Gabriel39 commented on code in PR #65851:
URL: https://github.com/apache/doris/pull/65851#discussion_r3670362669
##########
fe/fe-core/src/main/java/org/apache/doris/datasource/iceberg/source/IcebergScanNode.java:
##########
@@ -530,30 +594,646 @@ void enableCurrentIcebergScanSemantics() {
params.setIcebergScanSemanticsVersion(ICEBERG_SCAN_SEMANTICS_VERSION);
}
+ /**
+ * Build the schema metadata carrier used by both scanners and
equality-delete readers.
+ *
+ * <p>Batch-mode delete files are planned asynchronously after scan
parameters are sent to BE.
+ * The authenticated manifest preflight therefore supplies the live
equality field IDs before
+ * the schema carrier is serialized. Only historical fields referenced by
those delete files are
+ * added, so an unrelated dropped type cannot make an otherwise supported
scan fail.
+ */
+ @VisibleForTesting
+ List<NestedField> getSchemaFieldsForScan(
+ Schema scanSchema, Set<Integer> equalityDeleteFieldIds) throws
UserException {
+ List<NestedField> fields = new ArrayList<>(scanSchema.columns());
+ if (isSystemTable || equalityDeleteFieldIds.isEmpty()) {
+ return fields;
+ }
+
+ Set<Integer> missingFieldIds = new HashSet<>(equalityDeleteFieldIds);
+
missingFieldIds.removeAll(TypeUtil.indexById(scanSchema.asStruct()).keySet());
+ if (missingFieldIds.isEmpty()) {
+ return fields;
+ }
+
+ List<Schema> schemaHistory = getMetadataSchemaHistory();
+ // Schema IDs may be reused when evolution returns to an earlier
schema, while the metadata
+ // list may also contain schemas committed after a time-travel or
branch target. Follow the
+ // actual scan snapshot's parent chain first so the field definition
active on that lineage
+ // wins. Then use the complete metadata list as a fallback for
schema-only changes and
+ // expired ancestors. A fallback definition may come from a later
rename, so BE resolves an
+ // ID-less equality key through the target mapping first and the
delete file's original key
+ // name second. Initial-default and field identity remain bound to the
stable field ID.
+ Snapshot snapshot = createTableScan().snapshot();
+ while (snapshot != null) {
+ Integer schemaId = snapshot.schemaId();
+ if (schemaId != null) {
+ Schema historicalSchema = icebergTable.schemas().get(schemaId);
+ Preconditions.checkState(historicalSchema != null,
+ "Iceberg snapshot schema %s is absent from table
metadata", schemaId);
+ addHistoricalEqualityFields(fields, missingFieldIds,
historicalSchema);
+ }
+ Long parentId = snapshot.parentId();
+ snapshot = parentId == null ? null :
icebergTable.snapshot(parentId);
+ }
+ for (int index = schemaHistory.size() - 1; index >= 0; index--) {
+ addHistoricalEqualityFields(fields, missingFieldIds,
schemaHistory.get(index));
+ }
+ Preconditions.checkState(missingFieldIds.isEmpty(),
+ "Iceberg equality-delete fields are absent from schema
history: %s",
+ missingFieldIds);
+ return fields;
+ }
+
+ private List<Schema> getMetadataSchemaHistory() {
+ Preconditions.checkState(icebergTable instanceof HasTableOperations,
+ "Iceberg table does not expose metadata schema history: %s",
icebergTable.name());
+ return ((HasTableOperations)
icebergTable).operations().current().schemas();
+ }
+
+ /**
+ * Return only schemas that can describe files visible from the selected
target.
+ *
+ * <p>The query schema is included explicitly because a schema-only commit
does not create a
+ * snapshot. Other schemas are taken from the selected snapshot's parent
lineage and from
+ * cherry-picked source snapshots (including their ancestry), excluding
later main-branch and
+ * unrelated branch schemas from the rolling-upgrade fence. An empty
optional means snapshot
+ * expiration truncated any required lineage, so callers must
conservatively require current
+ * scan semantics.
