How-to: Persist Scheduler Jobs

Configure Scheduler to persist its database to make it resilient to restarts

The Scheduler service is responsible for writing jobs to its embedded Etcd database and scheduling them for execution. By default, the Scheduler service database writes data to a Persistent Volume Claim volume of size 1Gb, using the cluster’s default storage class. This means that there is no additional parameter required to run the scheduler service reliably on most Kubernetes deployments, although you will need additional configuration if a default StorageClass is not available or when running a production environment.

Warning

The default storage size for the Scheduler is 1Gi, which is likely not sufficient for most production deployments. Remember that the Scheduler is used for Actor Reminders & Workflows when the SchedulerReminders preview feature is enabled, and the Jobs API. You may want to consider reinstalling Dapr with a larger Scheduler storage of at least 16Gi or more. For more information, see the ETCD Storage Disk Size section below.

Production Setup

ETCD Storage Disk Size

The default storage size for the Scheduler is 1Gb. This size is likely not sufficient for most production deployments. When the storage size is exceeded, the Scheduler will log an error similar to the following:

  1. error running scheduler: etcdserver: mvcc: database space exceeded

Knowing the safe upper bound for your storage size is not an exact science, and relies heavily on the number, persistence, and the data payload size of your application jobs. The Job API and Actor Reminders (with the SchedulerReminders preview feature enabled) transparently maps one to one to the usage of your applications. Workflows (when the SchedulerReminders preview feature is enabled) create a large number of jobs as Actor Reminders, however these jobs are short lived- matching the lifecycle of each workflow execution. The data payload of jobs created by Workflows is typically empty or small.

The Scheduler uses Etcd as its storage backend database. By design, Etcd persists historical transactions and data in form of Write-Ahead Logs (WAL) and snapshots. This means the actual disk usage of Scheduler will be higher than the current observable database state, often by a number of multiples.

Setting the Storage Size on Installation

If you need to increase an existing Scheduler storage size, see the Increase Scheduler Storage Size section below. To increase the storage size (in this example- 16Gi) for a fresh Dapr instalation, you can use the following command:

  1. dapr init -k --set dapr_scheduler.cluster.storageSize=16Gi --set dapr_scheduler.etcdSpaceQuota=16Gi
  1. helm upgrade --install dapr dapr/dapr \
  2. --version=1.14 \
  3. --namespace dapr-system \
  4. --create-namespace \
  5. --set dapr_scheduler.cluster.storageSize=16Gi \
  6. --set dapr_scheduler.etcdSpaceQuota=16Gi \
  7. --wait

Increase existing Scheduler Storage Size

Warning

Not all storage providers support dynamic volume expansion. Please see your storage provider documentation to determine if this feature is supported, and what to do if it is not.

By default, each Scheduler will create a Persistent Volume and Persistent Volume Claim of size 1Gi against the default standard storage class for each Scheduler replica. These will look similar to the following, where in this example we are running Scheduler in HA mode.

  1. NAMESPACE     NAME                                              STATUS   VOLUME                                     CAPACITY   ACCESS MODES   STORAGECLASS   VOLUMEATTRIBUTESCLASS   AGE
  2. dapr-system   dapr-scheduler-data-dir-dapr-scheduler-server-0   Bound    pvc-9f699d2e-f347-43b0-aa98-57dcf38229c5   1Gi        RWO            standard       <unset>                 3m25s
  3. dapr-system   dapr-scheduler-data-dir-dapr-scheduler-server-1   Bound    pvc-f4c8be7b-ffbe-407b-954e-7688f2482caa   1Gi        RWO            standard       <unset>                 3m25s
  4. dapr-system   dapr-scheduler-data-dir-dapr-scheduler-server-2   Bound    pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f   1Gi        RWO            standard       <unset>                 3m25s
  1. NAME                                       CAPACITY   ACCESS MODES   RECLAIM POLICY   STATUS   CLAIM                                                         STORAGECLASS   VOLUMEATTRIBUTESCLASS   REASON   AGE
  2. pvc-9f699d2e-f347-43b0-aa98-57dcf38229c5   1Gi        RWO            Delete           Bound    dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-0   standard       <unset>                          4m24s
  3. pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f   1Gi        RWO            Delete           Bound    dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-2   standard       <unset>                          4m24s
  4. pvc-f4c8be7b-ffbe-407b-954e-7688f2482caa   1Gi        RWO            Delete           Bound    dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-1   standard       <unset>                          4m24s

To expand the storage size of the Scheduler, follow these steps:

  1. First, ensure that the storage class supports volume expansion, and that the allowVolumeExpansion field is set to true if it is not already.
  1. apiVersion: storage.k8s.io/v1
  2. kind: StorageClass
  3. metadata:
  4.   name: standard
  5. provisioner: my.driver
  6. allowVolumeExpansion: true
  7. ...
  1. Delete the Scheduler StatefulSet whilst preserving the Bound Persistent Volume Claims.
  1. kubectl delete sts -n dapr-system dapr-scheduler-server --cascade=orphan
  1. Increase the size of the Persistent Volume Claims to the desired size by editing the spec.resources.requests.storage field. Again in this case, we are assuming that the Scheduler is running in HA mode with 3 replicas.
  1. kubectl edit pvc -n dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-0 dapr-scheduler-data-dir-dapr-scheduler-server-1 dapr-scheduler-data-dir-dapr-scheduler-server-2
  1. Recreate the Scheduler StatefulSet by installing Dapr with the desired storage size.

Storage Class

In case your Kubernetes deployment does not have a default storage class or you are configuring a production cluster, defining a storage class is required.

A persistent volume is backed by a real disk that is provided by the hosted Cloud Provider or Kubernetes infrastructure platform. Disk size is determined by how many jobs are expected to be persisted at once; however, 64Gb should be more than sufficient for most production scenarios. Some Kubernetes providers recommend using a CSI driver to provision the underlying disks. Below are a list of useful links to the relevant documentation for creating a persistent disk for the major cloud providers:

Once the storage class is available, you can install Dapr using the following command, with Scheduler configured to use the storage class (replace my-storage-class with the name of the storage class):

Note

If Dapr is already installed, the control plane needs to be completely uninstalled in order for the Scheduler StatefulSet to be recreated with the new persistent volume.

  1. dapr init -k --set dapr_scheduler.cluster.storageClassName=my-storage-class
  1. helm upgrade --install dapr dapr/dapr \
  2. --version=1.14 \
  3. --namespace dapr-system \
  4. --create-namespace \
  5. --set dapr_scheduler.cluster.storageClassName=my-storage-class \
  6. --wait

Ephemeral Storage

Scheduler can be optionally made to use Ephemeral storage, which is in-memory storage which is not resilient to restarts, i.e. all Job data will be lost after a Scheduler restart. This is useful for deployments where storage is not available or required, or for testing purposes.

Note

If Dapr is already installed, the control plane needs to be completely uninstalled in order for the Scheduler StatefulSet to be recreated without the persistent volume.

  1. dapr init -k --set dapr_scheduler.cluster.inMemoryStorage=true
  1. helm upgrade --install dapr dapr/dapr \
  2. --version=1.14 \
  3. --namespace dapr-system \
  4. --create-namespace \
  5. --set dapr_scheduler.cluster.inMemoryStorage=true \
  6. --wait

Last modified October 11, 2024: Fixed typo (#4389) (fe17926)