velero/design/Implemented/node-agent-load-soothing.md

Node-agent Load Soothing Design

Glossary & Abbreviation

Velero Generic Data Path (VGDP): VGDP is the collective of modules that is introduced in Unified Repository design. Velero uses these modules to finish data transfer for various purposes (i.e., PodVolume backup/restore, Volume Snapshot Data Movement). VGDP modules include uploaders and the backup repository.

Background

As mentioned in node-agent Concurrency design, CSI Snapshot Data Movement design, VGDP Micro Service design and VGDP Micro Service for fs-backup design, all data movement activities for CSI snapshot data movement backups/restores and fs-backup respect the loadConcurrency settings configured in the node-agent-configmap. Once the number of existing loads exceeds the corresponding loadConcurrency setting, the loads will be throttled and some loads will be held until VGDP quotas are available.
However, this throttling only happens after the data mover pod is started and gets to running. As a result, when there are large number of concurrent volume backups, there may be many data mover pods get created but the VGDP instances inside them are actually on hold because of the VGDP throttling.
This could cause below problems:

• In some environments, there is a pod limit in each node of the cluster or a pod limit throughout the cluster, too many of the inactive data mover pods may block other pods from running
• In some environments, the system disk for each node of the cluster is limited, while pods also occupy system disk space, etc., many of the inactive data mover pods also take unnecessary space from system disk and cause other critical pods evicted
• For CSI snapshot data movement backup, before creation of the data mover pod, the volume snapshot has also created, this means excessive number of snapshots may also be created and live for longer time since the VGDP won't start until the quota is available. However, in some environments, large number of snapshots is not allowed or may cause degradation of the storage peroformance

On the other hand, the VGDP throttling mentioned in node-agent Concurrency design is an accurate controlling mechanism, that is, exactly the required number of data mover pods are throttled.

Therefore, another mechanism is required to soothe the creation of the data mover pods and volume snapshots before the VGDP throttling. It doesn't need to accurately control these creations but should effectively reduce the excessive number of inactive data mover pods and volume snapshots.
It is not practical to make an accurate control as it is almost impossible to predict which group of nodes a data mover pod is scheduled to, under the consideration of many complex factors, i.e., selected node, affinity, node OS, etc.

Goals

• Allow users to configure the expected number of loads pending on waiting for VGDP load concurrency quota
• Create a soothing mechanism to prevent new loads from starting if the number of existing loads excceds the expected number

Non-Goals

• Accurately controlling the loads from initiation is not a goal

Solution

We introduce a new field prepareQueueLength in loadConcurrency of node-agent-configmap as the allowed number of loads that are under preparing (expose). Specifically, loads are in this situation after its CR is in Accepted and Prepared phase. The prepareQueueLength should be a positive number, negative numbers will be ignored.
Once the value is set, the soothing mechanism takes effect, as the best effort, only the allowed number of CRs go into Accepted or Prepared phase, others will wait and stay as New state; and thereby only the allowed number of data mover pods, volume snapshots are created.
Otherwise, node-agent works the same as the legacy behavior, CRs go to Accepted or Prepared state as soon as the controllers process them and data mover pods and volume snapshots are also created without any constraints.
If users want to constrain the excessive number of pending data mover pods and volume snapshots, they could set a value by considering the VGDP load concurrency; otherwise, if they don't see constrains for pods or volume snapshots in their environment, they don't need to use this feature, in parallel preparing could also be beneficial for increasing the concurrency.

Node-agent server checks this configuration at startup time and use it to initiate the related VGDP modules. Therefore, users could edit this configMap any time, but in order to make the changes effective, node-agent server needs to be restarted.

The data structure is as below:

type LoadConcurrency struct {
    // GlobalConfig specifies the concurrency number to all nodes for which per-node config is not specified
    GlobalConfig int `json:"globalConfig,omitempty"`

    // PerNodeConfig specifies the concurrency number to nodes matched by rules
    PerNodeConfig []RuledConfigs `json:"perNodeConfig,omitempty"`

    // PrepareQueueLength specifies the max number of loads that are under expose
	PrepareQueueLength int `json:"prepareQueueLength,omitempty"`    
}

Sample

A sample of the ConfigMap is as below:

{
    "loadConcurrency": {
        "globalConfig": 2,
        "perNodeConfig": [
            {
                "nodeSelector": {
                    "matchLabels": {
                        "kubernetes.io/hostname": "node1"
                    }
                },
                "number": 3
            },
            {
                "nodeSelector": {
                    "matchLabels": {
                        "beta.kubernetes.io/instance-type": "Standard_B4ms"
                    }
                },
                "number": 5
            }
        ],
        "prepareQueueLength": 2
    }
}

To create the configMap, users need to save something like the above sample to a json file and then run below command:

kubectl create cm <ConfigMap name> -n velero --from-file=<json file name>

Detailed Design

Changes apply to the DataUpload Controller, DataDownload Controller, PodVolumeBackup Controller and PodVolumeRestore Controller, as below:

1. The soothe happens to data mover CRs (DataUpload, DataDownload, PodVolumeBackup or PodVolumeRestore) that are in New state
2. Before starting processing the CR, the corresponding controller counts the existing CRs under or pending for expose in the cluster, that is a total number of existing DataUpload, DataDownload, PodVolumeBackup and PodVolumeRestore that are in either Accepted or Preparing state
3. If the total number doesn't exceed the allowed number, the controller set the CR's phase to Accepted
4. Once the total number exceeds the allowed number, the controller gives up processing the CR and have it requeued later. The delay for the requeue is 5 seconds

The count happens for all the controllers in all nodes, to prevent the checks drain out the API server, the count happens to controller client caches for those CRs. And the count result is also cached, so that the count only happens whenever necessary. Below shows how it judges the necessity:

• When one or more CRs' phase change to Accepted
• When one or more CRs' phase change from Accepted to one of the terminal phases
• When one or more CRs' phase change from Prepared to one of the terminal phases
• When one or more CRs' phase change from Prepared to InProgress

Ideally, 2~3 in the above steps need to be synchornized among controllers in all nodes. However, this synchronization is not implemented, the consideration is as below:

1. It is impossible to accurately synchronize the count among controllers in different nodes, because the client cache is not coherrent among nodes.
2. It is possible to synchronize the count among controllers in the same node. However, it is too expensive to make this synchronization, because 2~3 are part of the expose workflow, the synchronization impacts the performance and stability of the existing workflow.
3. Even without the synchronization, the soothing mechanism still works eventually -- when the controllers see all the discharged loads (expected ones and over-discharged ones), they will stop creating new loads until the quota is available again.
4. Step 2~3 that need to be synchronized could complete very quickly.

This is why we say this mechanism is not an accurate control. Or in another word, it is possible that more loads than the number of prepareQueueLength are discharged if controllers make the count and expose in the overlapped time (step 2~3).
For example, when multiple controllers of the same type (DataUpload, DataDownload, PodVolumeBackup or PodVolumeRestore) from different nodes make the count:

max number of waiting loads = number defined by `prepareQueueLength` + number of nodes in cluster

As another example, when hybrid loads are running the count concurrently, e.g., mix of data mover backups, data mover restores, pod volume backups or pod volume restores, more loads may be discharged and the number depends on the number of concurrent hybrid loads.
In either case, because step 2~3 is short in time, it is less likely to reach the theoretically worset result.