xiangfu0 commented on code in PR #19998: URL: https://github.com/apache/pinot/pull/19998#discussion_r4366769930
########## pinot-segment-spi/src/main/java/org/apache/pinot/segment/spi/index/metadata/MappedSegmentMetadataStore.java: ########## @@ -0,0 +1,904 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package org.apache.pinot.segment.spi.index.metadata; + +import com.fasterxml.jackson.databind.node.ObjectNode; +import com.google.common.collect.Interner; +import com.google.common.collect.Interners; +import java.io.ByteArrayInputStream; +import java.io.ByteArrayOutputStream; +import java.io.Closeable; +import java.io.DataInputStream; +import java.io.DataOutputStream; +import java.io.File; +import java.io.IOException; +import java.math.BigDecimal; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; +import java.nio.charset.StandardCharsets; +import java.nio.file.AtomicMoveNotSupportedException; +import java.nio.file.Files; +import java.nio.file.Path; +import java.nio.file.StandardCopyOption; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Comparator; +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import java.util.Set; +import java.util.stream.Collectors; +import java.util.zip.CRC32C; +import javax.annotation.Nullable; +import org.apache.commons.configuration2.Configuration; +import org.apache.commons.configuration2.PropertiesConfiguration; +import org.apache.commons.configuration2.ex.ConfigurationException; +import org.apache.pinot.segment.spi.ColumnMetadata; +import org.apache.pinot.segment.spi.V1Constants; +import org.apache.pinot.segment.spi.creator.SegmentVersion; +import org.apache.pinot.segment.spi.index.IndexService; +import org.apache.pinot.segment.spi.index.IndexType; +import org.apache.pinot.segment.spi.memory.PinotDataBuffer; +import org.apache.pinot.segment.spi.partition.PartitionFunction; +import org.apache.pinot.segment.spi.store.ColumnIndexUtils; +import org.apache.pinot.segment.spi.store.SegmentDirectoryPaths; +import org.apache.pinot.segment.spi.utils.SegmentMetadataUtils; +import org.apache.pinot.spi.config.table.FieldConfig.EncodingType; +import org.apache.pinot.spi.data.ComplexFieldSpec; +import org.apache.pinot.spi.data.FieldSpec; +import org.apache.pinot.spi.env.CommonsConfigurationUtils; +import org.apache.pinot.spi.utils.ByteArray; +import org.apache.pinot.spi.utils.ColumnNameInterner; +import org.apache.pinot.spi.utils.JsonUtils; + +/// Versioned, file-backed metadata for a sealed segment. The source `metadata.properties` remains the canonical +/// representation for old readers and preprocessing. +/// +/// The big-endian file starts with a 64-byte header containing the magic, format version, total length, source +/// `metadata.properties` and `index_map` length/CRC fingerprints, payload CRC, segment-property blob reference, +/// column count, and record-table bounds. Sorted columns occupy fixed 180-byte records, followed by variable blobs; +/// each blob reference is an absolute 32-bit offset and length. Numeric column statistics are read directly from the +/// mapping. The column layout retains names and reusable [FieldSpec] references on heap. +/// +/// A stale or corrupt sidecar is rebuilt from the canonical source and replaced locally, using atomic rename where +/// supported. Change `VERSION` whenever the encoding or the values derived from canonical metadata change; old +/// versions are rebuilt rather than migrated. +final class MappedSegmentMetadataStore implements Closeable { + static final String FILE_NAME = "segment_metadata.bin"; + private static final String TEMP_FILE_PREFIX = ".segment_metadata."; + private static final String TEMP_FILE_SUFFIX = ".tmp"; + + static boolean isSidecarFileName(String fileName) { + return FILE_NAME.equals(fileName) + || (fileName.startsWith(TEMP_FILE_PREFIX) && fileName.endsWith(TEMP_FILE_SUFFIX)); + } + + private static final int MAGIC = 0x504d4431; // PMD1 + private static final int VERSION = 4; + private static final int HEADER_SIZE = 64; + private static final int RECORD_SIZE = 180; + + static final int NAME = 0; + static final int FIELD_SPEC = 8; + static final int TOTAL_DOCS = 16; + static final int CARDINALITY = 20; + static final int FLAGS = 24; + static final int ENCODING = 28; + static final int SHORTEST = 32; + static final int LONGEST = 36; + static final int TOTAL_ENTRIES = 40; + static final int MAX_MV = 44; + static final int MAX_ROW = 48; + static final int BITS = 52; + static final int MIN_TAG = 56; + static final int MIN_RAW = 60; + static final int MIN_BLOB = 68; + static final int MAX_TAG = 76; + static final int MAX_RAW = 80; + static final int MAX_BLOB = 88; + static final int PARTITION_FUNCTION = 96; + static final int PARTITIONS = 104; + static final int PARENT = 112; + static final int SPARSE_KEYS = 120; + static final int SPARSE_MULTI_VALUE_KEYS = 128; + static final int SPARSE_KEY_TYPES = 136; + static final int RAW_UNCOMPRESSED_SIZE = 144; + static final int DICTIONARY_UNCOMPRESSED_SIZE = 152; + static final int COMPRESSION_TYPE = 160; + static final int INDEXES = 168; + static final int INDEX_COUNT = 176; + + static final int HAS_DICTIONARY = 1; + static final int SORTED = 1 << 1; + static final int NON_NULL = 1 << 2; + static final int INVALID_MIN_MAX = 1 << 3; + static final int ASCII = 1 << 4; + static final int AUTO_GENERATED = 1 << 5; + + private static final int NO_VALUE = 0; + private static final int INT_VALUE = 1; + private static final int LONG_VALUE = 2; + private static final int FLOAT_VALUE = 3; + private static final int DOUBLE_VALUE = 4; + private static final int DECIMAL_VALUE = 5; + private static final int STRING_VALUE = 6; + private static final int BYTES_VALUE = 7; + + private static final Interner<FieldSpec> FIELD_SPECS = Interners.newWeakInterner(); + private static final Interner<ColumnLayout> COLUMN_LAYOUTS = Interners.newWeakInterner(); + + private final PinotDataBuffer _buffer; + private final ColumnLayout _columns; + private final int _tableOffset; + private final int _segmentConfigOffset; + private final int _segmentConfigLength; + + private MappedSegmentMetadataStore(PinotDataBuffer buffer, ColumnLayout columns, int tableOffset, + int segmentConfigOffset, int segmentConfigLength) { + _buffer = buffer; + _columns = columns; + _tableOffset = tableOffset; + _segmentConfigOffset = segmentConfigOffset; + _segmentConfigLength = segmentConfigLength; + } + + static MappedSegmentMetadataStore open(File indexDir) + throws IOException, ConfigurationException { + File source = SegmentDirectoryPaths.findMetadataFile(indexDir); + if (source == null) { + throw new IOException("Missing metadata.properties in " + indexDir); + } + File sidecar = new File(indexDir, FILE_NAME); + boolean rebuilt = false; + for (int attempt = 0; attempt < 3; attempt++) { + Fingerprint propertiesFingerprint = Fingerprint.of(source); + Fingerprint indexMapFingerprint = Fingerprint.of(findIndexMap(indexDir)); + // The validated heap copy is transient; the store retains only the mapping and shared column layout. + byte[] bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + writeAtomically(sidecar, build(indexDir, propertiesFingerprint, indexMapFingerprint)); + rebuilt = true; + bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + continue; + } + } + PinotDataBuffer buffer = PinotDataBuffer.mapFile(sidecar, true, 0, bytes.length, ByteOrder.BIG_ENDIAN, + "segment metadata"); + boolean keepBuffer = false; + try { + // Another loader may atomically replace the sidecar between validation and mapFile(). Check the exact mapped + // bytes and the canonical inputs before publishing a view over this mapping. The header alone cannot prove + // that the mapped payload is the same bytes we validated. + if (!matchesMapped(buffer, bytes) || !propertiesFingerprint.sameAs(Fingerprint.of(source)) + || !indexMapFingerprint.sameAs(Fingerprint.of(findIndexMap(indexDir)))) { + continue; + } + ByteBuffer header = ByteBuffer.wrap(bytes).order(ByteOrder.BIG_ENDIAN); + int count = header.getInt(52); + int tableOffset = header.getInt(56); + String[] names = new String[count]; + FieldSpec[] fieldSpecs = new FieldSpec[count]; + for (int i = 0; i < count; i++) { + int record = tableOffset + i * RECORD_SIZE; + names[i] = ColumnNameInterner.intern(readString(bytes, header.getInt(record + NAME), + header.getInt(record + NAME + 4))); + FieldSpec parsed = JsonUtils.stringToObject(readString(bytes, header.getInt(record + FIELD_SPEC), + header.getInt(record + FIELD_SPEC + 4)), FieldSpec.class); + fieldSpecs[i] = parsed instanceof ComplexFieldSpec ? parsed : FIELD_SPECS.intern(parsed); + } + ColumnLayout columns = COLUMN_LAYOUTS.intern(new ColumnLayout(names, fieldSpecs)); + MappedSegmentMetadataStore store = new MappedSegmentMetadataStore(buffer, columns, tableOffset, + header.getInt(44), header.getInt(48)); + keepBuffer = true; + return store; + } catch (IOException | RuntimeException e) { + if (!rebuilt) { + // A checksum-valid older sidecar can still contain FieldSpec JSON this reader cannot parse. Rebuild it + // once from canonical metadata instead of falling back to eager on every load. + writeAtomically(sidecar, build(indexDir, propertiesFingerprint, indexMapFingerprint)); + rebuilt = true; + continue; + } + throw new IOException("Invalid field specs in mapped segment metadata: " + sidecar, e); + } finally { + if (!keepBuffer) { + buffer.close(); + } + } + } + throw new IOException("Metadata changed while mapping segment: " + indexDir); + } + + private static