Scale Dataproc clusters

After creating a Dataproc cluster, you can adjust ("scale") the cluster by increasing or decreasing the number of primary or secondary worker nodes (horizontal scaling) in the cluster. You can scale a Dataproc cluster at any time, even when jobs are running on the cluster. You cannot change the machine type of an existing cluster (vertical scaling). To vertically scale, create a cluster using a supported machine type, then migrate jobs to the new cluster.

Use Dataproc Autoscaling. Instead of manually scaling clusters, enable Autoscaling to have Dataproc set the "right" number of workers for your workloads.

Why scale a Dataproc cluster?

1. to increase the number of workers to make a job run faster
2. to decrease the number of workers to save money (see Graceful Decommissioning as an option to use when downsizing a cluster to avoid losing work in progress).
3. to increase the number of nodes to expand available Hadoop Distributed Filesystem (HDFS) storage

Note: You cannot scale a single-node cluster.

Because clusters can be scaled more than once, you might want to increase/decrease the cluster size at one time, and then decrease/increase the size later.

Removing workers: If you remove workers from your cluster, make sure that the new cluster size is sufficient to handle your workload; if it isn't, your jobs may take a long time to complete or may stop responding.

Using Scaling

There are three ways you can scale your Dataproc cluster:

1. Use the gcloud command-line tool in the gcloud CLI.
2. Edit the cluster configuration in the Google Cloud console.
3. Use the REST API.

New workers added to a cluster will use the same machine type as existing workers. For example, if a cluster is created with workers that use the n1-standard-8 machine type, new workers will also use the n1-standard-8 machine type.

Note: Dataproc does not support changing worker size (vCPUs, memory, persistent disk) when you add workers to a cluster.

You can scale the number of primary workers or the number of secondary (preemptible) workers, or both. For example, if you only scale the number of preemptible workers, the number of primary workers remains the same.

gcloud

gcloud CLI setup: You must setup and configure the gcloud CLI to use the Google Cloud CLI.

To scale a cluster with gcloud dataproc clusters update, run the following command.

gcloud dataproc clusters update cluster-name \
    --region=region \
    [--num-workers and/or --num-secondary-workers]=new-number-of-workers

where cluster-name is the name of the cluster to update, and new-number-of-workers is the updated number of primary and/or secondary worker nodes. For example, to scale a cluster named "dataproc-1" to use five primary worker nodes, run the following command.

gcloud dataproc clusters update dataproc-1 \
    --region=region \
    --num-workers=5
...
Waiting on operation [operations/projects/project-id/operations/...].
Waiting for cluster update operation...done.
Updated [https://dataproc.googleapis.com/...].
clusterName: my-test-cluster
...
  masterDiskConfiguration:
    bootDiskSizeGb: 500
  masterName: dataproc-1-m
  numWorkers: 5
  ...
  workers:
  - my-test-cluster-w-0
  - my-test-cluster-w-1
  - my-test-cluster-w-2
  - my-test-cluster-w-3
  - my-test-cluster-w-4
...

REST API

See clusters.patch.

Example

PATCH /v1/projects/project-id/regions/us-central1/clusters/example-cluster?updateMask=config.worker_config.num_instances,config.secondary_worker_config.num_instances
{
  "config": {
    "workerConfig": {
      "numInstances": 4
    },
    "secondaryWorkerConfig": {
      "numInstances": 2
    }
  },
  "labels": null
}

Let the Console construct your cluster update request.: You can click the Equivalent REST link at the bottom of the Google Cloud console Dataproc Cluster details→Configuration page to have the Console construct an equivalent API REST PATCH request.

Console

After a cluster is created, you can scale a cluster by opening the Cluster details page for the cluster from the Google Cloud console Clusters page, then clicking the Edit button on the Configuration tab.

Enter a new value for the number of Worker nodes and/or Preemptible worker nodes (updated to "5" and " 2", respectively, in the following screenshot).

Click Save to update the cluster.

