add v3.6 to v3.5 downgrade support automatically

server/config/config.go

@@ -183,6 +183,12 @@ type ServerConfig struct {
 	// consider running defrag during bootstrap. Needs to be set to non-zero value to take effect.
 	ExperimentalBootstrapDefragThresholdMegabytes uint `json:"experimental-bootstrap-defrag-threshold-megabytes"`
 
+	// UnsafeAllowClusterVersionDowngrade is "true" to allow cluster version downgrade.
+	// "false" by default, since newer minor versions may introduce incompatible feature changes.
+	// For instance, lease checkpointer request to 3.4 will fail the remaining 3.3 nodes.
+	// But, if one does not use "lease checkpointer" feature, it can be safe to run 3.3 along with 3.4.
+	UnsafeAllowClusterVersionDowngrade bool `json:"unsafe-allow-cluster-version-downgrade"`
+
 	// V2Deprecation defines a phase of v2store deprecation process.
 	V2Deprecation V2DeprecationEnum `json:"v2-deprecation"`
 }

server/embed/config.go

@@ -89,6 +89,10 @@ const (
 	// v2 API is disabled by default.
 	DefaultEnableV2 = false
 
+	// DefaultUnsafeAllowClusterVersionDowngrade is the default value for "unsafe-allow-cluster-version-downgrade" flag.
+	// unsafe allow cluster version downgrade is disabled by default
+	DefaultUnsafeAllowClusterVersionDowngrade = false
+
 	// maxElectionMs specifies the maximum value of election timeout.
 	// More details are listed in ../Documentation/tuning.md#time-parameters.
 	maxElectionMs = 50000
@@ -392,6 +396,12 @@ type Config struct {
 	// ExperimentalTxnModeWriteWithSharedBuffer enables write transaction to use a shared buffer in its readonly check operations.
 	ExperimentalTxnModeWriteWithSharedBuffer bool `json:"experimental-txn-mode-write-with-shared-buffer"`
 
+	// UnsafeAllowClusterVersionDowngrade is "true" to allow cluster version downgrade.
+	// "false" by default, since newer minor versions may introduce incompatible feature changes.
+	// For instance, lease checkpointer request to 3.4 will fail the remaining 3.3 nodes.
+	// But, if one does not use "lease checkpointer" feature, it can be safe to run 3.3 along with 3.4.
+	UnsafeAllowClusterVersionDowngrade bool `json:"unsafe-allow-cluster-version-downgrade"`
+
 	// V2Deprecation describes phase of API & Storage V2 support
 	V2Deprecation config.V2DeprecationEnum `json:"v2-deprecation"`
 }
@@ -489,7 +499,8 @@ func NewConfig() *Config {
 		ExperimentalMemoryMlock:                  false,
 		ExperimentalTxnModeWriteWithSharedBuffer: true,
 
-		V2Deprecation: config.V2_DEPR_DEFAULT,
+		UnsafeAllowClusterVersionDowngrade: DefaultUnsafeAllowClusterVersionDowngrade,
+		V2Deprecation:                      config.V2_DEPR_DEFAULT,
 	}
 	cfg.InitialCluster = cfg.InitialClusterFromName(cfg.Name)
 	return cfg

server/embed/etcd.go

@@ -223,7 +223,8 @@ func StartEtcd(inCfg *Config) (e *Etcd, err error) {
 		ExperimentalMemoryMlock:                  cfg.ExperimentalMemoryMlock,
 		ExperimentalTxnModeWriteWithSharedBuffer: cfg.ExperimentalTxnModeWriteWithSharedBuffer,
 		ExperimentalBootstrapDefragThresholdMegabytes: cfg.ExperimentalBootstrapDefragThresholdMegabytes,
-		V2Deprecation: cfg.V2DeprecationEffective(),
+		UnsafeAllowClusterVersionDowngrade:            cfg.UnsafeAllowClusterVersionDowngrade,
+		V2Deprecation:                                 cfg.V2DeprecationEffective(),
 	}
 
