How-to: Persist Scheduler Jobs
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 is1Gi
, 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:
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:
dapr init -k --set dapr_scheduler.cluster.storageSize=16Gi --set dapr_scheduler.etcdSpaceQuota=16Gi
helm upgrade --install dapr dapr/dapr \
--version=1.14 \
--namespace dapr-system \
--create-namespace \
--set dapr_scheduler.cluster.storageSize=16Gi \
--set dapr_scheduler.etcdSpaceQuota=16Gi \
--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.
NAMESPACE NAME STATUS VOLUME CAPACITY ACCESS MODES STORAGECLASS VOLUMEATTRIBUTESCLASS AGE
dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-0 Bound pvc-9f699d2e-f347-43b0-aa98-57dcf38229c5 1Gi RWO standard <unset> 3m25s
dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-1 Bound pvc-f4c8be7b-ffbe-407b-954e-7688f2482caa 1Gi RWO standard <unset> 3m25s
dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-2 Bound pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f 1Gi RWO standard <unset> 3m25s
NAME CAPACITY ACCESS MODES RECLAIM POLICY STATUS CLAIM STORAGECLASS VOLUMEATTRIBUTESCLASS REASON AGE
pvc-9f699d2e-f347-43b0-aa98-57dcf38229c5 1Gi RWO Delete Bound dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-0 standard <unset> 4m24s
pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f 1Gi RWO Delete Bound dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-2 standard <unset> 4m24s
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:
- First, ensure that the storage class supports volume expansion, and that the
allowVolumeExpansion
field is set totrue
if it is not already.
apiVersion: storage.k8s.io/v1
kind: StorageClass
metadata:
name: standard
provisioner: my.driver
allowVolumeExpansion: true
...
- Delete the Scheduler StatefulSet whilst preserving the Bound Persistent Volume Claims.
kubectl delete sts -n dapr-system dapr-scheduler-server --cascade=orphan
- 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.
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
- 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:
- Google Cloud Persistent Disk
- Amazon EBS Volumes
- Azure AKS Storage Options
- Digital Ocean Block Storage
- VMWare vSphere Storage
- OpenShift Persistent Storage
- Alibaba Cloud Disk Storage
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 SchedulerStatefulSet
to be recreated with the new persistent volume.
dapr init -k --set dapr_scheduler.cluster.storageClassName=my-storage-class
helm upgrade --install dapr dapr/dapr \
--version=1.14 \
--namespace dapr-system \
--create-namespace \
--set dapr_scheduler.cluster.storageClassName=my-storage-class \
--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 SchedulerStatefulSet
to be recreated without the persistent volume.
dapr init -k --set dapr_scheduler.cluster.inMemoryStorage=true
helm upgrade --install dapr dapr/dapr \
--version=1.14 \
--namespace dapr-system \
--create-namespace \
--set dapr_scheduler.cluster.inMemoryStorage=true \
--wait
Feedback
Was this page helpful?
Glad to hear it! Please tell us how we can improve.
Sorry to hear that. Please tell us how we can improve.