boolean matchesMapped(PinotDataBuffer buffer, byte[] bytes) { + if (buffer.size() != bytes.length) { + return false; + } + byte[] block = new byte[Math.min(8192, bytes.length)]; + for (int offset = 0; offset < bytes.length; offset += block.length) { + int length = Math.min(block.length, bytes.length - offset); + buffer.copyTo(offset, block, 0, length); + for (int i = 0; i < length; i++) { + if (block[i] != bytes[offset + i]) { + return false; + } + } + } + return true; + } + + int size() { + return _columns._names.length; + } + + String name(int ordinal) { + return _columns._names[ordinal]; + } + + int ordinal(String column) { + return Arrays.binarySearch(_columns._names, column); + } + + int getInt(int ordinal, int offset) { + return _buffer.getInt((long) _tableOffset + (long) ordinal * RECORD_SIZE + offset); + } + + long getLong(int ordinal, int offset) { + return _buffer.getLong((long) _tableOffset + (long) ordinal * RECORD_SIZE + offset); + } + + boolean flag(int ordinal, int flag) { + return (getInt(ordinal, FLAGS) & flag) != 0; + } + + @Nullable + String string(int ordinal, int field) { + byte[] bytes = blob(ordinal, field); + return bytes != null ? new String(bytes, StandardCharsets.UTF_8) : null; + } + + @Nullable + byte[] blob(int ordinal, int field) { + int offset = getInt(ordinal, field); + if (offset < 0) { + return null; + } + int length = getInt(ordinal, field + 4); + byte[] bytes = new byte[length]; + _buffer.copyTo(offset, bytes); + return bytes; + } + + FieldSpec fieldSpec(int ordinal) { + return _columns._fieldSpecs[ordinal]; + } + + @Nullable + Comparable<?> value(int ordinal, boolean minimum) { + int tag = getInt(ordinal, minimum ? MIN_TAG : MAX_TAG); + long raw = getLong(ordinal, minimum ? MIN_RAW : MAX_RAW); + int blobField = minimum ? MIN_BLOB : MAX_BLOB; + switch (tag) { + case NO_VALUE: + return null; + case INT_VALUE: + return (int) raw; + case LONG_VALUE: + return raw; + case FLOAT_VALUE: + return Float.intBitsToFloat((int) raw); + case DOUBLE_VALUE: + return Double.longBitsToDouble(raw); + case DECIMAL_VALUE: + return new BigDecimal(string(ordinal, blobField)); + case STRING_VALUE: + return string(ordinal, blobField); + case BYTES_VALUE: + return new ByteArray(blob(ordinal, blobField)); + default: + throw new IllegalStateException("Invalid value tag: " + tag); + } + } + + PropertiesConfiguration segmentConfiguration() + throws IOException, ConfigurationException { + byte[] bytes = new byte[_segmentConfigLength]; + _buffer.copyTo(_segmentConfigOffset, bytes); + // Values were already split by the source parser. A second delimiter pass would change escaped commas. + PropertiesConfiguration config = CommonsConfigurationUtils.fromInputStream(null, false, null); + try (DataInputStream input = new DataInputStream(new ByteArrayInputStream(bytes))) { + int count = input.readInt(); + for (int i = 0; i < count; i++) { + String key = readString(input); + int valueCount = input.readInt(); + if (valueCount == 1) { + config.setProperty(key, readString(input)); + } else { + List<String> values = new ArrayList<>(valueCount); + for (int j = 0; j < valueCount; j++) { + values.add(readString(input)); + } + config.setProperty(key, values); + } + } + } + return config; + } + + @Override + public void close() + throws IOException { + _buffer.close(); + } + + private static byte[] build(File indexDir, Fingerprint propertiesFingerprint, Fingerprint indexMapFingerprint) + throws IOException, ConfigurationException { + SegmentMetadataImpl source = new SegmentMetadataImpl(indexDir); + PropertiesConfiguration config = SegmentMetadataUtils.getPropertiesConfiguration(indexDir); + List<String> columns = new ArrayList<>(source.getColumnMetadataMap().keySet()); + columns.sort(Comparator.naturalOrder()); + if ((long) columns.size() * RECORD_SIZE + HEADER_SIZE > Integer.MAX_VALUE) { + throw new IOException("Too many columns for mapped metadata: " + columns.size()); + } + int blobBase = HEADER_SIZE + columns.size() * RECORD_SIZE; + BlobWriter blobs = new BlobWriter(blobBase); + BlobRef segmentConfig = blobs.add(encodeSegmentConfiguration(config)); + // EmptyColumnMetadata does not load index sizes even if a stale index_map is still present. + Map<String, List<IndexSize>> indexSizes = source.getTotalDocs() > 0 + && source.getVersion() == SegmentVersion.v3 ? readIndexSizes(indexDir) : Map.of(); + byte[] records = new byte[columns.size() * RECORD_SIZE]; + ByteBuffer recordBuffer = ByteBuffer.wrap(records).order(ByteOrder.BIG_ENDIAN); + for (int ordinal = 0; ordinal < columns.size(); ordinal++) { + String column = columns.get(ordinal); + ColumnMetadata metadata = source.getColumnMetadataFor(column); + int offset = ordinal * RECORD_SIZE; + putRef(recordBuffer, offset + NAME, blobs.add(column)); + putRef(recordBuffer, offset + FIELD_SPEC, blobs.add(fieldSpecJson(metadata.getFieldSpec()).toString())); + recordBuffer.putInt(offset + TOTAL_DOCS, metadata.getTotalDocs()); + recordBuffer.putInt(offset + CARDINALITY, metadata.getCardinality()); + int flags = (metadata.hasDictionary() ? HAS_DICTIONARY : 0) | (metadata.isSorted() ? SORTED : 0) + | (metadata.isNonNull() ? NON_NULL : 0) | (metadata.isMinMaxValueInvalid() ? INVALID_MIN_MAX : 0) + | (metadata.isAscii() ? ASCII : 0) | (metadata.isAutoGenerated() ? AUTO_GENERATED : 0); + recordBuffer.putInt(offset + FLAGS, flags); + EncodingType encoding = metadata.getForwardIndexEncoding(); + recordBuffer.putInt(offset + ENCODING, encoding == null ? 0 : encoding == EncodingType.RAW ? 1 : 2); + recordBuffer.putInt(offset + SHORTEST, metadata.getLengthOfShortestElement()); + recordBuffer.putInt(offset + LONGEST, metadata.getLengthOfLongestElement()); + recordBuffer.putInt(offset + TOTAL_ENTRIES, metadata.getTotalNumberOfEntries()); + recordBuffer.putInt(offset + MAX_MV, metadata.getMaxNumberOfMultiValues()); + recordBuffer.putInt(offset + MAX_ROW, metadata.getMaxRowLengthInBytes()); + recordBuffer.putInt(offset + BITS, metadata.getBitsPerElement()); + writeValue(recordBuffer, offset + MIN_TAG, offset + MIN_RAW, offset + MIN_BLOB, metadata.getMinValue(), blobs); + writeValue(recordBuffer, offset + MAX_TAG, offset + MAX_RAW, offset + MAX_BLOB, metadata.getMaxValue(), blobs); + PartitionFunction partitionFunction = metadata.getPartitionFunction(); + putRef(recordBuffer, offset + PARTITION_FUNCTION, + partitionFunction != null ? blobs.add(encodePartitionFunction(column, config)) : BlobRef.ABSENT); + putRef(recordBuffer, offset + PARTITIONS, + metadata.getPartitions() != null ? blobs.add(encodePartitions(metadata.getPartitions())) : BlobRef.ABSENT); + if (metadata instanceof ColumnMetadataImpl) { + ColumnMetadataImpl impl = (ColumnMetadataImpl) metadata; + putRef(recordBuffer, offset + PARENT, blobs.add(impl.getParentColumn())); + putRef(recordBuffer, offset + SPARSE_KEYS, blobs.addJson(impl.getSparseKeys())); + putRef(recordBuffer, offset + SPARSE_MULTI_VALUE_KEYS, blobs.addJson(impl.getSparseMultiValueKeys())); + putRef(recordBuffer, offset + SPARSE_KEY_TYPES, blobs.addJson(impl.getSparseKeyTypes())); + } else { + putRef(recordBuffer, offset + PARENT, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_KEYS, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_MULTI_VALUE_KEYS, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_KEY_TYPES, BlobRef.ABSENT); + } + recordBuffer.putLong(offset + RAW_UNCOMPRESSED_SIZE, + metadata.getRawForwardIndexUncompressedValueSizeInBytes()); + recordBuffer.putLong(offset + DICTIONARY_UNCOMPRESSED_SIZE, + metadata.getDictionaryEncodedUncompressedValueSizeInBytes()); + putRef(recordBuffer, offset + COMPRESSION_TYPE, + blobs.add(metadata.getRawForwardIndexChunkCompressionType() != null + ? metadata.getRawForwardIndexChunkCompressionType().name() : null)); + List<IndexSize> sizes = indexSizes.getOrDefault(column, List.of()); + recordBuffer.putInt(offset + INDEX_COUNT, sizes.size()); + putRef(recordBuffer, offset + INDEXES, sizes.isEmpty() ? BlobRef.ABSENT : blobs.add(encodeIndexSizes(sizes))); + } + byte[] blobBytes = blobs.toByteArray(); + long totalSize = (long) HEADER_SIZE + records.length + blobBytes.length; + if (totalSize > Integer.MAX_VALUE) { + throw new IOException("Mapped metadata exceeds 2 GiB: " + totalSize); + } + byte[] bytes = new byte[(int) totalSize]; + ByteBuffer header = ByteBuffer.wrap(bytes).order(ByteOrder.BIG_ENDIAN); + header.putInt(0, MAGIC); + header.putInt(4, VERSION); + header.putLong(8, totalSize); + header.putLong(16, propertiesFingerprint._length); + header.putInt(24, propertiesFingerprint._crc); + header.putLong(28, indexMapFingerprint._length); + header.putInt(36, indexMapFingerprint._crc); + header.putInt(44, segmentConfig._offset); + header.putInt(48, segmentConfig._length); + header.putInt(52, columns.size()); + header.putInt(56, HEADER_SIZE); + header.putInt(60, blobBase); + System.arraycopy(records, 0, bytes, HEADER_SIZE, records.length); + System.arraycopy(blobBytes, 0, bytes, blobBase, blobBytes.length); + header.putInt(40, sidecarCrc(bytes)); + return bytes; + } + + private static ObjectNode fieldSpecJson(FieldSpec fieldSpec) { + ObjectNode json = fieldSpec.toJsonObject(); Review Comment: **BIG_DECIMAL default null values change in this round trip.