fix race condition

docker/controller.py

@@ -1110,7 +1110,8 @@ class JobManager(Manager):
 
             # pull the image to the swarm
             if info.ver_id is not None:
-                # FIXME: race condition: will fail if ver.state==sandbox
+                # NOTE: race condition: will fail if ver.state==sandbox
+                #  jobs must be submitted after the image is committed
                 yield from ctrl.image_manager.pull(info.ver_id, swarm=True)
 
             # lock the semaphore (to limit the total number of jobs)
@@ -1187,6 +1188,8 @@ class PushManager(Manager):
             log.info("push from the %-8s %s:%s", "sandbox", image, tag)
             yield from self.run_in_executor(docker_check_error, self.ctrl.sandbox.push, image, tag)
 
+            reset()
+
             with ses.begin():
                 prev = ses.query(WebappVersion).filter_by(
                         webapp_id = version.webapp_id,
@@ -1202,16 +1205,10 @@ class PushManager(Manager):
                     for key in "updated_at", "changelog", "published":
                         setattr(prev, key, getattr(version, key))
             
-                    # TODO get rid of this 'replaced' state (and delete the record immediately)
-                    #   The 'replaced' state was introduced to avoid a race condition
-                    #   if multiple tasks are calling push() concurrently. If
-                    #   the version row is deleted, then the second call to push()).
-                    #   However this 'replaced' state does not need to be
-                    #   written to the db (because is only important for the
-                    #   running controller process). We could use another way
-                    #   of storing this state.
+                    # mark this version as replaced
+                    # TODO: purge this entry after some time
                     version.state = int(VersionState.replaced)
-
+                ses.add(version)
 
 class ImageManager: