infoschema: fix issue of information schema cache miss cause by schema version gap (pingcap#53445)

close pingcap#53428

Signed-off-by: crazycs520 <[email protected]>

Author: crazycs520

domain/domain.go

@@ -417,6 +417,8 @@ func (do *Domain) tryLoadSchemaDiffs(m *meta.Meta, usedVersion, newVersion int64
 		if diff == nil {
 			// Empty diff means the txn of generating schema version is committed, but the txn of `runDDLJob` is not or fail.
 			// It is safe to skip the empty diff because the infoschema is new enough and consistent.
+			logutil.BgLogger().Info("diff load InfoSchema get empty schema diff", zap.Int64("version", usedVersion))
+			do.infoCache.InsertEmptySchemaVersion(usedVersion)
 			continue
 		}
 		diffs = append(diffs, diff)

infoschema/cache.go

@@ -30,6 +30,9 @@ type InfoCache struct {
 	mu sync.RWMutex
 	// cache is sorted by both SchemaVersion and timestamp in descending order, assume they have same order
 	cache []schemaAndTimestamp
+
+	// emptySchemaVersions stores schema version which has no schema_diff.
+	emptySchemaVersions map[int64]struct{}
 }
 
 type schemaAndTimestamp struct {
@@ -40,10 +43,18 @@ type schemaAndTimestamp struct {
 // NewCache creates a new InfoCache.
 func NewCache(capacity int) *InfoCache {
 	return &InfoCache{
-		cache: make([]schemaAndTimestamp, 0, capacity),
+		cache:               make([]schemaAndTimestamp, 0, capacity),
+		emptySchemaVersions: make(map[int64]struct{}),
 	}
 }
 
+// Size returns the size of the cache, export for test.
+func (h *InfoCache) Size() int {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return len(h.cache)
+}
+
 // Reset resets the cache.
 func (h *InfoCache) Reset(capacity int) {
 	h.mu.Lock()
@@ -63,6 +74,11 @@ func (h *InfoCache) GetLatest() InfoSchema {
 	return nil
 }
 
+// GetEmptySchemaVersions returns emptySchemaVersions, exports for testing.
+func (h *InfoCache) GetEmptySchemaVersions() map[int64]struct{} {
+	return h.emptySchemaVersions
+}
+
 func (h *InfoCache) getSchemaByTimestampNoLock(ts uint64) (InfoSchema, bool) {
 	logutil.BgLogger().Debug("SCHEMA CACHE get schema", zap.Uint64("timestamp", ts))
 	// search one by one instead of binary search, because the timestamp of a schema could be 0
@@ -80,11 +96,32 @@ func (h *InfoCache) getSchemaByTimestampNoLock(ts uint64) (InfoSchema, bool) {
 			// the first element is the latest schema, so we can return it directly.
 			return is.infoschema, true
 		}
-		if h.cache[i-1].infoschema.SchemaMetaVersion() == is.infoschema.SchemaMetaVersion()+1 && uint64(h.cache[i-1].timestamp) > ts {
-			// This first condition is to make sure the schema version is continuous. If last(cache[i-1]) schema-version is 10,
-			// but current(cache[i]) schema-version is not 9, then current schema is not suitable for ts.
-			// The second condition is to make sure the cache[i-1].timestamp > ts >= cache[i].timestamp, then the current schema is suitable for ts.
-			return is.infoschema, true
+
+		if uint64(h.cache[i-1].timestamp) > ts {
+			// The first condition is to make sure the cache[i-1].timestamp > ts >= cache[i].timestamp, then the current schema is suitable for ts.
+			lastVersion := h.cache[i-1].infoschema.SchemaMetaVersion()
+			currentVersion := is.infoschema.SchemaMetaVersion()
+			if lastVersion == currentVersion+1 {
+				// This condition is to make sure the schema version is continuous. If last(cache[i-1]) schema-version is 10,
+				// but current(cache[i]) schema-version is not 9, then current schema may not suitable for ts.
+				return is.infoschema, true
+			}
+			if lastVersion > currentVersion {
+				found := true
+				for ver := currentVersion + 1; ver < lastVersion; ver++ {
+					_, ok := h.emptySchemaVersions[ver]
+					if !ok {
+						found = false
+						break
+					}
+				}
+				if found {
+					// This condition is to make sure the schema version is continuous. If last(cache[i-1]) schema-version is 10, and
+					// current(cache[i]) schema-version is 8, then there is a gap exist, and if all the gap version can be found in cache.emptySchemaVersions
+					// which means those gap versions don't have schema info, then current schema is also suitable for ts.
+					return is.infoschema, true
+				}
+			}
 		}
 		// current schema is not suitable for ts, then break the loop to avoid the unnecessary search.
 		break
@@ -194,3 +231,25 @@ func (h *InfoCache) Insert(is InfoSchema, schemaTS uint64) bool {
 