+ */
+ @VisibleForTesting
+ Optional<List<Schema>> getRequiredFieldSchemaHistory(Schema scanSchema)
throws UserException {
+ List<Schema> schemas = new ArrayList<>();
+ Set<Integer> schemaIds = new HashSet<>();
+ schemas.add(scanSchema);
+ schemaIds.add(scanSchema.schemaId());
+
+ Snapshot selectedSnapshot = createTableScan().snapshot();
+ Deque<Snapshot> snapshots = new ArrayDeque<>();
+ if (selectedSnapshot != null) {
+ snapshots.add(selectedSnapshot);
+ }
+ Set<Long> visitedSnapshotIds = new HashSet<>();
+ while (!snapshots.isEmpty()) {
+ Snapshot snapshot = snapshots.removeFirst();
+ if (!visitedSnapshotIds.add(snapshot.snapshotId())) {
+ continue;
+ }
+ Integer schemaId = snapshot.schemaId();
+ if (schemaId != null && schemaIds.add(schemaId)) {
+ Schema lineageSchema = icebergTable.schemas().get(schemaId);
+ Preconditions.checkState(lineageSchema != null,
+ "Iceberg snapshot schema %s is absent from table
metadata", schemaId);
+ schemas.add(lineageSchema);
+ }
+ Long parentId = snapshot.parentId();
+ if (parentId != null) {
+ Snapshot parent = icebergTable.snapshot(parentId);
+ if (parent == null) {
+ return Optional.empty();
+ }
+ snapshots.addLast(parent);
+ }
+ String sourceSnapshotId =
+
snapshot.summary().get(SnapshotSummary.SOURCE_SNAPSHOT_ID_PROP);
+ if (sourceSnapshotId != null) {
+ Snapshot sourceSnapshot =
+
icebergTable.snapshot(Long.parseLong(sourceSnapshotId));
+ if (sourceSnapshot == null) {
+ return Optional.empty();
+ }
+ snapshots.addLast(sourceSnapshot);
+ }
+ }
+ return Optional.of(schemas);
+ }
+
+ private static void addHistoricalEqualityFields(List<NestedField> fields,
+ Set<Integer> missingFieldIds, Schema historicalSchema) {
+ for (NestedField field :
TypeUtil.indexById(historicalSchema.asStruct()).values()) {
+ if (missingFieldIds.remove(field.fieldId())) {
+ Preconditions.checkState(field.type().isPrimitiveType(),
+ "Iceberg equality-delete field %s must be primitive",
field.fieldId());
+ fields.add(field);
+ }
+ }
+ }
+
+ @VisibleForTesting
+ static boolean requiresRecursiveInitialDefaultMaterialization(
+ Schema scanSchema, List<SlotDescriptor> projectedSlots) {
+ return requiresProjectedIcebergField(scanSchema, projectedSlots,
+ (field, isTopLevel) -> field.initialDefault() != null
+ && (!isTopLevel || field.type().isNestedType()));
+ }
+
+ @VisibleForTesting
+ static boolean requiresMissingRequiredFieldRejection(
+ Schema scanSchema, List<SlotDescriptor> projectedSlots,
+ Optional<List<Schema>> historicalSchemas) {
+ return !historicalSchemas.isPresent()
+ || requiresMissingRequiredFieldRejection(
+ scanSchema, projectedSlots, historicalSchemas.get());
+ }
+
+ @VisibleForTesting
+ static boolean requiresMissingRequiredFieldRejection(
+ Schema scanSchema, List<SlotDescriptor> projectedSlots,
+ List<Schema> historicalSchemas) {
+ Map<Integer, NestedField> fieldById =
TypeUtil.indexById(scanSchema.asStruct());
+ Map<Integer, Integer> parentById =
TypeUtil.indexParents(scanSchema.asStruct());
+ Set<Integer> collectionWrapperFieldIds = new HashSet<>();
+ collectCollectionWrapperFieldIds(scanSchema.asStruct(),
collectionWrapperFieldIds);
+ Set<Integer> potentiallyMissingRequiredFieldIds = new HashSet<>();
+ for (Schema historicalSchema : historicalSchemas) {
+ Map<Integer, NestedField> historicalFieldById =
+ TypeUtil.indexById(historicalSchema.asStruct());
+ for (NestedField field : fieldById.values()) {