** `toJsonObject()` writes the default as a JSON number, and the parse at L202 uses `JsonUtils.DEFAULT_READER`, which reads floats as `double` before `FieldSpec.setDefaultNullValue(JsonNode)` calls `asText()`. I checked it with jshell against `FieldSpec` at this head: | `defaultNullValue` literal | eager | mapped | | --- | --- | --- | | `1.10` | `1.10` | `1.1` | | `12345678901234567890.123456789` | unchanged | `12345678901234567000` | So for a BIG_DECIMAL column with a non-default default null value, the mapped `getFieldSpec()` no longer equals the eager one. The segment metadata REST response shows the changed default. For auto-generated columns, `BaseDefaultColumnHandler.computeDefaultColumnActionMap()` compares `getDefaultNullValueString()` with the schema, and `isReloadNeeded()` runs that check on the live mapped directory. Such a segment keeps reporting `UPDATE_*_DEFAULT_VALUE`, while the reload itself (on eager metadata) finds nothing to change. `READER_WITH_BIG_DECIMAL` alone doesn't fix it: `1.10` still comes back as `1.1` unless `STRIP_TRAILING_BIGDECIMAL_ZEROES` is also disabled. Writing numeric defaults as text round-trips exactly, because the setter only uses `asText()`. I verified BIG_DECIMAL `1.10` and the high-precision value, FLOAT `-0.0`, and DOUBLE `NaN`: ```java JsonNode defaultNullValue = json.get("defaultNullValue"); if (defaultNullValue != null && defaultNullValue.isNumber()) { json.put("defaultNullValue", fieldSpec.getDefaultNullValueString()); } ``` Placed here, the recursion covers complex children too. It changes the encoding, so it needs `VERSION = 5`. A BIG_DECIMAL `1.10` column in the eager-vs-mapped parity test would have caught this. ########## pinot-segment-local/src/main/java/org/apache/pinot/segment/local/upsert/BasePartitionUpsertMetadataManager.java: ########## @@ -1001,9 +1001,21 @@ private void recordSnapshotPass(boolean consistent) { TreeMap<String, Long> segmentVersions = new TreeMap<>(); for (IndexSegment segment : _trackedSegments) { if (segment instanceof ImmutableSegmentImpl) { - SegmentMetadata segmentMetadata = segment.getSegmentMetadata(); - long dataCrc = segmentMetadata.getDataCrc(); - segmentVersions.put(segment.getSegmentName(), dataCrc >= 0 ? dataCrc : segmentMetadata.getCrc()); + if (!tryAcquireSegmentReadLock(segment)) { Review Comment: Nit: this block only reads `getDataCrc()`, `getCrc()`, and `getSegmentName()`, which are heap fields on `SegmentMetadataImpl` in both modes. The read lock here doesn't protect any mapped read (unlike `getWatermarkToPersist()`, which reads the mapped max). What it does change is eager behavior: a tracked segment that's already destroyed now marks the pass inconsistent instead of being counted. If that tightening is intended, a comment saying so would help. Otherwise I'd drop this hunk to keep the PR focused on the metadata change. ########## pinot-segment-spi/src/main/java/org/apache/pinot/segment/spi/index/metadata/MappedSegmentMetadataStore.java: ########## @@ -0,0 +1,904 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package org.apache.pinot.segment.spi.index.metadata; + +import com.fasterxml.jackson.databind.node.ObjectNode; +import com.google.common.collect.Interner; +import com.google.common.collect.Interners; +import java.io.ByteArrayInputStream; +import java.io.ByteArrayOutputStream; +import java.io.Closeable; +import java.io.DataInputStream; +import java.io.DataOutputStream; +import java.io.File; +import java.io.IOException; +import java.math.BigDecimal; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; +import java.nio.charset.StandardCharsets; +import java.nio.file.AtomicMoveNotSupportedException; +import java.nio.file.Files; +import java.nio.file.Path; +import java.nio.file.StandardCopyOption; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Comparator; +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import java.util.Set; +import java.util.stream.Collectors; +import java.util.zip.CRC32C; +import javax.annotation.Nullable; +import org.apache.commons.configuration2.Configuration; +import org.apache.commons.configuration2.PropertiesConfiguration; +import org.apache.commons.configuration2.ex.ConfigurationException; +import org.apache.pinot.segment.spi.ColumnMetadata; +import org.apache.pinot.segment.spi.V1Constants; +import org.apache.pinot.segment.spi.creator.SegmentVersion; +import org.apache.pinot.segment.spi.index.IndexService; +import org.apache.pinot.segment.spi.index.IndexType; +import org.apache.pinot.segment.spi.memory.PinotDataBuffer; +import org.apache.pinot.segment.spi.partition.PartitionFunction; +import org.apache.pinot.segment.spi.store.ColumnIndexUtils; +import org.apache.pinot.segment.spi.store.SegmentDirectoryPaths; +import org.apache.pinot.segment.spi.utils.SegmentMetadataUtils; +import org.apache.pinot.spi.config.table.FieldConfig.EncodingType; +import org.apache.pinot.spi.data.ComplexFieldSpec; +import org.apache.pinot.spi.data.FieldSpec; +import org.apache.pinot.spi.env.CommonsConfigurationUtils; +import org.apache.pinot.spi.utils.ByteArray; +import