placement: supports survival preferences (pingcap#40613)

close pingcap#38605

Author: nolouch

ddl/ddl_api.go

@@ -3026,6 +3026,8 @@ func SetDirectPlacementOpt(placementSettings *model.PlacementSettings, placement
 		placementSettings.FollowerConstraints = stringVal
 	case ast.PlacementOptionVoterConstraints:
 		placementSettings.VoterConstraints = stringVal
+	case ast.PlacementOptionSurvivalPreferences:
+		placementSettings.SurvivalPreferences = stringVal
 	default:
 		return errors.Trace(errors.New("unknown placement policy option"))
 	}

ddl/placement/bundle.go

@@ -29,6 +29,7 @@ import (
 	"github.com/pingcap/tidb/tablecodec"
 	"github.com/pingcap/tidb/util/codec"
 	"golang.org/x/exp/slices"
+	"gopkg.in/yaml.v2"
 )
 
 // Refer to https://github.com/tikv/pd/issues/2701 .
@@ -123,7 +124,13 @@ func NewBundleFromConstraintsOptions(options *model.PlacementSettings) (*Bundle,
 			rules = append(rules, rule)
 		}
 	}
-
+	labels, err := newLocationLabelsFromSurvivalPreferences(options.SurvivalPreferences)
+	if err != nil {
+		return nil, err
+	}
+	for _, rule := range rules {
+		rule.LocationLabels = labels
+	}
 	return &Bundle{Rules: rules}, nil
 }
 
@@ -155,9 +162,17 @@ func NewBundleFromSugarOptions(options *model.PlacementSettings) (*Bundle, error
 
 	var rules []*Rule
 
+	locationLabels, err := newLocationLabelsFromSurvivalPreferences(options.SurvivalPreferences)
+	if err != nil {
+		return nil, err
+	}
+
 	// in case empty primaryRegion and regions, just return an empty bundle
 	if primaryRegion == "" && len(regions) == 0 {
 		rules = append(rules, NewRule(Voter, followers+1, NewConstraintsDirect()))
+		for _, rule := range rules {
+			rule.LocationLabels = locationLabels
+		}
 		return &Bundle{Rules: rules}, nil
 	}
 
@@ -195,6 +210,11 @@ func NewBundleFromSugarOptions(options *model.PlacementSettings) (*Bundle, error
 		}
 	}
 
+	// set location labels
+	for _, rule := range rules {
+		rule.LocationLabels = locationLabels
+	}
+
 	return &Bundle{Rules: rules}, nil
 }
 
@@ -223,6 +243,19 @@ func newBundleFromOptions(options *model.PlacementSettings) (bundle *Bundle, err
 	return bundle, err
 }
 
+// newLocationLabelsFromSurvivalPreferences will parse the survival preferences into location labels.
+func newLocationLabelsFromSurvivalPreferences(survivalPreferenceStr string) ([]string, error) {
+	if len(survivalPreferenceStr) > 0 {
+		labels := []string{}
+		err := yaml.UnmarshalStrict([]byte(survivalPreferenceStr), &labels)
+		if err != nil {
+			return nil, ErrInvalidSurvivalPreferenceFormat
+		}
+		return labels, nil
+	}
+	return nil, nil
+}
+
 // NewBundleFromOptions will transform options into the bundle.
 func NewBundleFromOptions(options *model.PlacementSettings) (bundle *Bundle, err error) {
 	bundle, err = newBundleFromOptions(options)
@@ -257,6 +290,15 @@ func (b *Bundle) String() string {
 func (b *Bundle) Tidy() error {
 	extraCnt := map[PeerRoleType]int{}
 	newRules := b.Rules[:0]
+
+	// One Bundle is from one PlacementSettings, rule share same location labels, so we can use the first rule's location labels.
+	var locationLabels []string
+	for _, rule := range b.Rules {
+		if len(rule.LocationLabels) > 0 {
+			locationLabels = rule.LocationLabels
+			break
+		}
+	}
 	for i, rule := range b.Rules {
 		// useless Rule
 		if rule.Count <= 0 {
@@ -300,6 +342,8 @@ func (b *Bundle) Tidy() error {
 				Key:    EngineLabelKey,
 				Values: []string{EngineLabelTiFlash},
 			}},
+			// the merged rule should have the same location labels with the original rules.
+			LocationLabels: locationLabels,
 		})
 	}
 	b.Rules = newRules