+ NestedField historicalField =
historicalFieldById.get(field.fieldId());
+ if (historicalField != null) {
+ if (!collectionWrapperFieldIds.contains(field.fieldId())
+ && field.isRequired() && field.initialDefault() ==
null
+ && historicalField.isOptional()) {
+
potentiallyMissingRequiredFieldIds.add(field.fieldId());
+ }
+ continue;
+ }
+ NestedField highestMissingField = field;
+ Integer parentId = parentById.get(field.fieldId());
+ while (parentId != null &&
!historicalFieldById.containsKey(parentId)) {
+ highestMissingField =
Preconditions.checkNotNull(fieldById.get(parentId),
+ "Iceberg parent field %s is absent from scan
schema", parentId);
+ parentId = parentById.get(parentId);
+ }
+ // If the highest missing ancestor is optional, the old
physical subtree is NULL
+ // and no required descendant is materialized. A non-null
initial default is
+ // already covered by
requiresRecursiveInitialDefaultMaterialization().
+ if
(!collectionWrapperFieldIds.contains(highestMissingField.fieldId())
+ && highestMissingField.isRequired()
+ && highestMissingField.initialDefault() == null) {
+
potentiallyMissingRequiredFieldIds.add(highestMissingField.fieldId());
+ }
+ }
+ }
+ return requiresProjectedIcebergField(scanSchema, projectedSlots,
+ (field, isTopLevel) ->
potentiallyMissingRequiredFieldIds.contains(
+ field.fieldId()));
+ }
+
+ private static void collectCollectionWrapperFieldIds(
+ Type type, Set<Integer> collectionWrapperFieldIds) {
+ switch (type.typeId()) {
+ case STRUCT:
+ for (NestedField field : type.asStructType().fields()) {
+ collectCollectionWrapperFieldIds(field.type(),
collectionWrapperFieldIds);
+ }
+ break;
+ case LIST:
+ Types.ListType listType = (Types.ListType) type;
+ collectionWrapperFieldIds.add(listType.elementId());
+ collectCollectionWrapperFieldIds(
+ listType.elementType(), collectionWrapperFieldIds);
+ break;
+ case MAP:
+ Types.MapType mapType = (Types.MapType) type;
+ collectionWrapperFieldIds.add(mapType.keyId());
+ collectionWrapperFieldIds.add(mapType.valueId());
+ collectCollectionWrapperFieldIds(mapType.keyType(),
collectionWrapperFieldIds);
+ collectCollectionWrapperFieldIds(mapType.valueType(),
collectionWrapperFieldIds);
+ break;
+ default:
+ break;
+ }
+ }
+
+ private static boolean requiresProjectedIcebergField(
+ Schema scanSchema, List<SlotDescriptor> projectedSlots,
+ ProjectedFieldRequirement requirement) {
+ Map<Integer, NestedField> fieldById =
TypeUtil.indexById(scanSchema.asStruct());
+ Set<Integer> topLevelFieldIds = new HashSet<>();
+ for (NestedField field : scanSchema.columns()) {
+ topLevelFieldIds.add(field.fieldId());
+ }
+ for (SlotDescriptor slot : projectedSlots) {
+ Column column = slot.getColumn();
+ List<ColumnAccessPath> accessPaths = slot.getAllAccessPaths();
+ if (accessPaths != null && !accessPaths.isEmpty()) {
+ for (ColumnAccessPath accessPath : accessPaths) {
+ List<String> path = accessPath.getPath();
+ Preconditions.checkState(!path.isEmpty(),
+ "Iceberg column access path must not be empty");
+
Preconditions.checkState(matchesAccessPathComponent(column, path.get(0)),
+ "Iceberg access path root %s does not match column
%s", path.get(0),
+ column.getName());
+ if (requiresProjectedIcebergField(
+ column, path, 1, fieldById,
+ topLevelFieldIds.contains(column.getUniqueId()),
requirement)) {
+ return true;
+ }
+ }