org.apache.pinot.spi.utils.ColumnNameInterner; +import org.apache.pinot.spi.utils.JsonUtils; + +/// Versioned, file-backed metadata for a sealed segment. The source `metadata.properties` remains the canonical +/// representation for old readers and preprocessing. +/// +/// The big-endian file starts with a 64-byte header containing the magic, format version, total length, source +/// `metadata.properties` and `index_map` length/CRC fingerprints, payload CRC, segment-property blob reference, +/// column count, and record-table bounds. Sorted columns occupy fixed 180-byte records, followed by variable blobs; +/// each blob reference is an absolute 32-bit offset and length. Numeric column statistics are read directly from the +/// mapping. The column layout retains names and reusable [FieldSpec] references on heap. +/// +/// A stale or corrupt sidecar is rebuilt from the canonical source and replaced locally, using atomic rename where +/// supported. Change `VERSION` whenever the encoding or the values derived from canonical metadata change; old +/// versions are rebuilt rather than migrated. +final class MappedSegmentMetadataStore implements Closeable { + static final String FILE_NAME = "segment_metadata.bin"; + private static final String TEMP_FILE_PREFIX = ".segment_metadata."; + private static final String TEMP_FILE_SUFFIX = ".tmp"; + + static boolean isSidecarFileName(String fileName) { + return FILE_NAME.equals(fileName) + || (fileName.startsWith(TEMP_FILE_PREFIX) && fileName.endsWith(TEMP_FILE_SUFFIX)); + } + + private static final int MAGIC = 0x504d4431; // PMD1 + private static final int VERSION = 4; + private static final int HEADER_SIZE = 64; + private static final int RECORD_SIZE = 180; + + static final int NAME = 0; + static final int FIELD_SPEC = 8; + static final int TOTAL_DOCS = 16; + static final int CARDINALITY = 20; + static final int FLAGS = 24; + static final int ENCODING = 28; + static final int SHORTEST = 32; + static final int LONGEST = 36; + static final int TOTAL_ENTRIES = 40; + static final int MAX_MV = 44; + static final int MAX_ROW = 48; + static final int BITS = 52; + static final int MIN_TAG = 56; + static final int MIN_RAW = 60; + static final int MIN_BLOB = 68; + static final int MAX_TAG = 76; + static final int MAX_RAW = 80; + static final int MAX_BLOB = 88; + static final int PARTITION_FUNCTION = 96; + static final int PARTITIONS = 104; + static final int PARENT = 112; + static final int SPARSE_KEYS = 120; + static final int SPARSE_MULTI_VALUE_KEYS = 128; + static final int SPARSE_KEY_TYPES = 136; + static final int RAW_UNCOMPRESSED_SIZE = 144; + static final int DICTIONARY_UNCOMPRESSED_SIZE = 152; + static final int COMPRESSION_TYPE = 160; + static final int INDEXES = 168; + static final int INDEX_COUNT = 176; + + static final int HAS_DICTIONARY = 1; + static final int SORTED = 1 << 1; + static final int NON_NULL = 1 << 2; + static final int INVALID_MIN_MAX = 1 << 3; + static final int ASCII = 1 << 4; + static final int AUTO_GENERATED = 1 << 5; + + private static final int NO_VALUE = 0; + private static final int INT_VALUE = 1; + private static final int LONG_VALUE = 2; + private static final int FLOAT_VALUE = 3; + private static final int DOUBLE_VALUE = 4; + private static final int DECIMAL_VALUE = 5; + private static final int STRING_VALUE = 6; + private static final int BYTES_VALUE = 7; + + private static final Interner<FieldSpec> FIELD_SPECS = Interners.newWeakInterner(); + private static final Interner<ColumnLayout> COLUMN_LAYOUTS = Interners.newWeakInterner(); + + private final PinotDataBuffer _buffer; + private final ColumnLayout _columns; + private final int _tableOffset; + private final int _segmentConfigOffset; + private final int _segmentConfigLength; + + private MappedSegmentMetadataStore(PinotDataBuffer buffer, ColumnLayout columns, int tableOffset, + int segmentConfigOffset, int segmentConfigLength) { + _buffer = buffer; + _columns = columns; + _tableOffset = tableOffset; + _segmentConfigOffset = segmentConfigOffset; + _segmentConfigLength = segmentConfigLength; + } + + static MappedSegmentMetadataStore open(File indexDir) + throws IOException, ConfigurationException { + File source = SegmentDirectoryPaths.findMetadataFile(indexDir); + if (source == null) { + throw new IOException("Missing metadata.properties in " + indexDir); + } + File sidecar = new File(indexDir, FILE_NAME); + boolean rebuilt = false; + for (int attempt = 0; attempt < 3; attempt++) { + Fingerprint propertiesFingerprint = Fingerprint.of(source); + Fingerprint indexMapFingerprint = Fingerprint.of(findIndexMap(indexDir)); + // The validated heap copy is transient; the store retains only the mapping and shared column