 	return true
 }
+
+// InsertEmptySchemaVersion inserts empty schema version into a map. If exceeded the cache capacity, remove the oldest version.
+func (h *InfoCache) InsertEmptySchemaVersion(version int64) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+
+	h.emptySchemaVersions[version] = struct{}{}
+	if len(h.emptySchemaVersions) > cap(h.cache) {
+		// remove oldest version.
+		versions := make([]int64, 0, len(h.emptySchemaVersions))
+		for ver := range h.emptySchemaVersions {
+			versions = append(versions, ver)
+		}
+		sort.Slice(versions, func(i, j int) bool { return versions[i] < versions[j] })
+		for _, ver := range versions {
+			delete(h.emptySchemaVersions, ver)
+			if len(h.emptySchemaVersions) <= cap(h.cache) {
+				break
+			}
+		}
+	}
+}

infoschema/cache_test.go

@@ -244,4 +244,51 @@ func TestCacheWithSchemaTsZero(t *testing.T) {
 	checkFn(1, 84, false)
 	checkFn(85, 100, true)
 	require.Equal(t, 16, ic.Size())
+
+	// Test cache with schema version hole, which is cause by schema version doesn't has related schema-diff.
+	ic = infoschema.NewCache(16)
+	require.NotNil(t, ic)
+	for i := 1; i <= 8; i++ {
+		ic.Insert(infoschema.MockInfoSchemaWithSchemaVer(nil, int64(i)), uint64(i))
+	}
+	checkFn(1, 10, true)
+	// mock for schema version hole, schema-version 9 is missing.
+	ic.Insert(infoschema.MockInfoSchemaWithSchemaVer(nil, 10), 10)
+	checkFn(1, 7, true)
+	// without empty schema version map, get snapshot by ts 8, 9 will both failed.
+	checkFn(8, 9, false)
+	checkFn(10, 10, true)
+	// add empty schema version 9.
+	ic.InsertEmptySchemaVersion(9)
+	// after set empty schema version, get snapshot by ts 8, 9 will both success.
+	checkFn(1, 8, true)
+	checkFn(10, 10, true)
+	is := ic.GetBySnapshotTS(uint64(9))
+	require.NotNil(t, is)
+	// since schema version 9 is empty, so get by ts 9 will get schema which version is 8.
+	require.Equal(t, int64(8), is.SchemaMetaVersion())
+}
+
+func TestCacheEmptySchemaVersion(t *testing.T) {
+	ic := infoschema.NewCache(16)
+	require.NotNil(t, ic)
+	require.Equal(t, 0, len(ic.GetEmptySchemaVersions()))
+	for i := 0; i < 16; i++ {
+		ic.InsertEmptySchemaVersion(int64(i))
+	}
+	emptyVersions := ic.GetEmptySchemaVersions()
+	require.Equal(t, 16, len(emptyVersions))
+	for i := 0; i < 16; i++ {
+		_, ok := emptyVersions[int64(i)]
+		require.True(t, ok)
+	}
+	for i := 16; i < 20; i++ {
+		ic.InsertEmptySchemaVersion(int64(i))
+	}
+	emptyVersions = ic.GetEmptySchemaVersions()
+	require.Equal(t, 16, len(emptyVersions))
+	for i := 4; i < 20; i++ {
+		_, ok := emptyVersions[int64(i)]
+		require.True(t, ok)
+	}
 }

statistics/handle/handletest/globalstats/BUILD.bazel

@@ -9,7 +9,7 @@ go_test(
         "main_test.go",
     ],
     flaky = True,
-    shard_count = 7,
+    shard_count = 8,
     deps = [
         "//domain",
         "//testkit",