+ } else if (requiresProjectedIcebergField(
+ column, slot.getType(), fieldById,
+ topLevelFieldIds.contains(column.getUniqueId()),
requirement)) {
+ return true;
+ }
+ }
+ return false;
+ }
+
+ private static boolean requiresProjectedIcebergField(
+ Column column, org.apache.doris.catalog.Type projectedType,
+ Map<Integer, NestedField> fieldById, boolean isTopLevel,
+ ProjectedFieldRequirement requirement) {
+ if (requiresIcebergField(column, fieldById, isTopLevel, requirement)) {
+ return true;
+ }
+ if (column.getChildren() == null) {
+ return false;
+ }
+ if (projectedType.isStructType()) {
+ for (StructField projectedField : ((StructType)
projectedType).getFields()) {
+ Column child = findChildByName(column,
projectedField.getName());
+ Preconditions.checkState(child != null,
+ "Projected Iceberg child %s is absent from column %s",
+ projectedField.getName(), column.getName());
+ if (requiresProjectedIcebergField(
+ child, projectedField.getType(), fieldById, false,
requirement)) {
+ return true;
+ }
+ }
+ } else if (projectedType.isArrayType()) {
+ Preconditions.checkState(column.getChildren().size() == 1,
+ "Iceberg array column %s must have one child",
column.getName());
+ if (requiresProjectedIcebergField(
+ column.getChildren().get(0), ((ArrayType)
projectedType).getItemType(),
+ fieldById, false, requirement)) {
+ return true;
+ }
+ } else if (projectedType.isMapType()) {
+ Preconditions.checkState(column.getChildren().size() == 2,
+ "Iceberg map column %s must have two children",
column.getName());
+ MapType mapType = (MapType) projectedType;
+ if (requiresProjectedIcebergField(
+ column.getChildren().get(0), mapType.getKeyType(),
fieldById, false,
+ requirement)
+ || requiresProjectedIcebergField(
+ column.getChildren().get(1),
mapType.getValueType(), fieldById, false,
+ requirement)) {
+ return true;
+ }
+ }
+ return false;
+ }
+
+ private static boolean requiresProjectedIcebergField(
+ Column column, List<String> path, int pathIndex,
+ Map<Integer, NestedField> fieldById, boolean isTopLevel,
+ ProjectedFieldRequirement requirement) {
+ if (requiresIcebergField(column, fieldById, isTopLevel, requirement)) {
+ return true;
+ }
+ if (pathIndex == path.size()) {
+ return requiresProjectedIcebergField(column, fieldById,
requirement);
+ }
+
+ String component = path.get(pathIndex);
+ if (AccessPathInfo.ACCESS_NULL.equals(component)
+ || AccessPathInfo.ACCESS_OFFSET.equals(component)) {
+ return false;
+ }
+ Preconditions.checkState(column.getChildren() != null,
+ "Iceberg access path continues below primitive column %s",
column.getName());
+
+ if (AccessPathInfo.ACCESS_ALL.equals(component)) {
+ if (column.getType().isArrayType()) {
+ Preconditions.checkState(column.getChildren().size() == 1,
+ "Iceberg array column %s must have one child",
column.getName());
+ return requiresProjectedIcebergField(
+ column.getChildren().get(0), path, pathIndex + 1,
fieldById, false,
+ requirement);
+ }
+ Preconditions.checkState(column.getType().isMapType(),
+ "Unexpected Iceberg access-all path below column %s",
column.getName());
+ Preconditions.checkState(column.getChildren().size() == 2,
+ "Iceberg map column %s must have two children",
column.getName());
+ Column key = column.getChildren().get(0);
+ // element_at(map, key) reads the complete key subtree, while any
path after '*'
+ // describes only the selected value subtree.
