will be working.

The issue is in the design of task policies. In case if a task action (std.noop in our case since "action" is not explicitly specified) succeeds it will run the logic of task completion before the "retry" policy has a chance to change the flow. See how the method "complete" works at https://github.com/openstack/mistral/blob/master/mistral/engine/tasks.py#L283. It publishes variables before the the call "self._after_task_complete()" that triggers policies.