Skip to content

SeaweedFS is a distributed storage system for object storage (S3), file systems, Iceberg table buckets, and Lance vector buckets, designed to handle billions of files with O(1) disk access and effortless horizontal scaling.

placement: let callers reject candidates placement cannot judge (#10581)

Chris Lu03388c4beb17
diff --git a/weed/placement/placement.go b/weed/placement/placement.go
index 274fb3e132..29f333cc8a 100644
--- a/weed/placement/placement.go
+++ b/weed/placement/placement.go
@@ -20,6 +20,12 @@ type PlacementPreference struct {
// Exclude names nodes already spoken for by the same plan, so a caller
// placing several copies does not stack them on one node.
Exclude map[string]bool
+ // Accept rejects candidates the caller cannot use for reasons placement does
+ // not model -- replica placement rules, a node already holding the volume.
+ // It receives the candidate's rack and data center because the constraints
+ // that need them are exactly the ones a bare node cannot express. Nil
+ // accepts everything.
+ Accept func(node *master_pb.DataNodeInfo, dataCenter, rack string) bool
// VolumeBytes is what this move will actually consume on the destination.
// Zero falls back to the tier's average volume size, which is all a caller
// planning a not-yet-created volume can know.
@@ -73,6 +79,9 @@ func PickTarget(topo *master_pb.TopologyInfo, pref PlacementPreference) *master_
if d == nil || d.VolumeCount >= d.MaxVolumeCount {
continue
}
+ if pref.Accept != nil && !pref.Accept(n.info, n.dc, n.rack) {
+ continue
+ }
c := candidate{node: n}
c.bytes, c.hasBytes = d.DiskFreeBytes, d.DiskTotalBytes != 0
c.slots = d.MaxVolumeCount - d.VolumeCount
diff --git a/weed/placement/placement_test.go b/weed/placement/placement_test.go
index f36faa8fcd..1b553bc872 100644
--- a/weed/placement/placement_test.go
+++ b/weed/placement/placement_test.go
@@ -224,3 +224,35 @@ func TestPickTargetReservesTheVolumeSize(t *testing.T) {
t.Fatalf("second pick %q, want a2 after 800 bytes were spent on a1", second.GetId())
}
}
+
+func TestPickTargetHonoursTheCallerPredicate(t *testing.T) {
+ // Constraints placement does not model -- replica placement, a node already
+ // holding the volume -- stay with the caller, which keeps them out of the
+ // ranking rather than duplicating them here.
+ topo := placementTopo(
+ placementNode("a1", 10, 1, 1000, 900),
+ placementNode("a2", 10, 1, 1000, 500),
+ )
+ got := PickTarget(topo, PlacementPreference{
+ Source: "z9", DiskType: types.SsdType,
+ Accept: func(n *master_pb.DataNodeInfo, dc, rack string) bool { return n.Id != "a1" },
+ })
+ if got.GetId() != "a2" {
+ t.Fatalf("got %q, want a2 -- a1 was rejected despite being emptier", got.GetId())
+ }
+}
+
+func TestPickTargetPredicateSeesRackAndDataCenter(t *testing.T) {
+ topo := placementTopo(placementNode("a1", 10, 1, 1000, 900))
+ var sawDc, sawRack string
+ PickTarget(topo, PlacementPreference{
+ Source: "z9", DiskType: types.SsdType,
+ Accept: func(n *master_pb.DataNodeInfo, dc, rack string) bool {
+ sawDc, sawRack = dc, rack
+ return true
+ },
+ })
+ if sawDc != "dc1" || sawRack != "a" {
+ t.Fatalf("predicate saw dc=%q rack=%q, want dc1/a", sawDc, sawRack)
+ }
+}