 	if srvcfg.ExperimentalEnableDistributedTracing {

server/etcdmain/config.go

@@ -291,6 +291,7 @@ func newConfig() *config {
 
 	// unsafe
 	fs.BoolVar(&cfg.ec.UnsafeNoFsync, "unsafe-no-fsync", false, "Disables fsync, unsafe, will cause data loss.")
+	fs.BoolVar(&cfg.ec.UnsafeAllowClusterVersionDowngrade, "unsafe-allow-cluster-version-downgrade", embed.DefaultUnsafeAllowClusterVersionDowngrade, "true to allow cluster version downgrade, because newer minor versions may introduce incompatible feature changes like lease checkpointer introduced in v3.4")
 	fs.BoolVar(&cfg.ec.ForceNewCluster, "force-new-cluster", false, "Force to create a new one member cluster.")
 
 	// ignored

server/etcdmain/help.go

@@ -252,6 +252,9 @@ Unsafe feature:
     Force to create a new one-member cluster.
   --unsafe-no-fsync 'false'
     Disables fsync, unsafe, will cause data loss.
+  --unsafe-allow-cluster-version-downgrade 'false'
+    Allow cluster version downgrade, unsafe, newer minor versions may introduce incompatible feature changes.
+    For instance, experimental lease checkpointer is enabled in 3.4 and downgrade to 3.3 will fail. 
 
 CAUTIOUS with unsafe flag! It may break the guarantees given by the consensus protocol!
 `

server/etcdserver/api/membership/cluster.go

@@ -52,6 +52,9 @@ type RaftCluster struct {
 	v2store v2store.Store
 	be      backend.Backend
 
+	// Readonly field after initialization
+	unsafeAllowDowngrade bool
+
 	sync.Mutex // guards the fields below
 	version    *semver.Version
 	members    map[types.ID]*Member
@@ -268,7 +271,7 @@ func (c *RaftCluster) Recover(onSet func(*zap.Logger, *semver.Version)) {
 	if c.downgradeInfo != nil {
 		d = &DowngradeInfo{Enabled: c.downgradeInfo.Enabled, TargetVersion: c.downgradeInfo.TargetVersion}
 	}
-	mustDetectDowngrade(c.lg, c.version, d)
+	mustDetectDowngrade(c.lg, c.version, d, c.unsafeAllowDowngrade)
 	onSet(c.lg, c.version)
 
 	for _, m := range c.members {
@@ -536,7 +539,7 @@ func (c *RaftCluster) SetVersion(ver *semver.Version, onSet func(*zap.Logger, *s
 	}
 	oldVer := c.version
 	c.version = ver
-	mustDetectDowngrade(c.lg, c.version, c.downgradeInfo)
+	mustDetectDowngrade(c.lg, c.version, c.downgradeInfo, c.unsafeAllowDowngrade)
 	if c.v2store != nil {
 		mustSaveClusterVersionToStore(c.lg, c.v2store, ver)
 	}
@@ -550,6 +553,10 @@ func (c *RaftCluster) SetVersion(ver *semver.Version, onSet func(*zap.Logger, *s
 	onSet(c.lg, ver)
 }
 
+func (c *RaftCluster) AllowUnsafeDowngrade() {
+	c.unsafeAllowDowngrade = true
+}
+
 func (c *RaftCluster) IsReadyToAddVotingMember() bool {
 	nmembers := 1
 	nstarted := 0

server/etcdserver/api/membership/downgrade.go

@@ -38,7 +38,7 @@ func isValidDowngrade(verFrom *semver.Version, verTo *semver.Version) bool {
 }
 