ddl/placement/bundle_test.go

@@ -887,6 +887,9 @@ func TestTidy(t *testing.T) {
 	bundle.Rules = append(bundle.Rules, rules3...)
 	bundle.Rules = append(bundle.Rules, rules4...)
 
+	for _, r := range bundle.Rules {
+		r.LocationLabels = []string{"zone", "host"}
+	}
 	chkfunc := func() {
 		require.NoError(t, err)
 		require.Len(t, bundle.Rules, 3)
@@ -901,6 +904,7 @@ func TestTidy(t *testing.T) {
 				Values: []string{EngineLabelTiFlash},
 			},
 		}, bundle.Rules[2].Constraints)
+		require.Equal(t, []string{"zone", "host"}, bundle.Rules[2].LocationLabels)
 	}
 	err = bundle.Tidy()
 	chkfunc()

ddl/placement/errors.go

@@ -29,6 +29,8 @@ var (
 	ErrInvalidConstraintsMapcnt = errors.New("label constraints in map syntax have invalid replicas")
 	// ErrInvalidConstraintsFormat is from rule.go.
 	ErrInvalidConstraintsFormat = errors.New("invalid label constraints format")
+	// ErrInvalidSurvivalPreferenceFormat is from rule.go.
+	ErrInvalidSurvivalPreferenceFormat = errors.New("survival preference format should be in format [xxx=yyy, ...]")
 	// ErrInvalidConstraintsRelicas is from rule.go.
 	ErrInvalidConstraintsRelicas = errors.New("label constraints with invalid REPLICAS")
 	// ErrInvalidBundleID is from bundle.go.

ddl/placement/rule.go

@@ -45,15 +45,16 @@ type RuleGroupConfig struct {
 
 // Rule is the core placement rule struct. Check https://github.com/tikv/pd/blob/master/server/schedule/placement/rule.go.
 type Rule struct {
-	GroupID     string       `json:"group_id"`
-	ID          string       `json:"id"`
-	Index       int          `json:"index,omitempty"`
-	Override    bool         `json:"override,omitempty"`
-	StartKeyHex string       `json:"start_key"`
-	EndKeyHex   string       `json:"end_key"`
-	Role        PeerRoleType `json:"role"`
-	Count       int          `json:"count"`
-	Constraints Constraints  `json:"label_constraints,omitempty"`
+	GroupID        string       `json:"group_id"`
+	ID             string       `json:"id"`
+	Index          int          `json:"index,omitempty"`
+	Override       bool         `json:"override,omitempty"`
+	StartKeyHex    string       `json:"start_key"`
+	EndKeyHex      string       `json:"end_key"`
+	Role           PeerRoleType `json:"role"`
+	Count          int          `json:"count"`
+	Constraints    Constraints  `json:"label_constraints,omitempty"`
+	LocationLabels []string     `json:"location_labels,omitempty"`
 }
 
 // TiFlashRule extends Rule with other necessary fields.

ddl/placement_policy_test.go

@@ -155,7 +155,8 @@ func TestPlacementPolicy(t *testing.T) {
 		"LEARNERS=1 " +
 		"LEARNER_CONSTRAINTS=\"[+region=cn-west-1]\" " +
 		"FOLLOWERS=3 " +
-		"FOLLOWER_CONSTRAINTS=\"[+disk=ssd]\"")
+		"FOLLOWER_CONSTRAINTS=\"[+disk=ssd]\"" +
+		"SURVIVAL_PREFERENCES=\"[region, zone]\"")
 
 	checkFunc := func(policyInfo *model.PolicyInfo) {
 		require.Equal(t, true, policyInfo.ID != 0)
@@ -168,6 +169,7 @@ func TestPlacementPolicy(t *testing.T) {
 		require.Equal(t, "[+region=cn-west-1]", policyInfo.LearnerConstraints)
 		require.Equal(t, model.StatePublic, policyInfo.State)
 		require.Equal(t, "", policyInfo.Schedule)
+		require.Equal(t, "[region, zone]", policyInfo.SurvivalPreferences)
 	}
 
