Always manage your E-MapReduce (EMR) cluster through the EMR console. Operations performed directly in the Elastic Compute Service (ECS) console—or through the ECS API—can cause cluster instability, data loss, or make the cluster unrecoverable.
If you perform any of the prohibited operations listed below, you are responsible for the resulting consequences.
What to avoid in the ECS console
The following operations bypass EMR's internal state management. Performing them leads to the outcomes described.
Operations marked with a data-loss risk can result in permanent, unrecoverable data loss. Plan your cluster architecture upfront to avoid needing to perform these operations.
| Operation | What goes wrong | Alternative |
|---|---|---|
Delete or modify the etc/hosts file | EMR cannot resolve service endpoints on cluster nodes, causing service exceptions. | Add entries to the hosts file instead of modifying or deleting existing ones. |
| Modify component configuration files in the ECS console | When the affected service restarts, EMR overwrites the changes. | Modify component parameters in the EMR console. |
| Redeploy ECS instances via the ECS console or ECS API | The EMR service is disrupted. | None |
| Attach disks to cluster nodes via the ECS console or ECS API | EMR cannot recognize or initialize the attached disks, making them unavailable. | Plan disk capacity when creating the cluster. |
| Detach disks from cluster nodes via the ECS console or ECS API | Data loss risk. EMR is unaware of the detach operation, and data on the disk may be lost. | Plan disk capacity when creating the cluster. |
| Remove core nodes via the ECS console or ECS API | Data loss risk. In-flight jobs on the removed nodes fail, and data stored on those nodes is lost. | Scale the cluster through the EMR console. |
| Remove master nodes via the ECS console or ECS API | Data loss risk. For high availability (HA) clusters: HDFS NameNode HA, YARN ResourceManager, or HBase HMaster switchover fails—you must purchase a new cluster to recover. For non-HA clusters: the cluster becomes permanently unavailable with no migration path. | Do not remove master nodes. Purchase a new cluster if reconfiguration is needed. |
| Remove task nodes via the ECS console or ECS API | Jobs running on the removed nodes fail immediately. | Remove task nodes through the EMR console. |
| Stop the MySQL service on the master node (Built-in MySQL type) | The MySQL instance on master-1-1 backs Hive MetaStore, Oozie, and Ranger. Stopping it cuts off these components from their metadata database. | Do not stop the built-in MySQL service. |
Change the MySQL root password on master-1-1 (Built-in MySQL type) | Components such as Hue or Ranger lose access to MySQL and stop working. | Do not change the MySQL root password for the built-in MySQL instance. |
| Modify the security group, VPC, or vSwitch of ECS instances via the ECS console or ECS API | Inter-node network connectivity breaks, and cluster components become unavailable. | Configure networking before creating the cluster. Contact support if changes are required after creation. |
| Change the billing method via the ECS console or ECS API | The billing method change cannot be reverted. | Change the billing method in the EMR console instead. Switching from pay-as-you-go to subscription is supported; switching from subscription to pay-as-you-go is not. See Switch from pay-as-you-go to subscription. |
| Delete EMR agent-related directories | The cluster stops functioning as expected. | Do not delete agent directories. Contact support if you need to troubleshoot agent issues. |
Common capacity and configuration questions
My cluster is running out of disk space.
Increase the capacity of an existing disk, or add core nodes through the EMR console. EMR clusters do not support the addition of disks.
I have too much disk capacity and want to reduce it.
Purchase a new cluster sized for your actual needs, then release the original cluster. See Create a cluster. EMR clusters do not support disk scale-down.
My cluster does not have enough compute capacity.
Add task nodes through the EMR console. See Scale out a cluster.
I have excess compute capacity and want to scale down.
The approach depends on your billing method:
Pay-as-you-go: Remove task nodes through the EMR console.
Subscription: Stop the YARN NodeManager on the target task node, switch that ECS instance's billing method to pay-as-you-go in the ECS console, then release the instance.
My cluster is running outdated component versions.
Purchase a new cluster with the version you need. See Create a cluster. Upgrading individual components on a running cluster is not supported.
I want to convert a non-HA cluster to an HA cluster.
Non-HA clusters cannot be converted to HA clusters. Purchase an HA cluster from the start if high availability is required.
I need to run third-party software or services on EMR.
Use bootstrap actions to install third-party software when creating the cluster. If you install software manually after the cluster is created, you must reinstall it each time you add nodes.