diff --git a/mistral/db/v2/api.py b/mistral/db/v2/api.py
index de1d0b78c..589ee7443 100644
--- a/mistral/db/v2/api.py
+++ b/mistral/db/v2/api.py
@@ -62,6 +62,10 @@ def refresh(model):
     IMPL.refresh(model)
 
 
+def expire_all():
+    IMPL.expire_all()
+
+
 # Locking.
 
 
diff --git a/mistral/db/v2/sqlalchemy/api.py b/mistral/db/v2/sqlalchemy/api.py
index cbde52486..14b46531f 100644
--- a/mistral/db/v2/sqlalchemy/api.py
+++ b/mistral/db/v2/sqlalchemy/api.py
@@ -125,6 +125,11 @@ def refresh(model, session=None):
     session.refresh(model)
 
 
+@b.session_aware()
+def expire_all(session=None):
+    session.expire_all()
+
+
 @b.session_aware()
 def acquire_lock(model, id, session=None):
     # Expire all so all objects queried after lock is acquired
diff --git a/mistral/engine/workflows.py b/mistral/engine/workflows.py
index e2fd62ffe..0e8bef3e7 100644
--- a/mistral/engine/workflows.py
+++ b/mistral/engine/workflows.py
@@ -414,6 +414,16 @@ class Workflow(object):
         if incomplete_tasks_count > 0:
             return incomplete_tasks_count
 
+        LOG.debug("Workflow completed [id=%s]", self.wf_ex.id)
+
+        # NOTE(rakhmerov): Once we know that the workflow has completed,
+        # we need to expire all the objects in the DB session to make sure
+        # to read the most relevant data from the DB (that's already been
+        # committed in parallel transactions). Otherwise, some data like
+        # workflow context may be stale and decisions made upon it will be
+        # wrong.
+        db_api.expire_all()
+
         wf_ctrl = wf_base.get_controller(self.wf_ex, self.wf_spec)
 
         if wf_ctrl.any_cancels():
diff --git a/mistral/tests/unit/engine/test_direct_workflow.py b/mistral/tests/unit/engine/test_direct_workflow.py
index 0612f16ee..364542fef 100644
--- a/mistral/tests/unit/engine/test_direct_workflow.py
+++ b/mistral/tests/unit/engine/test_direct_workflow.py
@@ -816,6 +816,44 @@ class DirectWorkflowEngineTest(base.EngineTestCase):
 
             self.assertDictEqual({}, wf_ex.output)
 
+    def test_output_expression(self):
+        wf_text = """---
+        version: '2.0'
+
+        wf:
+          output:
+            continue_flag: <% $.continue_flag %>
+
+          task-defaults:
+            on-error:
+              - task2
+
+          tasks:
+            task1:
+              action: std.fail
+              on-success: task3
+
+            task2:
+              action: std.noop
+              publish:
+                continue_flag: false
+
+            task3:
+              action: std.noop
+        """
+
+        wf_service.create_workflows(wf_text)
+
+        wf_ex = self.engine.start_workflow('wf')
+
+        self.await_workflow_success(wf_ex.id)
+
+        with db_api.transaction():
+            wf_ex = db_api.get_workflow_execution(wf_ex.id)
+
+            self.assertEqual(2, len(wf_ex.task_executions))
+            self.assertDictEqual({'continue_flag': False}, wf_ex.output)
+
     def test_triggered_by(self):
         wf_text = """---
         version: '2.0'
diff --git a/releasenotes/notes/fix_workflow_output-cee5df431679de6b.yaml b/releasenotes/notes/fix_workflow_output-cee5df431679de6b.yaml
new file mode 100644
index 000000000..46dec0a4d
--- /dev/null
+++ b/releasenotes/notes/fix_workflow_output-cee5df431679de6b.yaml
@@ -0,0 +1,13 @@
+---
+
+fixes:
+  - |
+    Workflow output sometimes was not calculated correctly due to
+    the race condition between different transactions: the one that
+    checks workflow completion (i.e. calls "check_and_complete") and
+    the one that processes action execution completion (i.e. calls
+    "on_action_complete"). Calculating output sometimes was based on
+    stale data cached by the SQLAlchemy session. To fix this, we just
+    need to expire all objects in the session so that they are
+    refreshed automatically if we read their state in order to make
+    required calculations. The corresponding change was made.