openstack中主机重新调度

举个例子方便大家理解： 假设 A,B,C 三个节点都通过了过滤，最终 A 因为权重值最大被选中执行操作。 但由于某个原因，操作在 A 上失败了。 默认情况下，nova-scheduler 会重新执行过滤操作（重复次数由 scheduler_max_attempts 选项指定，默认是 3）。 那么这时候 RetryFilter 就会将 A 直接刷掉，避免操作再次失败。 RetryFilter 通常作为第一个 filter。
---- 但是，有些操作失败，是不会进行Retry的。 记录一下遇到的问题，在创建网络时提示抛出PortBindingFailed异常
    def _build_resources(self, context, instance, requested_networks,
                         security_groups, image_meta, block_device_mapping):
        .......
        try:
        ......
            network_info = self._build_networks_for_instance(context, instance,
                    requested_networks, security_groups)
        ......
        except (exception.InstanceNotFound,
                exception.UnexpectedDeletingTaskStateError):
            raise
        except exception.UnexpectedTaskStateError as e:
            raise exception.BuildAbortException(instance_uuid=instance.uuid,
                    reason=e.format_message())
        except Exception:
            # Because this allocation is async any failures are likely to occur
            # when the driver accesses network_info during spawn().
            LOG.exception(_LE('Failed to allocate network(s)'),
                          instance=instance)
            msg = _('Failed to allocate the network(s), not rescheduling.')
            raise exception.BuildAbortException(instance_uuid=instance.uuid,
                    reason=msg)
这里如果不是前几类异常（InstanceNotFound，UnexpectedDeletingTaskStateError，UnexpectedTaskStateError），就会抛出BuildAbortException异常，而这类异常不会进行Retry。
---- 回退到之前调用_build_resources函数的地方，可以发现，函数的异常捕获都是分类的，如果是BuildAbortException，则直接将return build_results.FAILED，不会尝试。
---- 如果内部调用函数抛出的异常没有在指定异常类型中的，默认抛出BuildAbortException异常。
---- 只有在compute节点上捕获到RescheduledException，才会去给conductor发消息重新调度并创建VM。
    @hooks.add_hook('build_instance')
    @wrap_exception()
    @reverts_task_state
    @wrap_instance_event(prefix='compute')
    @wrap_instance_fault
    def _do_build_and_run_instance(self, context, instance, image,
            request_spec, filter_properties, admin_password, injected_files,
            requested_networks, security_groups, block_device_mapping,
            node=None, limits=None):
        ......

        try:
            with timeutils.StopWatch() as timer:
                self._build_and_run_instance(context, instance, image,
                        decoded_files, admin_password, requested_networks,
                        security_groups, block_device_mapping, node, limits,
                        filter_properties)
            LOG.info(_LI('Took %0.2f seconds to build instance.'),
                     timer.elapsed(), instance=instance)
            return build_results.ACTIVE
        except exception.RescheduledException as e:
            retry = filter_properties.get('retry')
            ......
            self.compute_task_api.build_instances(context, [instance],
                    image, filter_properties, admin_password,
                    injected_files, requested_networks, security_groups,
                    block_device_mapping)
            return build_results.RESCHEDULED

再看发现能够抛出RescheduledException异常的地方其实很少。
        def _do_validation(context, instance, group_hint):
            ....
            if 'anti-affinity' in group.policies:
                ......
                    raise exception.RescheduledException(
                            instance_uuid=instance.uuid,
                            reason=msg)
            elif 'affinity' in group.policies:
                ......
                    raise exception.RescheduledException(
                            instance_uuid=instance.uuid,
                            reason=msg)


    def _build_and_run_instance(self, context, instance, image, injected_files,
            admin_password, requested_networks, security_groups,
            block_device_mapping, node, limits, filter_properties):
        .......
        try:
            rt = self._get_resource_tracker(node)
            with rt.instance_claim(context, instance, limits):
                ......
                with self._build_resources(context, instance,
                        requested_networks, security_groups, image_meta,
                        block_device_mapping) as resources:
                    ......
        except (exception.InstanceNotFound,
                exception.UnexpectedDeletingTaskStateError) as e:
        ............

        except Exception as e:
            ......
            raise exception.RescheduledException(
                    instance_uuid=instance.uuid, reason=six.text_type(e))

注：上面的代码是N版本的代码