layout. + byte[] bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + writeAtomically(sidecar, build(indexDir, propertiesFingerprint, indexMapFingerprint)); + rebuilt = true; + bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + continue; + } + } + PinotDataBuffer buffer = PinotDataBuffer.mapFile(sidecar, true, 0, bytes.length, ByteOrder.BIG_ENDIAN, + "segment metadata"); + boolean keepBuffer = false; + try { + // Another loader may atomically replace the sidecar between validation and mapFile(). Check the exact mapped + // bytes and the canonical inputs before publishing a view over this mapping. The header alone cannot prove + // that the mapped payload is the same bytes we validated. + if (!matchesMapped(buffer, bytes) || !propertiesFingerprint.sameAs(Fingerprint.of(source)) + || !indexMapFingerprint.sameAs(Fingerprint.of(findIndexMap(indexDir)))) { + continue; + } + ByteBuffer header = ByteBuffer.wrap(bytes).order(ByteOrder.BIG_ENDIAN); + int count = header.getInt(52); + int tableOffset = header.getInt(56); + String[] names = new String[count]; + FieldSpec[] fieldSpecs = new FieldSpec[count]; + for (int i = 0; i < count; i++) { + int record = tableOffset + i * RECORD_SIZE; + names[i] = ColumnNameInterner.intern(readString(bytes, header.getInt(record + NAME), + header.getInt(record + NAME + 4))); + FieldSpec parsed = JsonUtils.stringToObject(readString(bytes, header.getInt(record + FIELD_SPEC), + header.getInt(record + FIELD_SPEC + 4)), FieldSpec.class); + fieldSpecs[i] = parsed instanceof ComplexFieldSpec ? parsed : FIELD_SPECS.intern(parsed); + } + ColumnLayout columns = COLUMN_LAYOUTS.intern(new ColumnLayout(names, fieldSpecs)); + MappedSegmentMetadataStore store = new MappedSegmentMetadataStore(buffer, columns, tableOffset, + header.getInt(44), header.getInt(48)); + keepBuffer = true; + return store; + } catch (IOException | RuntimeException e) { + if (!rebuilt) { + // A checksum-valid older sidecar can still contain FieldSpec JSON this reader cannot parse. Rebuild it + // once from canonical metadata instead of falling back to eager on every load. + writeAtomically(sidecar, build(indexDir, propertiesFingerprint, indexMapFingerprint)); + rebuilt = true; + continue; + } + throw new IOException("Invalid field specs in mapped segment metadata: " + sidecar, e); + } finally { + if (!keepBuffer) { + buffer.close(); + } + } + } + throw new IOException("Metadata changed while mapping segment: " + indexDir); + } + + private static boolean matchesMapped(PinotDataBuffer buffer, byte[] bytes) { + if (buffer.size() != bytes.length) { + return false; + } + byte[] block = new byte[Math.min(8192, bytes.length)]; + for (int offset = 0; offset < bytes.length; offset += block.length) { + int length = Math.min(block.length, bytes.length - offset); + buffer.copyTo(offset, block, 0, length); + for (int i = 0; i < length; i++) { + if (block[i] != bytes[offset + i]) { + return false; + } + } + } + return true; + } + + int size() { + return _columns._names.length; + } + + String name(int ordinal) { + return _columns._names[ordinal]; + } + + int ordinal(String column) { + return Arrays.binarySearch(_columns._names, column); + } + + int getInt(int ordinal, int offset) { + return _buffer.getInt((long) _tableOffset + (long) ordinal * RECORD_SIZE + offset); + } + + long getLong(int ordinal, int offset) { + return _buffer.getLong((long) _tableOffset + (long) ordinal * RECORD_SIZE + offset); + } + + boolean flag(int ordinal, int flag) { + return (getInt(ordinal, FLAGS) & flag) != 0; + } + + @Nullable + String string(int ordinal, int field) { + byte[] bytes = blob(ordinal, field); + return bytes != null ? new String(bytes, StandardCharsets.UTF_8) : null; + } + + @Nullable + byte[] blob(int ordinal, int field) { + int offset = getInt(ordinal, field); + if (offset < 0) { + return null; + } + int length = getInt(ordinal, field + 4); + byte[] bytes = new byte[length]; + _buffer.copyTo(offset, bytes); + return bytes; + } + + FieldSpec fieldSpec(int ordinal) { + return _columns._fieldSpecs[ordinal]; + } + + @Nullable + Comparable<?