feat(1-1-restore): add unpin-agent-cpu flag

docs/source/sctool/partials/sctool_restore_1-1-restore.yaml

@@ -117,6 +117,10 @@ options:
       usage: |
         Timezone of --cron and --window flag values.
         The default value is taken from this system, namely 'TZ' envvar or '/etc/localtime' file.
+    - name: unpin-agent-cpu
+      default_value: "false"
+      usage: |
+        Defines if ScyllaDB Manager Agent should be unpinned from CPUs during restore.
     - name: window
       default_value: '[]'
       usage: |

docs/source/sctool/partials/sctool_restore_1-1-restore_update.yaml

@@ -118,6 +118,10 @@ options:
       usage: |
         Timezone of --cron and --window flag values.
         The default value is taken from this system, namely 'TZ' envvar or '/etc/localtime' file.
+    - name: unpin-agent-cpu
+      default_value: "false"
+      usage: |
+        Defines if ScyllaDB Manager Agent should be unpinned from CPUs during restore.
     - name: window
       default_value: '[]'
       usage: |

pkg/command/one2onerestore/cmd.go

@@ -32,6 +32,7 @@ type command struct {
 	keyspace      []string
 	snapshotTag   string
 	nodesMapping  nodesMapping
+	unpinAgentCPU bool
 	dryRun        bool
 }
 
@@ -83,6 +84,7 @@ func (cmd *command) init() {
 
 	// Common configuration for restore procedures
 	w.Unwrap().BoolVar(&cmd.dryRun, "dry-run", false, "")
+	w.Unwrap().BoolVar(&cmd.unpinAgentCPU, "unpin-agent-cpu", false, "")
 }
 
 func (cmd *command) run(args []string) error {
@@ -205,6 +207,11 @@ func flagsToTaskProperties(cmd *command, task *models.Task) (updated bool, err e
 			flagName: "nodes-mapping",
 			value:    cmd.nodesMapping,
 		},
+		{
+			flagName:     "unpin-agent-cpu",
+			value:        cmd.unpinAgentCPU,
+			canBeUpdated: true,
+		},
 	}
 
 	props := task.Properties.(map[string]interface{})

pkg/command/one2onerestore/res.yaml

@@ -28,3 +28,6 @@ nodes-mapping:
 dry-run: |
   Validates and displays restore information without actually running the restore.
   This allows you to display what will happen should the restore run with the parameters you set.
+
+unpin-agent-cpu: |
+  Defines if ScyllaDB Manager Agent should be unpinned from CPUs during restore.

pkg/service/one2onerestore/model.go

@@ -19,6 +19,7 @@ type Target struct {
 	SourceClusterID uuid.UUID             `json:"source_cluster_id"`
 	SnapshotTag     string                `json:"snapshot_tag"`
 	NodesMapping    []nodeMapping         `json:"nodes_mapping"`
+	UnpinAgentCPU   bool                  `json:"unpin_agent_cpu"`
 }
 
 func defaultTarget() Target {

pkg/service/one2onerestore/worker.go

@@ -79,6 +79,18 @@ func (w *worker) restore(ctx context.Context, workload []hostWorkload, target Ta
 		return errors.Wrap(err, "disable auto compaction")
 	}
 
+	// We always want to pin agent to CPUs outside the 1-1-restore.
+	defer func() {
+		if err := w.pinAgentCPU(context.Background(), workload, true); err != nil {
+			w.logger.Error(ctx, "Can't pin agent to CPU", "error", err)
+		}
+	}()
+	if target.UnpinAgentCPU {
+		if err := w.pinAgentCPU(ctx, workload, false); err != nil {
+			return errors.Wrap(err, "unpin agent from CPU")
+		}
+	}
+
 	if err := w.setTombstoneGCModeRepair(ctx, workload); err != nil {
 		return errors.Wrap(err, "tombstone_gc mode")
 	}
@@ -209,6 +221,22 @@ func (w *worker) setAutoCompaction(ctx context.Context, workload []hostWorkload,
 	return nil
 }
 
+func (w *worker) pinAgentCPU(ctx context.Context, workload []hostWorkload, pin bool) error {
+	setPinFunc := w.client.PinCPU
+	if !pin {
+		setPinFunc = w.client.UnpinFromCPU
+	}
+	return parallel.Run(len(workload), len(workload), func(i int) error {
+		host := workload[i].host
+		return errors.Wrapf(setPinFunc(ctx, host.Addr), "set CPU pinning on %s", host.Addr)
+	}, func(i int, err error) {
+		w.logger.Error(ctx, "Failed to change agent CPU pinning",
+			"host", workload[i].host.Addr,
+			"pinned", pin,
+			"error", err)
+	})
+}
+
 // alterSchemaRetryWrapper is useful when executing many statements altering schema,
 // as it might take more time for Scylla to process them one after another.
 // This wrapper exits on: success, context cancel, op returned non-timeout error or after maxTotalTime has passed.