 	// Check the policy is correctly reloaded in the information schema.
@@ -590,11 +592,16 @@ func TestCreateTableWithPlacementPolicy(t *testing.T) {
 	tk.MustExec("create placement policy x " +
 		"FOLLOWERS=2 " +
 		"CONSTRAINTS=\"[+disk=ssd]\" ")
+	tk.MustExec("create placement policy z " +
+		"FOLLOWERS=1 " +
+		"SURVIVAL_PREFERENCES=\"[region, zone]\"")
 	tk.MustExec("create placement policy y " +
 		"FOLLOWERS=3 " +
 		"CONSTRAINTS=\"[+region=bj]\" ")
 	tk.MustExec("create table t(a int)" +
 		"PLACEMENT POLICY=\"x\"")
+	tk.MustExec("create table tt(a int)" +
+		"PLACEMENT POLICY=\"z\"")
 	tk.MustQuery("SELECT TABLE_CATALOG, TABLE_SCHEMA, TABLE_NAME, TIDB_PLACEMENT_POLICY_NAME FROM information_schema.Tables WHERE TABLE_SCHEMA='test' AND TABLE_NAME = 't'").Check(testkit.Rows(`def test t x`))
 	tk.MustExec("create table t_range_p(id int) placement policy x partition by range(id) (" +
 		"PARTITION p0 VALUES LESS THAN (100)," +
@@ -617,7 +624,18 @@ func TestCreateTableWithPlacementPolicy(t *testing.T) {
 	require.Equal(t, "y", policyY.Name.L)
 	require.Equal(t, true, policyY.ID != 0)
 
-	tbl := external.GetTableByName(t, tk, "test", "t")
+	policyZ := testGetPolicyByName(t, tk.Session(), "z", true)
+	require.Equal(t, "z", policyZ.Name.L)
+	require.Equal(t, true, policyZ.ID != 0)
+	require.Equal(t, "[region, zone]", policyZ.SurvivalPreferences)
+
+	tbl := external.GetTableByName(t, tk, "test", "tt")
+	require.NotNil(t, tbl)
+	require.NotNil(t, tbl.Meta().PlacementPolicyRef)
+	require.Equal(t, "z", tbl.Meta().PlacementPolicyRef.Name.L)
+	require.Equal(t, policyZ.ID, tbl.Meta().PlacementPolicyRef.ID)
+
+	tbl = external.GetTableByName(t, tk, "test", "t")
 	require.NotNil(t, tbl)
 	require.NotNil(t, tbl.Meta().PlacementPolicyRef)
 	require.Equal(t, "x", tbl.Meta().PlacementPolicyRef.Name.L)

docs/design/2020-06-24-placement-rules-in-sql.md

@@ -413,6 +413,26 @@ If a table is imported when `tidb_placement_mode='IGNORE'`, and the placement po
 
 The default value for `tidb_placement_mode` is `STRICT`. The option is an enum, and in future we may add support for a `WARN` mode.
 
