Cluster with HA
Overview
An Alluxio cluster with High Availability (HA) is achieved by running multiple Alluxio master processes on different nodes in the system. One master is elected as the leading master which serves all workers and clients as the primary point of contact. The other masters act as standby masters and maintain the same file system state as the leading master by reading a shared journal. Standby masters do not serve any client or worker requests; however, if the leading master fails, one standby master will automatically be elected as the new leading master. Once the new leading master starts serving, Alluxio clients and workers proceed as usual. During the failover to a standby master, clients may experience brief delays or transient errors.
The major challenges to achieving high-availability are maintaining a shared file system state across service restarts and maintaining consensus among masters about the identity of the leading master after failover. In Alluxio 2.0, there are two different ways to achieve these two goals:
Raft-based Journal: Uses an internal replicated state machine based on the Raft protocol to both store the file system journal and run leader elections. This approach is introduced in Alluxio 2.0 and requires no dependency on external services.
Zookeeper with a shared Journal: Uses an external Zookeeper service for leader elections in conjunction with a shared storage (e.g. HDFS) for the shared journal. See journal management documentation for more information about choosing and configuring Alluxio journal system.
Prerequisites
To deploy an Alluxio cluster, first download the pre-compiled Alluxio binary file, extract the tarball and copy the extracted directory to all nodes (including nodes running masters and workers).
Enable SSH login without password from all master nodes to all worker nodes. You can add a public SSH key for the host into
~/.ssh/authorized_keys
. See this tutorial for more details.TCP traffic across all nodes is allowed. For basic functionality make sure RPC port (default :19998) is open on all nodes.
Basic Setup
Raft-based Embedded Journal
The minimal configuration to set up a HA cluster is to give the embedded journal addresses to all nodes inside the cluster. On each Alluxio node, create the conf/alluxio-site.properties
configuration file from the template.
Add the following properties to the conf/alluxio-site.properties
file:
Explanation:
The first property
alluxio.master.hostname=<MASTER_HOSTNAME>
is required on each master node to be its own externally visible hostname. This is required on each individual component of the master quorum to have its own address set. On worker nodes, this parameter will be ignored. Examples includealluxio.master.hostname=1.2.3.4
,alluxio.master.hostname=node1.a.com
.The second property
alluxio.master.embedded.journal.addresses
sets the sets of masters to participate Alluxio's internal leader election and determine the leading master. The default embedded journal port is19200
. An example:alluxio.master.embedded.journal.addresses=master_hostname_1:19200,master_hostname_2:19200,master_hostname_3:19200
Note that embedded journal feature relies on Ratis which uses leader election based on the Raft protocol and has its own format for storing journal entries. The built-in leader election cannot work with Zookeeper since the journal formats between these configurations may not match. Enabling embedded journal enables Alluxio's internal leader election. See embedded journal configuration documentation for more details and alternative ways to set up HA cluster with internal leader election.
Zookeeper and Shared Journal Storage
When using Zookeeper for HA there are additional prerequisites:
A ZooKeeper cluster. Alluxio masters use ZooKeeper for leader election, and Alluxio clients and workers use ZooKeeper to inquire about the identity of the current leading master.
A shared storage system on which to place the journal (accessible by all Alluxio masters). The leading master writes to the journal on this shared storage system, while the standby masters continually replay the journal entries to stay up-to-date. The journal storage system is recommended to be:
Highly available. All metadata modifications on the master requires writing to the journal, so any downtime of the journal storage system will directly impact the Alluxio master availability.
A filesystem, not an object store. The Alluxio master writes to journal files to this storage system, and utilizes filesystem operations such as rename and flush. Object stores do not support these operations, and/or perform them slowly, so when the journal is stored on an object store, the Alluxio master operation throughput is significantly reduced.
The minimal configuration parameters which must be set are:
Explanation:
alluxio.zookeeper.enabled=true
enables the HA mode for the masters, and informs workers that HA mode is enabled.alluxio.zookeeper.address=<ZOOKEEPER_ADDRESS>
sets the ZooKeeper address whenalluxio.zookeeper.enabled
is enabled. The HA masters use ZooKeeper for leader election, and the workers use ZooKeeper to discover the leading master. Multiple ZooKeeper addresses can be specified by delimiting with commas. Examples includealluxio.zookeeper.address=1.2.3.4:2181
,alluxio.zookeeper.address=zk1:2181, zk2:2181,zk3:2181
alluxio.master.journal.type=UFS
indicates UFS is used as the journal place. Note that Zookeeper cannot work with journal typeEMBEDDED
(use a journal embedded in the masters).alluxio.master.journal.folder=<JOURNAL_URI>
sets the URI of the shared journal location for the Alluxio leading master to write the journal to, and for standby masters to replay journal entries from. This shared shared storage system must be accessible by all master nodes. Examples includealluxio.master.journal.folder=hdfs://1.2.3.4:9000/alluxio/journal/
oralluxio.master.journal.folder=/mnt/nfs/journal/
.
