10.3. Recovering from System Failures


VoltDB Home » Documentation » Using VoltDB

10.3. Recovering from System Failures

When running without K-safety (in other words, a K-safety value of zero) any node failure is fatal and will bring down the database (since there are no longer enough partitions to maintain operation). When running with K-safety on, if a node goes down, the remaining nodes of the database cluster log an error indicating that a node has failed.

By default, these error messages are logged to the console terminal. Since the loss of one or more nodes reduces the reliability of the cluster, you may want to increase the urgency of these messages. For example, you can configure a separate Log4J appender (such as the SMTP appender) to report node failure messages. To do this, you should configure the appender to handle messages of class HOST and severity level ERROR or greater. See the chapter on Logging in the VoltDB Administrator's Guide for more information about configuring logging.

When a node fails with K-safety enabled, the database continues to operate. But at the earliest possible convenience, you should repair (or replace) the failed node.

To replace a failed node to a running VoltDB cluster, you restart the VoltDB server process specifying the deployment file, rejoin as the start action, and the address of one of the remaining nodes of the cluster as the host. For example, to rejoin a node to the VoltDB cluster where myclusternode5 is one of the current member nodes, you use the following command:

$ voltdb  rejoin --host=myclusternode5 \

Note that the node you specify may be any active cluster node; it does not have to be the node identified as the host when the cluster was originally started. Also, the deployment file you specify must be the currently active deployment settings for the running database cluster.

10.3.1. What Happens When a Node Rejoins the Cluster

When you issue the rejoin command, the node first rejoins the cluster, then retrieves a copy of the database schema and the appropriate data for its partitions from other nodes in the cluster. Rejoining the cluster only takes seconds and once this is done and the schema is received, the node can accept and distribute stored procedure requests like any other member.

However, the new node will not actively participate in the work until a full working copy of its partition data is received. The rejoin process can happen in two different ways: blocking and "live".

During a blocking rejoin, the update process for each partition operates as a single transaction and will block further transactions on the partition which is providing the data. While the node is rejoining and being updated, the cluster continues to accept work. If the work queue gets filled (because the update is blocking further work), the client applications will experience back pressure. Under normal conditions, this means the calls to submit stored procedures with the callProcedure method (either synchronously or asynchronously) will wait until the back pressure clears before returning control to the calling application. The time this update process takes varies in length depending on the volume of data involved and network bandwidth. However, the process should not take more than a few minutes.

During a live rejoin, the update separates the rejoin process from the standard transactional workflow, allowing the database to continue operating with a minimal impact to throughput or latency. The advantage of a live rejoin is that the database remains available and responsive to client applications throughout the rejoin procedure. The deficit of a live rejoin is that, for large datasets, the rejoin process can take longer to complete than with a blocking rejoin.

By default, VoltDB performs live rejoins, allowing the work of the database to continue. If, for any reason, you choose to perform a blocking rejoin, you can do this by using the --blocking flag on the command line. For example, the following command performs a blocking rejoin to the database cluster including the node myclusternode5:

$ voltdb rejoin --blocking --host=myclusternode5 \
         --deployment mydeployment.xml

In rare cases, if the database is near capacity in terms of throughput, a live rejoin cannot keep up with the ongoing changes made to the data. If this happens, VoltDB reports that the live rejoin cannot complete and you must wait until database activity subsides or you can safely perform a blocking rejoin to reconnect the server.

It is important to remember that the cluster is not fully K-safe until the restoration is complete. For example, if the cluster was established with a K-safety value of two and one node failed, until that node rejoins and is updated, the cluster is operating with a K-safety value of one. Once the node is up to date, the cluster becomes fully operational and the original K-safety is restored.

10.3.2. Where and When Recovery May Fail

It is possible to rejoin any appropriately configured node to the cluster. It does not have to be the same physical machine that failed. This way, if a node fails for hardware reasons, it is possible to replace it in the cluster immediately with a new node, giving you time to diagnose and repair the faulty hardware without endangering the database itself.

It is also possible, when doing blocking rejoins, to rejoin multiple nodes simultaneously, if multiple nodes fail. That is, assuming the cluster is still viable after the failures. As long as there is at least one active copy of every partition, the cluster will continue to operate and be available for nodes to rejoin. Note that with live rejoin, only one node can rejoin at a time.

There are a few conditions in which the rejoin operation may fail. Those situations include the following:

  • Insufficient K-safety

    If the database is running without K-safety, or more nodes fail simultaneously than the cluster is capable of sustaining, the entire cluster will fail and must be restarted from scratch. (At a minimum, a VoltDB database running with K-safety can withstand at least as many simultaneous failures as the K-safety value. It may be able to withstand more node failures, depending upon the specific situation. But the K-safety value tells you the minimum number of node failures that the cluster can withstand.)

  • Mismatched deployment file

    If the deployment file that you specify when issuing the rejoin command does not match the current deployment configuration of the database, the cluster will refuse to let the node rejoin.

  • More nodes attempt to rejoin than have failed

    If one or more nodes fail, the cluster will accept rejoin requests from as many nodes as failed. For example, if one node fails, the first node requesting to rejoin will be accepted. Once the cluster is back to the correct number of nodes, any further requests to rejoin will be rejected. (This is the same behavior as if you tried to add more nodes than specified in the deployment file when initially starting the database.)