How Dataproc selects cluster nodes for removal

On clusters created with image versions 1.5.83+, 2.0.57+, and 2.1.5+, when scaling down a cluster, Dataproc attempts to minimize the impact of node removal on running YARN applications by first removing inactive, unhealthy, and idle nodes, then removing nodes with the fewest running YARN application masters and running containers.

Graceful Decommissioning

When you downscale a cluster, work in progress may stop before completion. If you are using Dataproc v 1.2 or later, you can use Graceful Decommissioning, which incorporates Graceful Decommission of YARN Nodes to finish work in progress on a worker before it is removed from the Cloud Dataproc cluster.

Note: A worker is not removed from a cluster until running YARN applications are finished.

Graceful Decommissioning and Secondary Workers

The preemptible (secondary) worker group continues to provision or delete workers to reach its expected size even after a cluster scaling operation is marked complete. If you attempt to gracefully decommission a secondary worker and receive an error message similar to the following:

"Secondary worker group cannot be modified outside of Dataproc. If you recently created or updated this cluster, wait a few minutes before gracefully decommissioning to allow all secondary instances to join or leave the cluster. Expected secondary worker group size: x, actual size: y",

wait a few minutes then repeat the graceful decommissioning request.

Also note:

• You can forcefully decommission preemptible workers at any time.
• You gracefully decommission primary workers at any time

Using Graceful Decommissioning

Dataproc Graceful Decommissioning incorporates Graceful Decommission of YARN Nodes to finish work in progress on a worker before it is removed from the Cloud Dataproc cluster. As a default, graceful decommissioning is disabled. You enable it by setting a timeout value when you update your cluster to remove one or more workers from the cluster.

gcloud

When you update a cluster to remove one or more workers, use the gcloud dataproc clusters update command with the --graceful-decommission-timeout flag. The timeout (string) values can be a value of "0s" (the default; forceful not graceful decommissioning) or a positive duration relative to the current time (for example, "3s"). The maximum duration is 1 day.

Durations: Specify a graceful decommissioning duration as a non-negative integer with one of the following lower-case suffixes: "s", "m", "h", or "d" for seconds, minutes, hours, or days, respectively. The maximum duration for graceful decommissioning is "1d" (one day). If the suffix is omitted, seconds is assumed.

gcloud dataproc clusters update cluster-name \
    --region=region \
    --graceful-decommission-timeout="timeout-value" \
    [--num-workers and/or --num-secondary-workers]=decreased-number-of-workers \
    ... other args ...

REST API

See clusters.patch.gracefulDecommissionTimeout. The timeout (string) values can be a value of "0" (the default; forceful not graceful decommissioning) or a duration in seconds (for example, "3s"). The maximum duration is 1 day.

Console

After a cluster is created, you can select graceful decommissioning of a cluster by opening the Cluster details page for the cluster from the Google Cloud console Clusters page, then clicking the Edit button on the Configuration tab.

In the Graceful Decommissioning section, check the "Use graceful decommissioning" box, then select a timeout value.

Click Save to update the cluster.

How to cancel a graceful decommissioning scaledown operation

On Dataproc clusters created with image versions 2.0.57+ or 2.1.5+, you can run the gcloud dataproc operations cancel command or issue a Dataproc API operations.cancel request to cancel a graceful decommissioning scaledown operation.

You can obtain the resource name, including the ID, of a scaledown operation by running gcloud dataproc operations list or issuing a Dataproc API operations.list request.

When you cancel a graceful decommissioning scaledown operation:

• workers in a DECOMMISSIONING state are re-commissioned and become ACTIVE at the completion of the operation's cancellation.

• if the scaledown operation includes label updates, the updates may not take effect.

Note: Cancellation of a scaledown operation is asynchronous. Dataproc makes a best effort to cancel the operation, but success of the cancellation is not guaranteed.

To verify the status of the cancellation request, you can run the gcloud dataproc operations describe command or issue a Dataproc API operations.get request. If the cancel operation succeeds, the inner operation status is marked as CANCELLED.