 // mustDetectDowngrade will detect unexpected downgrade when the local server is recovered.
-func mustDetectDowngrade(lg *zap.Logger, cv *semver.Version, d *DowngradeInfo) {
+func mustDetectDowngrade(lg *zap.Logger, cv *semver.Version, d *DowngradeInfo, unsafeAllowDowngrade bool) {
 	lv := semver.Must(semver.NewVersion(version.Version))
 	// only keep major.minor version for comparison against cluster version
 	lv = &semver.Version{Major: lv.Major, Minor: lv.Minor}
@@ -63,14 +63,25 @@ func mustDetectDowngrade(lg *zap.Logger, cv *semver.Version, d *DowngradeInfo) {
 		)
 	}
 
+	// if downgrade is enabled, and it's one minor version down
+	// safe to not fail (e.g., local version 3.4, cluster version 3.5)
 	// if the cluster disables downgrade, check local version against determined cluster version.
 	// the validation passes when local version is not less than cluster version
 	if cv != nil && lv.LessThan(*cv) {
-		lg.Fatal(
-			"invalid downgrade; server version is lower than determined cluster version",
-			zap.String("current-server-version", version.Version),
-			zap.String("determined-cluster-version", version.Cluster(cv.String())),
-		)
+		if unsafeAllowDowngrade && isValidDowngrade(cv, lv) {
+			lg.Warn("allowing unsafe downgrade; local server version is lower than determined cluster version",
+				zap.String("current-server-version", version.Version),
+				zap.String("determined-cluster-version", version.Cluster(cv.String())),
+				zap.String("target-cluster-version", version.Cluster(lv.String())),
+			)
+			// overwrite the cluster version with local version determined by the etcd binary version
+			*cv = *lv
+		} else {
+			lg.Fatal("invalid downgrade, not allowed; local server version is lower than determined cluster version",
+				zap.String("current-server-version", version.Version),
+				zap.String("determined-cluster-version", version.Cluster(cv.String())),
+			)
+		}
 	}
 }
 

server/etcdserver/api/membership/downgrade_test.go

@@ -40,51 +40,66 @@ func TestMustDetectDowngrade(t *testing.T) {
 	downgradeDisabled := &DowngradeInfo{Enabled: false}
 
 	tests := []struct {
-		name           string
-		clusterVersion *semver.Version
-		downgrade      *DowngradeInfo
-		success        bool
-		message        string
+		name                 string
+		clusterVersion       *semver.Version
+		downgrade            *DowngradeInfo
+		unsafeAllowDowngrade bool
+		success              bool
+		message              string
 	}{
 		{
 			"Succeeded when downgrade is disabled and cluster version is nil",
 			nil,
 			downgradeDisabled,
+			false,
 			true,
 			"",
 		},
 		{
 			"Succeeded when downgrade is disabled and cluster version is one minor lower",
 			oneMinorLower,
 			downgradeDisabled,
+			false,
 			true,
 			"",
 		},
 		{
 			"Succeeded when downgrade is disabled and cluster version is server version",
 			lv,
 			downgradeDisabled,
+			false,
 			true,
 			"",
 		},
+		{
+			"Succeed when downgrade is disabled, unsafeDowngrade is enabled and cluster version is one minor higher",
+			oneMinorHigher,
+			downgradeDisabled,
+			true,
+			true,
+			"allowing unsafe downgrade; local server version is lower than determined cluster version",
+		},
 		{
 			"Failed when downgrade is disabled and server version is lower than determined cluster version ",
 			oneMinorHigher,
 			downgradeDisabled,
 			false,
-			"invalid downgrade; server version is lower than determined cluster version",
+			false,
+			"invalid downgrade, not allowed; local server version is lower than determined cluster version",
 		},
 		{
 			"Succeeded when downgrade is enabled and cluster version is nil",
 			nil,
 			downgradeEnabledEqualVersion,
+			false,
 			true,
 			"",
 		},
 		{
 			"Failed when downgrade is enabled and server version is target version",
 			lv,
 			downgradeEnabledEqualVersion,
+			false,
 			true,
 			"cluster is downgrading to target version",
 		},
@@ -93,21 +108,23 @@ func TestMustDetectDowngrade(t *testing.T) {
 			lv,
 			downgradeEnabledLowerVersion,
 			false,
+			false,
 			"invalid downgrade; server version is not allowed to join when downgrade is enabled",
 		},
 		{
 			"Failed when downgrade is enabled and local version is out of range and cluster version is nil",
 			nil,
 			downgradeEnabledHigherVersion,
 			false,
+			false,
 			"invalid downgrade; server version is not allowed to join when downgrade is enabled",
 		},
-
 		{
 			"Failed when downgrade is enabled and local version is out of range",
 			lv,
 			downgradeEnabledHigherVersion,
 			false,
+			false,
 			"invalid downgrade; server version is not allowed to join when downgrade is enabled",
 		},
 	}
@@ -122,7 +139,7 @@ func TestMustDetectDowngrade(t *testing.T) {
 		lcfg.ErrorOutputPaths = []string{logPath}
 		lg, _ := lcfg.Build()
 