+
+#### Survival preference
+
+Some important data may need to store multiple copies across availability zones, so as to have high disaster recovery survivability, such as region-level survivability, `SURVIVAL_PREFERENCES` can provide survivability preference settings.
+
+The following example sets a constraint that the data try to satisfy the Survival Preferences setting:
+
+``` sql
+CREATE PLACEMENT POLICY multiregion
+    follower=4
+    PRIMARY_REGION="region1"
+    SURVIVAL_PREFERENCES="[region, zone]";
+```
+
+For tables with this policy set, the data will first meet the survival goal of cross-region data isolation, and then ensure the survival goal of cross-zone data isolation.
+
+> **Note:**
+>
+> `SURVIVAL_PREFERENCES` is equivalent to `LOCATION_LABELS` in PD. For more information, please refer to [Replica scheduling by topology label](https://docs.pingcap.com/tidb/dev/schedule-replicas-by-topology-labels#schedule-replicas-by-topology-labels).
+
 #### Ambiguous and edge cases
 
 The following two policies are not identical:
@@ -491,6 +511,7 @@ In this case the default rules will apply to placement, and the output from `SHO
 - `FOLLOWER_CONSTRAINTS`
 - `LEARNER_CONSTRAINTS`
 - `PLACEMENT POLICY`
+- `SURVIVAL_PREFERENCE`
 
 For a more complex rule using partitions, consider the following example:
 

domain/infosync/BUILD.bazel

@@ -43,6 +43,7 @@ go_library(
         "@com_github_pingcap_failpoint//:failpoint",
         "@com_github_pingcap_kvproto//pkg/metapb",
         "@com_github_pingcap_kvproto//pkg/resource_manager",
+        "@com_github_pingcap_log//:log",
         "@com_github_tikv_client_go_v2//oracle",
         "@com_github_tikv_pd_client//:client",
         "@io_etcd_go_etcd_client_v3//:client",

domain/infosync/placement_manager.go

@@ -21,9 +21,11 @@ import (
 	"path"
 	"sync"
 
+	"github.com/pingcap/log"
 	"github.com/pingcap/tidb/ddl/placement"
 	"github.com/pingcap/tidb/util/pdapi"
 	clientv3 "go.etcd.io/etcd/client/v3"
+	"go.uber.org/zap"
 )
 
 // PlacementManager manages placement settings
@@ -72,6 +74,7 @@ func (m *PDPlacementManager) PutRuleBundles(ctx context.Context, bundles []*plac
 		return err
 	}
 
+	log.Debug("Put placement rule bundles", zap.String("rules", string(b)))
 	_, err = doRequest(ctx, "PutPlacementRules", m.etcdCli.Endpoints(), path.Join(pdapi.Config, "placement-rule")+"?partial=true", "POST", bytes.NewReader(b))
 	return err
 }

parser/ast/ddl.go

@@ -2025,6 +2025,7 @@ const (
 	PlacementOptionLearnerConstraints
 	PlacementOptionFollowerConstraints
 	PlacementOptionVoterConstraints
+	PlacementOptionSurvivalPreferences
 	PlacementOptionPolicy
 )
 
@@ -2089,6 +2090,10 @@ func (n *PlacementOption) Restore(ctx *format.RestoreCtx) error {
 			ctx.WriteKeyWord("PLACEMENT POLICY ")
 			ctx.WritePlain("= ")
 			ctx.WriteName(n.StrValue)
+		case PlacementOptionSurvivalPreferences:
+			ctx.WriteKeyWord("SURVIVAL_PREFERENCES ")
+			ctx.WritePlain("= ")
+			ctx.WriteString(n.StrValue)
 		default:
 			return errors.Errorf("invalid PlacementOption: %d", n.Tp)
 		}

parser/misc.go

@@ -715,6 +715,7 @@ var tokenMap = map[string]int{
 	"SUBSTRING":                substring,
 	"SUM":                      sum,
 	"SUPER":                    super,
+	"SURVIVAL_PREFERENCES":     survivalPreferences,
 	"SWAPS":                    swaps,
 	"SWITCHES":                 switchesSym,
 	"SYSTEM":                   system,

parser/model/model.go

@@ -1747,6 +1747,7 @@ type PlacementSettings struct {
 	LearnerConstraints  string `json:"learner_constraints"`
 	FollowerConstraints string `json:"follower_constraints"`
 	VoterConstraints    string `json:"voter_constraints"`
+	SurvivalPreferences string `json:"survival_preferences"`
 }
 
 // PolicyInfo is the struct to store the placement policy.