元数据恢复

功能目标

在Pegasus bootstrap的过程中,meta server需要先从zookeeper上拉取table的元信息以及所有replica的拓扑结构,再开始服务。

元数据恢复的目标就是:让Pegasus可以不依赖zookeeper的任何信息,完成系统的bootstrap

具体流程就是:用户只需提供集群有效的replica server的集合;meta server通过和这些replica server交互,尝试重建出table元信息和replica拓扑结构,并写入到新的zookeeper节点,完成bootstrap。

注意:元数据恢复功能只是zookeeper数据损坏或者丢失之后的补救措施,运维人员要尽力避免这种情况的发生。

操作流程

使用onebox集群演示

  1. 初始化onebox集群

    只启动一个meta server:

    ./run.sh clear_onebox
    ./run.sh start_onebox -m 1 -w
    

    此时通过shell的cluster_info命令,可以看到zookeeper的节点路径:

    zookeeper_root      : /pegasus/onebox/x.x.x.x
    
  2. 使用bench工具灌数据

    灌数据是为了测试集群元数据恢复前后数据的完整性:

    ./run.sh bench --app_name temp -t fillseq_pegasus -n 10000
    
  3. 修改配置文件

    使用以下命令修改meta server的配置文件:

    sed -i 's@/pegasus/onebox@/pegasus/onebox_recovery@' onebox/meta1/config.ini
    sed -i 's@recover_from_replica_server = false@recover_from_replica_server = true@' onebox/meta1/config.ini
    

    以上命令将配置文件onebox/meta1/config.ini中的zookeeper路径进行修改,并设置为recovery模式:

    • cluster_root = /pegasus/onebox/x.x.x.x 改为 cluster_root = /pegasus/onebox_recovery/x.x.x.x
    • distributed_lock_service_parameters = /pegasus/onebox/x.x.x.x 改为 distributed_lock_service_parameters = /pegasus/onebox_recovery/x.x.x.x
    • recover_from_replica_server = false 改为 recover_from_replica_server = true
  4. 重启meta

    ./run.sh stop_onebox_instance -m 1
    ./run.sh start_onebox_instance -m 1
    

    重启成功后,meta server会进入recovery模式,此时除了start_recovery请求,所有其他RPC请求都会返回ERR_UNDER_RECOVERY,譬如,使用shell的ls命令得到结果如下:

    >>> ls
    list apps failed, error=ERR_UNDER_RECOVERY
    
  5. 通过shell发送recover命令

    首先准备recover_node_list文件,用于指定有效的replica server节点,格式为每行一个节点,譬如:

    # comment line
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    

    通过shell的recover命令,向meta server发送start_recovery请求:

    >>> recover -f recover_node_list
    Wait seconds: 100
    Skip bad nodes: false
    Skip lost partitions: false
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    =============================
    Recover result: ERR_OK
    

    当返回结果为ERR_OK时,表示恢复成功,可以通过shell的ls命令看到正常的表信息。

    通过shell的cluster_info命令,可以看到zookeeper节点路径已经改变:

    zookeeper_root      : /pegasus/onebox_recovery/x.x.x.x
    
  6. 检查数据完整性

    使用bench工具查询之前写入的数据是否完整存在:

    ./run.sh bench --app_name temp -t readrandom_pegasus -n 10000
    

    在最后的统计结果中能看到(10000 of 10000 found),表示恢复后数据完整存在。

  7. 修改配置文件并重启meta

    恢复成功后,需要修改配置文件,重新改回非recovery模式:

    • recover_from_replica_server = true 改为 recover_from_replica_server = false

    重启meta server:

    ./run.sh stop_onebox_instance -m 1
    ./run.sh start_onebox_instance -m 1
    

    这样做是避免meta server发生重启时,再次进入recovery模式,使集群变得不可用。

线上集群恢复

对线上集群进行元数据恢复时,请遵循上面的3~7步骤,需注意以下几点:

  • recover_node_list中指定有效的replica server节点时,请保证所有节点都是正常可用的。
  • 恢复之前不要忘记将配置文件中recover_from_replica_server设置为true。
  • 只能恢复到zookeeper的新节点或者空节点。
  • 恢复完成后重新将配置文件中recover_from_replica_server设置为false。

常见问题整理

  • 恢复到zookeeper的非空节点

    此时MetaServer应当启动失败并coredump:

    F12:16:26.793 (1488341786793734532 26cc)   meta.default0.0000269c00010001: /home/mi/git.xiaomi/Pegasus/pegasus/rdsn/src/dist/replication/meta_server/server_state.cpp:698:initialize_data_structure(): assertion expression: false
    F12:16:26.793 (1488341786793754317 26cc)   meta.default0.0000269c00010001: /home/mi/git.xiaomi/Pegasus/pegasus/rdsn/src/dist/replication/meta_server/server_state.cpp:698:initialize_data_structure(): find apps from remote storage, but [meta_server].recover_from_replica_server = true
    
  • 忘记设置recover_from_replica_server为true

    meta server会进入正常启动逻辑,从zookeeper上获取的apps为空,但是在config sync过程中发现replica server上有无法识别的replica,出现元数据不一致的情况,会coredump:

    F12:22:21.228 (1488342141228270056 2764)   meta.meta_state0.0102000000000001: /home/mi/git.xiaomi/Pegasus/pegasus/rdsn/src/dist/replication/meta_server/server_state.cpp:823:on_config_sync(): assertion expression: false
    F12:22:21.228 (1488342141228314857 2764)   meta.meta_state0.0102000000000001: /home/mi/git.xiaomi/Pegasus/pegasus/rdsn/src/dist/replication/meta_server/server_state.cpp:823:on_config_sync(): gpid(2.7) on node(10.235.114.240:34801) is not exist on meta server, administrator should check consistency of meta data
    
  • 恢复时连不上replica server

    如果恢复时连不上replica server,recover命令就会执行失败:

    >>> recover -f recover_node_list
    Wait seconds: 100
    Skip bad nodes: false
    Skip lost partitions: false
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    x.x.x.x:34804
    =============================
    Recover result: ERR_TRY_AGAIN
    =============================
    ERROR: collect app and replica info from node(x.x.x.x:34804) failed with err(ERR_NETWORK_FAILURE), you can skip it by set skip_bad_nodes option
    =============================
    

    可以通过指定--skip_bad_nodes参数,强制忽略有问题的节点。但是要注意,忽略问题节点可能造成部分partition的备份数不全,有丢数据风险。

    >>> recover -f recover_node_list --skip_bad_nodes
    Wait seconds: 100
    Skip bad nodes: true
    Skip lost partitions: false
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    =============================
    Recover result: ERR_OK
    =============================
    WARNING: collect app and replica info from node(x.x.x.x:34804) failed with err(ERR_NETWORK_FAILURE), skip the bad node
    WARNING: partition(1.0) only collects 2/3 of replicas, may lost data
    WARNING: partition(1.1) only collects 2/3 of replicas, may lost data
    WARNING: partition(1.3) only collects 2/3 of replicas, may lost data
    WARNING: partition(1.5) only collects 2/3 of replicas, may lost data
    WARNING: partition(1.7) only collects 2/3 of replicas, may lost data
    =============================
    
  • 恢复时发现某个partition的备份数不全

    当遇到partition备份数不全的情况时,依然能正常恢复,但是会打印告警提示:

    >>> recover -f recover_node_list
    Wait seconds: 100
    Skip bad nodes: false
    Skip lost partitions: false
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    =============================
    Recover result: ERR_OK
    =============================
    WARNING: partition(1.0) only collects 1/3 of replicas, may lost data
    =============================
    
  • 恢复时发现某个partition没有可用的replica

    当某个partition完全没有收集到可用的replica备份时,恢复会失败:

    >>> recover -f recover_node_list
    Wait seconds: 100
    Skip bad nodes: false
    Skip lost partitions: false
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    =============================
    Recover result: ERR_TRY_AGAIN
    =============================
    ERROR: partition(1.0) has no replica collected, you can force recover it by set skip_lost_partitions option
    =============================
    

    可以通过指定--skip_lost_partitions参数,强制其继续执行恢复,此时partition(1.0)会初始化为空replica。此操作需慎重,需确定能够容忍数据丢失

    >>> recover -f recover_node_list --skip_lost_partitions
    Wait seconds: 100
    Skip bad nodes: false
    Skip lost partitions: true
    Node list:
    =============================
    x.x.x.x:34801
    x.x.x.x:34802
    x.x.x.x:34803
    =============================
    Recover result: ERR_OK
    =============================
    WARNING: partition(1.0) has no replica collected, force recover the lost partition to empty
    =============================
    
  • 恢复软删除的表

    对于已经删除的表,由于有Table软删除功能,只要没有超过保留时间,replica server上的replica数据就不会被清理,所以该表能被恢复,且被认为是一个正常的未删除的表,也就是说丢掉了删除信息,但是不会丢数据。

    由于表删除后可以创建新的同名表,所以在恢复过程中可能会发现多个表都使用相同表名,出现表名冲突。此时,id最大的那个表会使用原始表名,其他表的表名都改为{name}-{id}的形式。

设计与实现

元数据恢复功能的设计与实现:

  • meta server提供一个配置项,用来标识“当从zookeeper上获取不到任何元数据信息时,是否进入元数据恢复的模式”。
  • shell端提供一个recovery的命令,来触发meta启动元数据恢复流程。
  • 如果进入了元数据恢复流程,meta server会接收replica server的探活心跳,并只响应一个特殊的start_recoveryRPC,而不响应任何其他类型的RPC。
  • 用户需要指定一个replica server的集合;meta server只和该集合中节点进行通信,响应其start_recoveryRPC,以进行信息收集,用于bootstrap;meta server和任何节点的交互失败都会导致恢复流程的失败。
Apache Incubator
Copyright © 2020 The Apache Software Foundation. Licensed under the Apache License, Version 2.0.

Apache Pegasus is an effort undergoing incubation at The Apache Software Foundation (ASF), sponsored by the Apache Incubator. Incubation is required of all newly accepted projects until a further review indicates that the infrastructure, communications, and decision making process have stabilized in a manner consistent with other successful ASF projects. While incubation status is not necessarily a reflection of the completeness or stability of the code, it does indicate that the project has yet to be fully endorsed by the ASF.