-		mustDetectDowngrade(lg, tests[iint].clusterVersion, tests[iint].downgrade)
+		mustDetectDowngrade(lg, tests[iint].clusterVersion, tests[iint].downgrade, tests[iint].unsafeAllowDowngrade)
 		return
 	}
 

server/etcdserver/cluster_util.go

@@ -200,9 +200,10 @@ func decideClusterVersion(lg *zap.Logger, vers map[string]*version.Versions) *se
 }
 
 // allowedVersionRange decides the available version range of the cluster that local server can join in;
-// if the downgrade enabled status is true, the version window is [oneMinorHigher, oneMinorHigher]
-// if the downgrade is not enabled, the version window is [MinClusterVersion, localVersion]
-func allowedVersionRange(downgradeEnabled bool) (minV *semver.Version, maxV *semver.Version) {
+// if the downgrade enabled status is true, the version window is [oneMinorHigherThanLocalVersion, oneMinorHigherThanLocalVersion]
+// otherwise, if the unsafeDowngrade enabled status is true, the version window is [MinClusterVersion, oneMinorHigherThanLocalVersion],
+// if the both downgrade and unsafeDowngrade is not enabled, the version window is [MinClusterVersion, localVersion]
+func allowedVersionRange(downgradeEnabled bool, unsafeDowngradeEnabled bool) (minV *semver.Version, maxV *semver.Version) {
 	minV = semver.Must(semver.NewVersion(version.MinClusterVersion))
 	maxV = semver.Must(semver.NewVersion(version.Version))
 	maxV = &semver.Version{Major: maxV.Major, Minor: maxV.Minor}
@@ -211,7 +212,15 @@ func allowedVersionRange(downgradeEnabled bool) (minV *semver.Version, maxV *sem
 		// Todo: handle the case that downgrading from higher major version(e.g. downgrade from v4.0 to v3.x)
 		maxV.Minor = maxV.Minor + 1
 		minV = &semver.Version{Major: maxV.Major, Minor: maxV.Minor}
+		return minV, maxV
 	}
+
+	// if unsafeDowngrade is enabled, and one minor version down
+	// safe to not fail (e.g., local version 3.4, cluster version 3.5)
+	if unsafeDowngradeEnabled {
+		maxV.Minor = maxV.Minor + 1
+	}
+
 	return minV, maxV
 }
 
@@ -221,9 +230,9 @@ func allowedVersionRange(downgradeEnabled bool) (minV *semver.Version, maxV *sem
 // cluster version in the range of [MinV, MaxV] and no known members has a cluster version
 // out of the range.
 // We set this rule since when the local member joins, another member might be offline.
-func isCompatibleWithCluster(lg *zap.Logger, cl *membership.RaftCluster, local types.ID, rt http.RoundTripper) bool {
+func isCompatibleWithCluster(lg *zap.Logger, cl *membership.RaftCluster, local types.ID, rt http.RoundTripper, unsafeAllowClusterVersionDowngrade bool) bool {
 	vers := getVersions(lg, cl, local, rt)
-	minV, maxV := allowedVersionRange(getDowngradeEnabledFromRemotePeers(lg, cl, local, rt))
+	minV, maxV := allowedVersionRange(getDowngradeEnabledFromRemotePeers(lg, cl, local, rt), unsafeAllowClusterVersionDowngrade)
 	return isCompatibleWithVers(lg, vers, local, minV, maxV)
 }
 
