HDFS rolling upgrade allows upgrading individual HDFS daemons. For examples, the datanodes can be upgraded independent of the namenodes. A namenode can be upgraded independent of the other namenodes. The namenodes can be upgraded independent of datanods and journal nodes.
In Hadoop v2, HDFS supports highly-available (HA) namenode services and wire compatibility. These two capabilities make it feasible to upgrade HDFS without incurring HDFS downtime. In order to upgrade a HDFS cluster without downtime, the cluster must be setup with HA.
In a HA cluster, there are two or more NameNodes (NNs), many DataNodes (DNs), a few JournalNodes (JNs) and a few ZooKeeperNodes (ZKNs). JNs is relatively stable and does not require upgrade when upgrading HDFS in most of the cases. In the rolling upgrade procedure described here, only NNs and DNs are considered but JNs and ZKNs are not. Upgrading JNs and ZKNs may incur cluster downtime.
Suppose there are two namenodes NN1 and NN2, where NN1 and NN2 are respectively in active and standby states. The following are the steps for upgrading a HA cluster:
In a federated cluster, there are multiple namespaces and a pair of active and standby NNs for each namespace. The procedure for upgrading a federated cluster is similar to upgrading a non-federated cluster except that Step 1 and Step 4 are performed on each namespace and Step 2 is performed on each pair of active and standby NNs, i.e.
For non-HA clusters, it is impossible to upgrade HDFS without downtime since it requires restarting the namenodes. However, datanodes can still be upgraded in a rolling manner.
In a non-HA cluster, there are a NameNode (NN), a SecondaryNameNode (SNN) and many DataNodes (DNs). The procedure for upgrading a non-HA cluster is similar to upgrading a HA cluster except that Step 2 "Upgrade Active and Standby NNs" is changed to below:
When the upgraded release is undesirable or, in some unlikely case, the upgrade fails (due to bugs in the newer release), administrators may choose to downgrade HDFS back to the pre-upgrade release, or rollback HDFS to the pre-upgrade release and the pre-upgrade state. Both downgrade and rollback require cluster downtime and are not done in a rolling fashion.
Note that downgrade and rollback are possible only after a rolling upgrade is started and before the upgrade is terminated. An upgrade can be terminated by either finalize, downgrade or rollback. Therefore, it may not be possible to perform rollback after finalize or downgrade, or to perform downgrade after finalize.
Downgrade restores the software back to the pre-upgrade release and preserves the user data. Suppose time T is the rolling upgrade start time and the upgrade is terminated by downgrade. Then, the files created before or after T remain available in HDFS. The files deleted before or after T remain deleted in HDFS.
A newer release is downgradable to the pre-upgrade release only if both the namenode layout version and the datenode layout version are not changed between these two releases. Below are the steps for downgrade:
Rollback restores the software back to the pre-upgrade release but also reverts the user data back to the pre-upgrade state. Suppose time T is the rolling upgrade start time and the upgrade is terminated by rollback. The files created before T remain available in HDFS but the files created after T become unavailable. The files deleted before T remain deleted in HDFS but the files deleted after T are restored.
Rollback from a newer release to the pre-upgrade release is always supported. Below are the steps for rollback:
hdfs dfsadmin -rollingUpgrade <query|start|finalize>
Execute a rolling upgrade action.
|query||Query the current rolling upgrade status.|
|prepare||Prepare a new rolling upgrade.|
|finalize||Finalize the current rolling upgrade.|
hdfs dfsadmin -getDatanodeInfo <DATANODE_HOST:IPC_PORT>
Get the information about the given datanode. This command can be used for checking if a datanode is alive like the Unix ping command.
hdfs dfsadmin -shutdownDatanode <DATANODE_HOST:IPC_PORT> [upgrade]
Submit a shutdown request for the given datanode. If the optional upgrade argument is specified, clients accessing the datanode will be advised to wait for it to restart and the fast start-up mode will be enabled. When the restart does not happen in time, clients will timeout and ignore the datanode. In such case, the fast start-up mode will also be disabled.
Note that the command does not wait for the datanode shutdown to complete. The "dfsadmin -getDatanodeInfo" command can be used for checking if the datanode shutdown is complete.
hdfs namenode -rollingUpgrade <downgrade|rollback|started>
When a rolling upgrade is in progress, the -rollingUpgrade namenode startup option is used to specify various rolling upgrade options.
|downgrade||Restores the namenode back to the pre-upgrade release and preserves the user data.|
|rollback||Restores the namenode back to the pre-upgrade release but also reverts the user data back to the pre-upgrade state.|
|started||Specifies a rolling upgrade already started so that the namenode should allow image directories with different layout versions during startup.|