Make sure all master nodes and all worker nodes have configured their respective conf/alluxio-site.properties
configuration file appropriately.
Once all the Alluxio masters and workers are configured in this way, Alluxio is ready to be formatted and started.
Advanced Zookeeper setup
For clusters with large namespaces, increased CPU overhead on leader could cause delays on Zookeeper client heartbeats. For this reason, we recommend setting Zookeeper client session timeout to at least 2 minutes on large clusters with namespace size more than several hundred million of files.
alluxio.zookeeper.session.timeout=120s
The Zookeeper server's min/max session timeout values must also be configured as such to allow this timeout. The defaults requires that the timeout be a minimum of 2 times the
tickTime
(as set in the server configuration) and a maximum of 20 times thetickTime
. You could also manually configureminSessionTimeout
andmaxSessionTimeout
.
Alluxio supports pluggable error handling policy on Zookeeper leader election.
alluxio.zookeeper.leader.connection.error.policy
specifies how connection errors are handled. It can be eitherSESSION
orSTANDARD
. It is setSESSION
as default.The
SESSION
policy makes use of Zookeeper sessions to determine whether leader state is dirty. This means suspended connections won't trigger stepping down of a current leader as long as it was able to reestablish the zookeeper connection with the same session. It provides more stability in maintaining the leadership state.The
STANDARD
policy treats any interruption to zookeeper server as an error. Thus leader will step down upon missing a heartbeat, even though its internal zookeeper session was still intact with the zookeeper server. It provides more security against bugs and issues in zookeeper setup.
Start an Alluxio Cluster with HA
Format Alluxio
Before Alluxio can be started for the first time, the Alluxio master journal and worker storage must be formatted.
Formatting the journal will delete all metadata from Alluxio. Formatting the worker storage will delete all data from the configured Alluxio storage. However, the data in under storage will be untouched.
On all the Alluxio master nodes, list all the worker hostnames in the conf/workers
file, and list all the masters in the conf/masters
file. This will allow alluxio scripts to run operations on the cluster nodes. format
Alluxio cluster with the following command in one of the master nodes:
Launch Alluxio
In one of the master nodes, start the Alluxio cluster with the following command:
This will start Alluxio masters on all the nodes specified in conf/masters
, and start the workers on all the nodes specified in conf/workers
. Argument SudoMount
indicates to mount the RamFS on each worker using sudo
privilege, if it is not already mounted. On MacOS, make sure your terminal has full disk access (tutorial here).
Verify Alluxio Cluster
To verify that Alluxio is running, you can visit the web UI of the leading master. To determine the leading master, run:
Then, visit http://<LEADER_HOSTNAME>:19999
to see the status page of the Alluxio leading master.
Alluxio comes with a simple program that writes and reads sample files in Alluxio. Run the sample program with:
Access an Alluxio Cluster with HA
When an application interacts with Alluxio in HA mode, the client must know about the connection information of Alluxio HA cluster, so that the client knows how to discover the Alluxio leading master. The following sections list three ways to specify the HA Alluxio service address on the client side.
Specify Alluxio Service in Configuration Parameters or Java Options
Users can pre-configure the service address of an Alluxio HA cluster in environment variables or site properties, and then connect to the service using an Alluxio URI such as alluxio:///path
. For example, with Alluxio connection information in core-site.xml
of Hadoop, Hadoop CLI can connect to the Alluxio cluster.
Depending on the different approaches to achieve HA, different properties are required:
If using embedded journal, set alluxio.master.rpc.addresses
.
Or specify the properties in Java option. For example, for Spark applications, add the following to spark.executor.extraJavaOptions
and spark.driver.extraJavaOptions
:
If using Zookeeper, set the following Zookeeper related properties
Note that, the ZooKeeper address (alluxio.zookeeper.address
) must be specified when alluxio.zookeeper.enabled
is enabled and vise versa. Multiple ZooKeeper addresses can be specified by delimiting with commas.
Specify Alluxio Service with URL Authority
Users can also fully specify the HA cluster information in the URI to connect to an Alluxio HA cluster. Configuration derived from the HA authority takes precedence over all other forms of configuration, e.g. site properties or environment variables.
When using embedded journal, use
alluxio://master_hostname_1:19998, master_hostname_2:19998,master_hostname_3:19998/path
When using Zookeeper leader election, use
alluxio://zk@<ZOOKEEPER_ADDRESS>/path
.
