The mechanism of cluster scaling is designed to enable user to change the number of running instances without creating a new cluster. User may change number of instances in existing Node Groups or add new Node Groups.
If cluster fails to scale properly, all changes will be rolled back.
In order to leverage Swift within Hadoop, including using Swift data sources from within EDP, Hadoop requires the application of a patch. For additional information about using Swift with Sahara, including patching Hadoop and configuring Sahara, please refer to the Swift Integration documentation.
Cinder is a block storage service that can be used as an alternative for an ephemeral drive. Using Cinder volumes increases reliability of data which is important for HDFS service.
User can set how many volumes will be attached to each node in a Node Group and the size of each volume.
All volumes are attached during Cluster creation/scaling operations.
OpenStack Cluster may use Nova Network or Neutron as a networking service. Sahara supports both, but when deployed, a special configuration for networking should be set explicitly. By default Sahara will behave as if Nova Network is used. If OpenStack Cluster uses Neutron, then use_neutron option should be set to True in Sahara configuration file. In addition, if the OpenStack Cluster supports network namespaces, set the use_namespaces option to True
use_neutron=True
use_namespaces=True
Sahara needs to access instances through ssh during a Cluster setup. To establish a connection Sahara may use both: fixed and floating IP of an Instance. By default use_floating_ips parameter is set to True, so Sahara will use Floating IP of an Instance to connect. In this case, user has two options for how to make all instances get a floating IP:
Note: When using floating IPs for management (use_floating_ip=True) every instance in the Cluster should have a floating IP, otherwise Sahara will not be able to work with it.
If use_floating_ips parameter is set to False Sahara will use Instances’ fixed IPs for management. In this case the node where Sahara is running should have access to Instances’ fixed IP network. When OpenStack uses Neutron for networking, user will be able to choose fixed IP network for all instances in a Cluster.
One of the problems in Hadoop running on OpenStack is that there is no ability to control where machine is actually running. We cannot be sure that two new virtual machines are started on different physical machines. As a result, any replication with cluster is not reliable because all replicas may turn up on one physical machine. Anti-affinity feature provides an ability to explicitly tell Sahara to run specified processes on different compute nodes. This is especially useful for Hadoop datanode process to make HDFS replicas reliable.
Starting with Juno release Sahara creates server groups with anti-affinity policy to enable anti affinity feature. Sahara creates one server group per cluster and assigns all instances with affected processes to this server group. Refer to Nova documentation on how server groups work.
This feature is supported by all plugins out of the box.
It is extremely important for data processing to do locally (on the same rack, openstack compute node or even VM) as much work as possible. Hadoop supports data-locality feature and can schedule jobs to tasktracker nodes that are local for input stream. In this case tasktracker could communicate directly with local data node.
Sahara supports topology configuration for HDFS and Swift data sources.
To enable data-locality set enable_data_locality parameter to True in Sahara configuration file
enable_data_locality=True
In this case two files with topology must be provided to Sahara. Options compute_topology_file and swift_topology_file parameters control location of files with compute and swift nodes topology descriptions correspondingly.
compute_topology_file should contain mapping between compute nodes and racks in the following format:
compute1 /rack1
compute1 /rack2
compute1 /rack2
Note that compute node name must be exactly the same as configured in openstack (host column in admin list for instances).
swift_topology_file should contain mapping between swift nodes and racks in the following format:
node1 /rack1
node2 /rack2
node3 /rack2
Note that swift node must be exactly the same as configures in object.builder swift ring. Also make sure that VMs with tasktracker service has direct access to swift nodes.
Hadoop versions after 1.2.0 support four-layer topology (https://issues.apache.org/jira/browse/HADOOP-8468). To enable this feature set enable_hypervisor_awareness option to True in Sahara configuration file. In this case Sahara will add compute node ID as a second level of topology for Virtual Machines.
Sahara allows you to control which security groups will be used for created instances. This can be done by providing the security_groups parameter for the Node Group or Node Group Template. By default an empty list is used that will result in using the default security group.
