longhorn/enhancements/20221205-concurrent-backup-restore-limit.md
Chin-Ya Huang d57db31395 feat(lep): concurrent backup restore design
Ref: 4558

Signed-off-by: Chin-Ya Huang <chin-ya.huang@suse.com>
2022-12-08 22:37:40 +08:00

79 lines
2.6 KiB
Markdown

# Concurrent Backup Restore Per Node Limit
## Summary
Longhorn has no boundary on the number of concurrent volume backup restoring.
Having a new `concurrent-backup-restore-per-node-limit` setting allows the user to limit the concurring backup restoring. Setting this restriction lowers the potential risk of overloading the cluster when volumes restoring from backup concurrently. For ex: during the Longhorn system restore.
### Related Issues
https://github.com/longhorn/longhorn/issues/4558
## Motivation
### Goals
Introduce a new `concurrent-backup-restore-per-node-limit` setting to define the boundary of the concurrent volume backup restoring.
### Non-goals
`None`
## Proposal
1. Introduce a new `concurrent-backup-restore-per-node-limit` setting.
1. Track the number of per-node volumes restoring from backup with atomic count (thread-safe) in the engine monitor.
### User Stories
Allow the user to set the concurrent backup restore per node limit to control the risk of cluster overload when Longhorn volume is restoring from backup concurrently.
### User Experience In Detail
1. Longhorn holds the engine backup restore when the number of volume backups restoring on a node reaches the `concurrent-backup-restore-per-node-limit`.
1. The volume backup restore continues when the number of volume backups restoring on a node is below the limit.
## Design
### Implementation Overview
#### The `concurrent-backup-restore-per-node-limit` Setting
This setting controls how many engines on a node can restore the backup concurrently.
Longhorn engine monitor backs off when the volume [backup restoring count](#track-the-volume-backup-restoring-per-node) reaches the setting limit.
Set the value to **0** to disable backup restore.
```
Category = SettingCategoryGeneral,
Type = integer
Default = 5 # same as the default replica rebuilding number
```
#### Track the volume backup restoring per node
1. Create a new atomic counter in the engine controller.
```
type EngineController struct {
restoringCounter util.Counter
}
```
1. Pass the restoring counter to each of its engine monitors.
```
type EngineMonitor struct {
restoringCounter util.Counter
}
```
1. Increase the restoring counter before backup restore.
> Ignore DR volumes (volume.Status.IsStandby).
1. Decrease the restoring counter when the backup restore caller method ends
### Test plan
- Test the setting should block backup restore when creating multiple volumes from the backup at the same time.
- Test the setting should be per-node limited.
- Test the setting should not have effect on DR volumes.
### Upgrade strategy
`None`
## Note [optional]
`None`