+ if (requiresIcebergField(key, fieldById, false, requirement)
+ || requiresProjectedIcebergField(key, fieldById,
requirement)) {
+ return true;
+ }
+ return requiresProjectedIcebergField(
+ column.getChildren().get(1), path, pathIndex + 1,
fieldById, false,
+ requirement);
+ }
+ if (column.getType().isMapType()) {
+ Preconditions.checkState(column.getChildren().size() == 2,
+ "Iceberg map column %s must have two children",
column.getName());
+ int childIndex;
+ if (AccessPathInfo.ACCESS_MAP_KEYS.equals(component)) {
+ childIndex = 0;
+ } else {
+
Preconditions.checkState(AccessPathInfo.ACCESS_MAP_VALUES.equals(component),
+ "Unexpected Iceberg map access path component %s",
component);
+ childIndex = 1;
+ }
+ return requiresProjectedIcebergField(
+ column.getChildren().get(childIndex), path, pathIndex + 1,
fieldById, false,
+ requirement);
+ }
+
+ Column child = findAccessPathChild(column, component);
+ Preconditions.checkState(child != null,
+ "Iceberg access path child %s is absent from column %s",
component,
+ column.getName());
+ return requiresProjectedIcebergField(
+ child, path, pathIndex + 1, fieldById, false, requirement);
+ }
+
+ private static boolean requiresProjectedIcebergField(
+ Column column, Map<Integer, NestedField> fieldById,
+ ProjectedFieldRequirement requirement) {
+ if (column.getChildren() == null) {
+ return false;
+ }
+ for (Column child : column.getChildren()) {
+ if (requiresIcebergField(child, fieldById, false, requirement)
+ || requiresProjectedIcebergField(child, fieldById,
requirement)) {
+ return true;
+ }
+ }
+ return false;
+ }
+
+ private static boolean requiresIcebergField(
+ Column column, Map<Integer, NestedField> fieldById, boolean
isTopLevel,
+ ProjectedFieldRequirement requirement) {
+ NestedField field = fieldById.get(column.getUniqueId());
+ return field != null && requirement.requires(field, isTopLevel);
+ }
+
+ private interface ProjectedFieldRequirement {
+ boolean requires(NestedField field, boolean isTopLevel);
+ }
+
+ /**
+ * Detect a reused name that current BEs resolve before an older sibling's
historical alias.
+ *
+ * <p>A smooth-upgrade source BE recognizes only the original semantics
marker and performs one
+ * ordered name/alias pass. If a sibling retains another sibling's current
name as an alias, the
+ * two BE generations can bind the same projected path to different field
IDs and types.
+ */
+ @VisibleForTesting
+ static boolean hasCurrentNameAliasCollision(
+ Schema schema, Optional<Map<Integer, List<String>>> nameMapping) {
+ return nameMapping.isPresent()
+ && hasCurrentNameAliasCollision(schema.asStruct(),
nameMapping.get());
+ }
+
+ @VisibleForTesting
+ static void checkNameMappingBackendCompatibility(
+ Schema schema,
+ Optional<Map<Integer, List<String>>> nameMapping,
+ Iterable<Backend> backends) throws UserException {
+ if (hasCurrentNameAliasCollision(schema, nameMapping)) {
+ checkCurrentIcebergScanSemanticsBackendCompatibility(backends);
+ }
+ }
+
+ private static boolean hasCurrentNameAliasCollision(
+ Type type, Map<Integer, List<String>> nameMapping) {
+ switch (type.typeId()) {
+ case STRUCT:
+ List<NestedField> fields = type.asStructType().fields();
+ for (NestedField field : fields) {
+ List<String> aliases =
+ nameMapping.getOrDefault(field.fieldId(),
Collections.emptyList());
+ for (String alias : aliases) {
+ for (NestedField sibling : fields) {
+ if (sibling.fieldId() != field.fieldId()
+ && sibling.name().equalsIgnoreCase(alias))
{
+ return true;
+ }
+ }
+ }
+ if (hasCurrentNameAliasCollision(field.type(),
nameMapping)) {
+ return true;
+ }
+ }
+ return false;
+ case LIST:
+ return hasCurrentNameAliasCollision(
+ type.asListType().elementType(), nameMapping);