Sahara may also create a security group for instances in node group automatically. This security group will only have open ports which are required by instance processes or the Sahara engine. This option is useful for development and secured from outside environments, but for production environments it is recommended to control security group policy manually.
Sahara may use OpenStack Orchestration engine (aka Heat) to provision nodes for Hadoop cluster. To make Sahara work with Heat the following steps are required:
# An engine which will be used to provision infrastructure for Hadoop cluster. (string value)
infrastructure_engine=heat
There is a feature parity between direct and heat infrastructure engines. It is recommended to use heat engine since direct engine will be deprecated at some point.
The below tables provides a plugin capability matrix:
| Feature | Plugin | |||
|---|---|---|---|---|
| Vanilla | HDP | Cloudera | Spark | |
| Nova and Neutron network | x | x | x | x |
| Cluster Scaling | x | Scale Up | x | x |
| Swift Integration | x | x | x | N/A |
| Cinder Support | x | x | x | x |
| Data Locality | x | x | N/A | x |
| EDP | x | x | x | x |
Warning
Currently distributed mode for Sahara is in alpha state. We do not recommend using it in production environment.
The installation guide suggests to launch Sahara as a single ‘sahara-all’ process. It is also possible to run Sahara in distributed mode with ‘sahara-api’ and ‘sahara-engine’ processes running on several machines simultaneously.
Sahara-api works as a frontend and serves users’ requests. It offloads ‘heavy’ tasks to sahara-engine via RPC mechanism. While sahara-engine could be loaded, sahara-api by design stays free and hence may quickly respond on user queries.
If Sahara runs on several machines, the API requests could be balanced between several sahara-api instances using a load balancer. It is not required to balance load between different sahara-engine instances, as that will be automatically done via a message queue.
If a single machine goes down, others will continue serving users’ requests. Hence a better scalability is achieved and some fault tolerance as well. Note that the proposed solution is not a true High Availability. While failure of a single machine does not affect work of other machines, all of the operations running on the failed machine will stop. For example, if a cluster scaling is interrupted, the cluster will be stuck in a half-scaled state. The cluster will probably continue working, but it will be impossible to scale it further or run jobs on it via EDP.
To run Sahara in distributed mode pick several machines on which you want to run Sahara services and follow these steps:
On each machine install and configure Sahara using the installation guide except:
- Do not run ‘sahara-db-manage’ or launch Sahara with ‘sahara-all’
- Make sure sahara.conf provides database connection string to a single database on all machines.
Run ‘sahara-db-manage’ as described in the installation guide, but only on a single (arbitrarily picked) machine.
sahara-api and sahara-engine processes use oslo.messaging to communicate with each other. You need to configure it properly on each node (see below).
run sahara-api and sahara-engine on the desired nodes. On a node you can run both sahara-api and sahara-engine or you can run them on separate nodes. It does not matter as long as they are configured to use the same message broker and database.
To configure oslo.messaging, first you need to pick the driver you are going to use. Right now three drivers are provided: Rabbit MQ, Qpid or Zmq. To use Rabbit MQ or Qpid driver, you will have to setup messaging broker. The picked driver must be supplied in sahara.conf in [DEFAULT]/rpc_backend parameter. Use one the following values: rabbit, qpid or zmq. Next you have to supply driver-specific options.
Unfortunately, right now there is no documentation with description of drivers’ configuration. The options are available only in source code.
- For Rabbit MQ see
- rabbit_opts variable in impl_rabbit.py
- amqp_opts variable in amqp.py
- For Qpid see
- qpid_opts variable in impl_qpid.py
- amqp_opts variable in amqp.py
- For Zmq see
- zmq_opts variable in impl_zmq.py
- matchmaker_opts variable in matchmaker.py
- matchmaker_redis_opts variable in matchmaker_redis.py
- matchmaker_opts variable in matchmaker_ring.py
You can find the same options defined in sahara.conf.sample. You can use it to find section names for each option (matchmaker options are defined not in [DEFAULT])