From 6eb44eeab93ff0893295038df7a700a984218110 Mon Sep 17 00:00:00 2001 From: joshvanl Date: Mon, 16 Sep 2024 13:53:13 +0100 Subject: [PATCH] Helm: Revert Scheduler storage quota size to `1Gi` In v1.14.3, the storage quota size for the Scheduler volume was increased from `1Gi` to `16Gi`. This is because users where encountering disk exhaustion fatal errors on the Scheduler under normal usage. Because the volume size request field is protected from updates, Dapr version upgrades to v1.14.3 failed without manual intervention. Reverts the Scheduler storage quota size back to `1Gi`, and adds warnings that the volume size may need to be increased for production deployments. See: https://github.com/dapr/dapr/pull/8107 Signed-off-by: joshvanl --- .../kubernetes-persisting-scheduler.md | 115 +++++++++++++++++- 1 file changed, 114 insertions(+), 1 deletion(-) diff --git a/daprdocs/content/en/operations/hosting/kubernetes/kubernetes-persisting-scheduler.md b/daprdocs/content/en/operations/hosting/kubernetes/kubernetes-persisting-scheduler.md index 9172a28feb9..b4e8f02e64e 100644 --- a/daprdocs/content/en/operations/hosting/kubernetes/kubernetes-persisting-scheduler.md +++ b/daprdocs/content/en/operations/hosting/kubernetes/kubernetes-persisting-scheduler.md @@ -7,10 +7,123 @@ description: "Configure Scheduler to persist its database to make it resilient t --- The [Scheduler]({{< ref scheduler.md >}}) service is responsible for writing jobs to its embedded Etcd database and scheduling them for execution. -By default, the Scheduler service database writes this data to a Persistent Volume Claim of 1Gb of size using the cluster's default [storage class](https://kubernetes.io/docs/concepts/storage/storage-classes/). 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 in some deployments or for a production environment. +By default, the Scheduler service database writes data to a Persistent Volume Claim volume of size `1Gb`, using the cluster's default [storage class](https://kubernetes.io/docs/concepts/storage/storage-classes/). +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](#storage-class) if a default StorageClass is not available or when running a production environment. + +{{% alert title="Warning" color="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]({{< ref actors-timers-reminders.md >}}) & [Workflows]({{< ref workflow-overview.md >}}) when the [SchedulerReminders]({{< ref support-preview-features.md >}}) preview feature is enabled, and the [Jobs API]({{< ref jobs_api.md >}}). +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](#etcd-storage-disk-size) section below. +{{% /alert %}} ## 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]({{< ref jobs_api.md >}}) and [Actor Reminders]({{< ref actors-timers-reminders.md >}}) (with the [SchedulerReminders]({{< ref support-preview-features.md >}}) preview feature enabled) transparently maps one to one to the usage of your applications. +Workflows (when the [SchedulerReminders]({{< ref support-preview-features.md >}}) 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](https://etcd.io/docs/v3.5/learning/persistent-storage-files/). +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](#increase-existing-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: + +{{< tabs "Dapr CLI" "Helm" >}} + +{{% codetab %}} + +```bash +dapr init -k --set dapr_scheduler.cluster.storageSize=16Gi --set dapr_scheduler.etcdSpaceQuota=16Gi +``` + +{{% /codetab %}} + + +{{% codetab %}} + +```bash +helm upgrade --install dapr dapr/dapr \ +--version={{% dapr-latest-version short="true" %}} \ +--namespace dapr-system \ +--create-namespace \ +--set dapr_scheduler.cluster.storageSize=16Gi \ +--set dapr_scheduler.etcdSpaceQuota=16Gi \ +--wait +``` + +{{% /codetab %}} +{{< /tabs >}} + +#### Increase existing Scheduler Storage Size + +{{% alert title="Warning" color="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. +{{% /alert %}} + +By default, each Scheduler will create a Persistent Volume and Persistent Volume Claim of size `1Gi` against the [default `standard` storage class](#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 3m25s +dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-1 Bound pvc-f4c8be7b-ffbe-407b-954e-7688f2482caa 1Gi RWO standard 3m25s +dapr-system dapr-scheduler-data-dir-dapr-scheduler-server-2 Bound pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f 1Gi RWO standard 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 4m24s +pvc-eaad5fb1-98e9-42a5-bcc8-d45dba1c4b9f 1Gi RWO Delete Bound dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-2 standard 4m24s +pvc-f4c8be7b-ffbe-407b-954e-7688f2482caa 1Gi RWO Delete Bound dapr-system/dapr-scheduler-data-dir-dapr-scheduler-server-1 standard 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. + +```yaml +apiVersion: storage.k8s.io/v1 +kind: StorageClass +metadata: + name: standard +provisioner: my.driver +allowVolumeExpansion: true +... +``` + +2. Delete the Scheduler StatefulSet whilst preserving the Bound Persistent Volume Claims. + +```bash +kubectl delete sts -n dapr-system dapr-scheduler-server --cascade=orphan +``` + +3. 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. + +```bash +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 +``` + +4. Recreate the Scheduler StatefulSet by [installing Dapr with the desired storage size](#setting-the-storage-size-on-installation). + +### 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.