+ case MAP:
+ return
hasCurrentNameAliasCollision(type.asMapType().keyType(), nameMapping)
+ || hasCurrentNameAliasCollision(
+ type.asMapType().valueType(), nameMapping);
+ default:
+ return false;
+ }
+ }
+
+ private static boolean matchesAccessPathComponent(Column column, String
component) {
+ return Integer.toString(column.getUniqueId()).equals(component)
+ || column.getName().equalsIgnoreCase(component);
+ }
+
+ private static Column findAccessPathChild(Column column, String component)
{
+ for (Column child : column.getChildren()) {
+ if (matchesAccessPathComponent(child, component)) {
+ return child;
+ }
+ }
+ return null;
+ }
+
+ private static Column findChildByName(Column column, String childName) {
+ for (Column child : column.getChildren()) {
+ if (child.getName().equalsIgnoreCase(childName)) {
+ return child;
+ }
+ }
+ return null;
+ }
+
+ @VisibleForTesting
+ Set<Integer> getEqualityDeleteFieldIdsForScan() throws UserException {
+ TableScan scan = createTableScan();
+ if (scan.snapshot() == null) {
+ return Collections.emptySet();
+ }
+ try {
+ return preExecutionAuthenticator.execute(
+ () -> loadEqualityDeleteFieldIds(scan));
+ } catch (Exception e) {
+ Optional<NotSupportedException> opt =
checkNotSupportedException(e);
+ if (opt.isPresent()) {
+ throw opt.get();
+ }
+ throw new UserException(ExceptionUtils.getRootCauseMessage(e), e);
+ }
+ }
+
+ /**
+ * Skip exhaustive delete-file planning when the exact snapshot summary
already proves that
+ * metadata-only COUNT(*) is safe. A usable count requires the summary's
equality-delete total
+ * to be zero, so no equality field IDs can affect this scan.
+ */
+ @VisibleForTesting
+ Set<Integer> getEqualityDeleteFieldIdsForPlanning() throws UserException {
+ if (prepareTableLevelSnapshotCount()) {
+ return Collections.emptySet();
+ }
+ return getEqualityDeleteFieldIdsForScan();
+ }
+
+ @VisibleForTesting
+ Set<Integer> loadEqualityDeleteFieldIds(TableScan scan) {
+ ConnectContext context = ConnectContext.get();
+ Preconditions.checkNotNull(context);
+ Preconditions.checkNotNull(context.getStatementContext());
+ List<FileScanTask> rewriteTasks =
+ context.getStatementContext().getIcebergRewriteFileScanTasks();
+ if (rewriteTasks != null) {
+ return collectEqualityDeleteFieldIdsFromTasks(rewriteTasks);
+ }
+ // This is a hidden metadata preflight. Using Doris's manifest cache
here would populate
+ // the cache before the real scan is planned, making the scan's cache
profile report hits
+ // for work performed only by this preflight.
+ try (CloseableIterable<FileScanTask> tasks = scan.planFiles()) {
Review Comment:
[P1] Avoid exhaustively planning files just to build the schema carrier
Except for the metadata-only COUNT(*) shortcut, every scan reaches this
preflight and fully consumes `scan.planFiles()` to collect equality-delete IDs.
A normal non-batch scan then calls `planFiles()` again from `splitFiles()`, so
manifest/file planning is performed twice. In batch mode this synchronous FE
preflight defeats the asynchronous planning contract described above:
large-table planning is forced back onto the FE critical path even when the
snapshot has no applicable equality deletes.
Please share the exact planned tasks with the real split path in non-batch
mode, and avoid exhausting `FileScanTask`s before dispatch in batch mode (for
example, by integrating equality-key discovery with the shared/asynchronous
planning path or using a metadata-level carrier that does not require full file
planning). Add coverage that counts `planFiles()` calls for ordinary and batch
scans, not only the metadata-only COUNT(*) case.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
