Deploy Alluxio on a Cluster with HA

Slack Docker Pulls GitHub edit source

Overview

An Alluxio service with High availability (or HA) achieved by running multiple Alluxio master processes on different nodes in the system. One of these masters is elected the leading master which serves all workers and clients as the primary point of contact. The other master processes act as standby masters which maintain the same file system state with the leading master by trailing a shared journal. Note that, standby masters do not serve any client or worker requests; however, if the leading master fails, one standby master will automatically be chosen to take over and become 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 here to achieve HA are maintaining the shared file system state across service restarts and maintain consensus among masters on leading master after failover. In Alluxio 2.0, there are two different ways to achieve these two goals:

  • Option1: Use an internal replicated state machine based on RAFT to both store the file system journal and select a leading master. This approach is introduced in Alluxio 2.0 and requires no dependency on external services.
  • Option2: Leverage an external Zookeeper service for leader election on the leading master and a shared storage (e.g., the root UFS) for shared journal. See journal management documentation for more information about choosing and configuring Alluxio journal system.

Prerequisites

  • To deploy a 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 master node to 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

Option1: 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.

  1. $ cp conf/alluxio-site.properties.template conf/alluxio-site.properties

Add the following properties to the conf/alluxio-site.properties file:

  1. alluxio.master.hostname=<MASTER_HOSTNAME> # Only needed on master node
  2. alluxio.master.mount.table.root.ufs=<STORAGE_URI>
  3. alluxio.master.embedded.journal.addresses=<EMBEDDED_JOURNAL_ADDRESS>

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 include alluxio.master.hostname=1.2.3.4, alluxio.master.hostname=node1.a.com.
  • The second property alluxio.master.mount.table.root.ufs=<STORAGE_URI> sets the URI of the shared under store to mount to the Alluxio root. This shared under store must be accessible by all master nodes and all worker nodes. Examples include alluxio.master.mount.table.root.ufs=hdfs://1.2.3.4:9000/alluxio/root/, alluxio.master.mount.table.root.ufs=s3://bucket/dir/
  • The third 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 is 19200. 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 Copycat which has built-in leader election. The built-in leader election cannot work with Zookeeper since we cannot have two leaders which might 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.

Option2: Zookeeper and Shared Journal Storage

The additional prerequisites of setting up Zookeeper HA cluster are:

  1. 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.
  2. 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.
    • 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:

  1. alluxio.zookeeper.enabled=true
  2. alluxio.zookeeper.address=<ZOOKEEPER_ADDRESS>
  3. alluxio.master.journal.type=UFS
  4. alluxio.master.journal.folder=<JOURNAL_URI>

Explanation:

  • Property alluxio.zookeeper.enabled=true enables the HA mode for the masters, and informs workers that HA mode is enabled.
  • Property alluxio.zookeeper.address=<ZOOKEEPER_ADDRESS> sets the ZooKeeper address when alluxio.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 include alluxio.zookeeper.address=1.2.3.4:2181, alluxio.zookeeper.address=zk1:2181, zk2:2181,zk3:2181
  • Property alluxio.master.journal.type=UFS indicates UFS is used as the journal place. Note that Zookeeper cannot work with journal type EMBEDDED (use a journal embedded in the masters).
  • Property 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 include alluxio.master.journal.folder=hdfs://1.2.3.4:9000/alluxio/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 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 millions of files.

  • alluxio.zookeeper.session.timeout=120s
    • 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 the tickTime. You could also manually configure minSessionTimeout and maxSessionTimeout.

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 either SESSION or STANDARD. It is set SESSION 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 journal must be formatted.

Formatting the journal will delete all metadata from Alluxio. However, the data in under storage will be untouched.

On any master node, format Alluxio with the following command:

  1. $ ./bin/alluxio format

Launch Alluxio

To start the Alluxio cluster with the provided scripts, on any master node, list all the worker hostnames in the conf/workers file, and list all the masters in the conf/masters file. This will allow the start scripts to start the appropriate processes on the appropriate nodes.

On the master node, start the Alluxio cluster with the following command:

  1. $ ./bin/alluxio-start.sh all SudoMount

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.

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:

  1. $ ./bin/alluxio fs leader

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:

  1. $ ./bin/alluxio runTests

Access an Alluxio Cluster with HA

When an application interacts with Alluxio in HA mode, the client must know about the Alluxio HA cluster, so that the client knows how to discover the Alluxio leading master. There are two ways to specify the HA Alluxio service address on the client side:

Specify Alluxio Service in Configuration Parameters

Users can also pre-configure the service address of an Alluxio HA cluster in environment variables or site properties, and then connect to the service using Alluxio URI like to alluxio:///path where the required connection details to the HA cluster information is already configured with these parameters. For example, if using Hadoop, you can configure the properties in core-site.xml, and then use the Hadoop CLI with an Alluxio URI.

  1. $ hadoop fs -ls alluxio:///directory

Depending on the different approach to achieve HA, different properties are required to set:

  • When connecting to an Alluxio HA cluster using embedded journal, set property alluxio.master.rpc.addresses to decide the node addresses to query. For examples,

    1. alluxio.master.rpc.addresses=master_hostname_1:19998,master_hostname_2:19998,
    2.  master_hostname_3:19998`

  • When connecting to an Alluxio HA cluster using Zookeeper, the following properties are needed to connect to Zookeeper to get the leading master information. 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

    1. alluxio.zookeeper.enabled=true
    2. alluxio.zookeeper.address=<ZOOKEEPER_ADDRESS>

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, HBase, 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.

Common Operations

Below are common operations to perform on an Alluxio cluster.

Stop Alluxio

To stop an Alluxio service, run:

  1. $ ./bin/alluxio-stop.sh all

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:

  1. $ ./bin/alluxio-stop.sh masters # stops all masters in conf/masters
  2. $ ./bin/alluxio-stop.sh workers # stops all workers in conf/workers

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:

  1. $ ./bin/alluxio-stop.sh master # stops the local master
  2. $ ./bin/alluxio-stop.sh worker # stops the local worker

Restart Alluxio

Starting Alluxio is similar. If conf/workers and conf/masters are both populated, you can start the cluster with:

  1. $ ./bin/alluxio-start.sh all

You can start just the masters and just the workers with the following commands:

  1. $ ./bin/alluxio-start.sh masters # starts all masters in conf/masters
  2. $ ./bin/alluxio-start.sh workers # starts all workers in conf/workers

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:

  1. $ ./bin/alluxio-start.sh master # starts the local master
  2. $ ./bin/alluxio-start.sh worker # starts the local worker

Format the Journal

On any master node, format the Alluxio journal with the following command:

  1. $ ./bin/alluxio format

Formatting the journal will delete all metadata from Alluxio. However, the data in under storage will be untouched.

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

  1. $ ./bin/alluxio-start.sh worker SudoMount # starts the local worker

Once the worker is started, it will register itself with the Alluxio master, and become part of the Alluxio cluster.

Removing a worker is as simple as stopping the worker process.

  1. $ ./bin/alluxio-stop.sh worker # stops the local worker

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 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:

  1. Update the master configuration on all the master nodes without restarting any master.
  2. Restart the leading master (can be determined by running bin/alluxio leader). Note that, a new leading master will be elected to continue the service.
  3. Wait for the previous leading master to come up successfully but as a standby master.
  4. Update and restart all remaining standby masters.
  5. 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. One can simply stop the local worker, update the configuration (e.g., conf/alluxio-site.properties) file on this worker, and then restart the worker.