For many applications (e.g., Hadoop, Hive and Flink), you can use a comma as the delimiter for multiple addresses in the URI, like alluxio://master_hostname_1:19998,master_hostname_2:19998,master_hostname_3:19998/path
and alluxio://zk@zkHost1:2181,zkHost2:2181,zkHost3:2181/path
.
For some other applications (e.g., Spark) where comma is not accepted inside a URL authority, you need to use semicolons as the delimiter for multiple addresses, like alluxio://master_hostname_1:19998;master_hostname_2:19998;master_hostname_3:19998
and alluxio://zk@zkHost1:2181;zkHost2:2181;zkHost3:2181/path
.
Specify Alluxio Service with logical URL Authority
Some frameworks may not accept either of these ways to connect to a highly available Alluxio HA cluster, so Alluxio also supports connecting to an Alluxio HA cluster via a logical name. In order to use logical names, the following configuration options need to be set in your environment variables or site properties.
Use logical name when using embedded journal
If you are using embedded journal, you need to configure the following configuration options and connect to the highly available alluxio node via alluxio://ebj@[logical-name]
, for example alluxio://ebj@my-alluxio-cluster
.
alluxio.master.nameservices.[logical-name] unique identifier for each alluxio master node
A comma-separated ID of the alluxio master node that determine all the alluxio master nodes in the cluster. For example, if you previously used my-alluxio-cluster
as the logical name and wanted to use master1,master2,master3
as individual IDs for each alluxio master, you configure this as such:
alluxio.master.rpc.address.[logical name]. [master node ID] RPC Address for each alluxio master node
For each alluxio master node previously configured, set the full address of each alluxio master node, for example:
Use logical name when using Zookeeper
If you are using zookeeper for leader election, you need to configure the following values and connect to the highly available alluxio node via alluxio://zk@[logical-name]
, for example alluxio://zk@my-alluxio-cluster
.
alluxio.master.zookeeper.nameservices.[logical-name] unique identifier for each Zookeeper node
A comma-separated zookeeper node ID that determine all the Zookeeper nodes in the cluster. For example, if you previously used my-alluxio-cluster
as the logical name and wanted to use node1,node2,node3
as individual IDs for each Zookeeper, you would configure this as such:
alluxio.master.zookeeper.address.[logical-domain]. [Zookeeper node ID] Address foreach Zookeeper node
For each Zookeeper node previously configured, set the full address of each Zookeeper node, for example:
Common Operations
Below are common operations to perform on an Alluxio cluster.
Stop Alluxio
To stop an Alluxio service, run:
This will stop all the processes on all nodes listed in conf/workers
and conf/masters
.
You can stop just the masters and just the workers with the following commands:
If you do not want to use ssh
to login to all the nodes and stop all the processes, you can run commands on each node individually to stop each component. For any node, you can stop a master or worker with:
Restart Alluxio
Starting Alluxio is similar. If conf/workers
and conf/masters
are both populated, you can start the cluster with:
You can start just the masters and just the workers with the following commands:
If you do not want to use ssh
to login to all the nodes and start all the processes, you can run commands on each node individually to start each component. For any node, you can start a master or worker with:
Add/Remove Workers Dynamically
Adding a worker to an Alluxio cluster dynamically is as simple as starting a new Alluxio worker process, with the appropriate configuration. In most cases, the new worker's configuration should be the same as all the other workers' configuration. Run the following command on the new worker to add
Once the worker is started, it will register itself with the Alluxio leading master and become part of the Alluxio cluster.
Removing a worker is as simple as stopping the worker process.
Once the worker is stopped, and after a timeout on the master (configured by master parameter alluxio.master.worker.timeout
), the master will consider the worker as "lost", and no longer consider it as part of the cluster.
Add/Remove Masters
In order to add a master, the Alluxio cluster must operate in HA mode. If you are running the cluster as a single master cluster, you must configure it to be an HA cluster before having more than one master.
See the journal management documentation for more information about adding and removing masters.
Update Master-side Configuration
In order to update the master-side configuration, you can first stop the service, update the conf/alluxio-site.properties
file on master node, and then restart the service. Note that, this approach introduces downtime of the Alluxio service.
Alternatively, one benefit of running Alluxio in HA mode is to use rolling restarts to minimize downtime when updating configurations:
Update the master configuration on all the master nodes without restarting any master.
Restart standby masters one by one (the cluster cannot survive more than
floor(n/2)
simultaneous restarts).Elect a standby master as the leading master (tutorial here).
Restart the old leading master that is now a standby master.
Verify the configuration update.
Update Worker-side Configuration
If you only need to update some local configuration for a worker (e.g., change the mount of storage capacity allocated to this worker or update the storage directory), the master node does not need to be stopped and restarted. Simply stop the desired worker, update the configuration (e.g., conf/alluxio-site.properties
) file on that node, and then restart the process.
Last updated