@@ -256,12 +265,13 @@ func isCompatibleWithVers(lg *zap.Logger, vers map[string]*version.Versions, loc
 			)
 			return false
 		}
+
 		if maxV.LessThan(*clusterv) {
 			lg.Warn(
 				"cluster version of remote member is not compatible; too high",
 				zap.String("remote-member-id", id),
 				zap.String("remote-member-cluster-version", clusterv.String()),
-				zap.String("minimum-cluster-version-supported", minV.String()),
+				zap.String("maximum-cluster-version-supported", maxV.String()),
 			)
 			return false
 		}

server/etcdserver/cluster_util_test.go

@@ -183,28 +183,45 @@ func TestDecideAllowedVersionRange(t *testing.T) {
 	localV = &semver.Version{Major: localV.Major, Minor: localV.Minor}
 
 	tests := []struct {
-		name             string
-		downgradeEnabled bool
-		expectedMinV     *semver.Version
-		expectedMaxV     *semver.Version
+		name                   string
+		downgradeEnabled       bool
+		unsafeDowngradeEnabled bool
+		expectedMinV           *semver.Version
+		expectedMaxV           *semver.Version
 	}{
 		{
 			"When cluster enables downgrade",
 			true,
+			false,
+			&semver.Version{Major: localV.Major, Minor: localV.Minor + 1},
+			&semver.Version{Major: localV.Major, Minor: localV.Minor + 1},
+		},
+		{
+			"When cluster enables downgrade and unsafeDowngrade",
+			true,
+			true,
 			&semver.Version{Major: localV.Major, Minor: localV.Minor + 1},
 			&semver.Version{Major: localV.Major, Minor: localV.Minor + 1},
 		},
+		{
+			"When cluster disables downgrade and enables unsafeDowngrade",
+			false,
+			true,
+			minClusterV,
+			&semver.Version{Major: localV.Major, Minor: localV.Minor + 1},
+		},
 		{
 			"When cluster disables downgrade",
 			false,
+			false,
 			minClusterV,
 			localV,
 		},
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			minV, maxV := allowedVersionRange(tt.downgradeEnabled)
+			minV, maxV := allowedVersionRange(tt.downgradeEnabled, tt.unsafeDowngradeEnabled)
 			if !minV.Equal(*tt.expectedMinV) {
 				t.Errorf("Expected minV is %v; Got %v", tt.expectedMinV.String(), minV.String())
 			}

server/etcdserver/raft.go

@@ -488,6 +488,9 @@ func restartNode(cfg config.ServerConfig, snapshot *raftpb.Snapshot) (types.ID,
 	)
 	cl := membership.NewCluster(cfg.Logger)
 	cl.SetID(id, cid)
+	if cfg.UnsafeAllowClusterVersionDowngrade {
+		cl.AllowUnsafeDowngrade()
+	}
 	s := raft.NewMemoryStorage()
 	if snapshot != nil {
 		s.ApplySnapshot(*snapshot)
@@ -562,6 +565,9 @@ func restartAsStandaloneNode(cfg config.ServerConfig, snapshot *raftpb.Snapshot)
 
 	cl := membership.NewCluster(cfg.Logger)
 	cl.SetID(id, cid)
+	if cfg.UnsafeAllowClusterVersionDowngrade {
+		cl.AllowUnsafeDowngrade()
+	}
 	s := raft.NewMemoryStorage()
 	if snapshot != nil {
 		s.ApplySnapshot(*snapshot)