> value(int ordinal, boolean minimum) { + int tag = getInt(ordinal, minimum ? MIN_TAG : MAX_TAG); + long raw = getLong(ordinal, minimum ? MIN_RAW : MAX_RAW); + int blobField = minimum ? MIN_BLOB : MAX_BLOB; + switch (tag) { + case NO_VALUE: + return null; + case INT_VALUE: + return (int) raw; + case LONG_VALUE: + return raw; + case FLOAT_VALUE: + return Float.intBitsToFloat((int) raw); + case DOUBLE_VALUE: + return Double.longBitsToDouble(raw); + case DECIMAL_VALUE: + return new BigDecimal(string(ordinal, blobField)); + case STRING_VALUE: + return string(ordinal, blobField); + case BYTES_VALUE: + return new ByteArray(blob(ordinal, blobField)); + default: + throw new IllegalStateException("Invalid value tag: " + tag); + } + } + + PropertiesConfiguration segmentConfiguration() + throws IOException, ConfigurationException { + byte[] bytes = new byte[_segmentConfigLength]; + _buffer.copyTo(_segmentConfigOffset, bytes); + // Values were already split by the source parser. A second delimiter pass would change escaped commas. + PropertiesConfiguration config = CommonsConfigurationUtils.fromInputStream(null, false, null); + try (DataInputStream input = new DataInputStream(new ByteArrayInputStream(bytes))) { + int count = input.readInt(); + for (int i = 0; i < count; i++) { + String key = readString(input); + int valueCount = input.readInt(); + if (valueCount == 1) { + config.setProperty(key, readString(input)); + } else { + List<String> values = new ArrayList<>(valueCount); + for (int j = 0; j < valueCount; j++) { + values.add(readString(input)); + } + config.setProperty(key, values); + } + } + } + return config; + } + + @Override + public void close() + throws IOException { + _buffer.close(); + } + + private static byte[] build(File indexDir, Fingerprint propertiesFingerprint, Fingerprint indexMapFingerprint) + throws IOException, ConfigurationException { + SegmentMetadataImpl source = new SegmentMetadataImpl(indexDir); + PropertiesConfiguration config = SegmentMetadataUtils.getPropertiesConfiguration(indexDir); + List<String> columns = new ArrayList<>(source.getColumnMetadataMap().keySet()); + columns.sort(Comparator.naturalOrder()); + if ((long) columns.size() * RECORD_SIZE + HEADER_SIZE > Integer.MAX_VALUE) { + throw new IOException("Too many columns for mapped metadata: " + columns.size()); + } + int blobBase = HEADER_SIZE + columns.size() * RECORD_SIZE; + BlobWriter blobs = new BlobWriter(blobBase); + BlobRef segmentConfig = blobs.add(encodeSegmentConfiguration(config)); + // EmptyColumnMetadata does not load index sizes even if a stale index_map is still present. + Map<String, List<IndexSize>> indexSizes = source.getTotalDocs() > 0 + && source.getVersion() == SegmentVersion.v3 ? readIndexSizes(indexDir) : Map.of(); + byte[] records = new byte[columns.size() * RECORD_SIZE]; + ByteBuffer recordBuffer = ByteBuffer.wrap(records).order(ByteOrder.BIG_ENDIAN); + for (int ordinal = 0; ordinal < columns.size(); ordinal++) { + String column = columns.get(ordinal); + ColumnMetadata metadata = source.getColumnMetadataFor(column); + int offset = ordinal * RECORD_SIZE; + putRef(recordBuffer, offset + NAME, blobs.add(column)); + putRef(recordBuffer, offset + FIELD_SPEC, blobs.add(fieldSpecJson(metadata.getFieldSpec()).toString())); + recordBuffer.putInt(offset + TOTAL_DOCS, metadata.getTotalDocs()); + recordBuffer.putInt(offset + CARDINALITY, metadata.getCardinality()); + int flags = (metadata.hasDictionary() ? HAS_DICTIONARY : 0) | (metadata.isSorted() ? SORTED : 0) + | (metadata.isNonNull() ? NON_NULL : 0) | (metadata.isMinMaxValueInvalid() ? INVALID_MIN_MAX : 0) + | (metadata.isAscii() ? ASCII : 0) | (metadata.isAutoGenerated() ? AUTO_GENERATED : 0); + recordBuffer.putInt(offset + FLAGS, flags); + EncodingType encoding = metadata.getForwardIndexEncoding(); + recordBuffer.putInt(offset + ENCODING, encoding == null ? 0 : encoding == EncodingType.RAW ? 1 : 2); + recordBuffer.putInt(offset + SHORTEST, metadata.getLengthOfShortestElement()); + recordBuffer.putInt(offset + LONGEST, metadata.getLengthOfLongestElement()); + recordBuffer.putInt(offset + TOTAL_ENTRIES, metadata.getTotalNumberOfEntries()); + recordBuffer.putInt(offset + MAX_MV, metadata.getMaxNumberOfMultiValues()); + recordBuffer.putInt(offset + MAX_ROW, metadata.getMaxRowLengthInBytes()); + recordBuffer.putInt(offset + BITS, metadata.getBitsPerElement()); + writeValue(recordBuffer, offset + MIN_TAG, offset + MIN_RAW, offset + MIN_BLOB, metadata.getMinValue(), blobs); + writeValue(recordBuffer, offset + MAX_TAG, offset + MAX_RAW, offset + MAX_BLOB, metadata.getMaxValue(), blobs); + PartitionFunction partitionFunction = metadata.getPartitionFunction(); + putRef(recordBuffer, offset + PARTITION_FUNCTION, + partitionFunction != null ? blobs.add(encodePartitionFunction(column, config)) : BlobRef.ABSENT); + putRef(recordBuffer, offset + PARTITIONS, + metadata.getPartitions() != null ? blobs.add(encodePartitions(metadata.getPartitions())) : BlobRef.ABSENT); + if (metadata instanceof ColumnMetadataImpl) { + ColumnMetadataImpl impl = (ColumnMetadataImpl) metadata; + putRef(recordBuffer, offset + PARENT, blobs.add(impl.getParentColumn())); + putRef(recordBuffer, offset + SPARSE_KEYS, blobs.addJson(impl.getSparseKeys())); + putRef(recordBuffer, offset + SPARSE_MULTI_VALUE_KEYS, blobs.addJson(impl.getSparseMultiValueKeys())); + putRef(recordBuffer, offset + SPARSE_KEY_TYPES, blobs.addJson(impl.getSparseKeyTypes())); + } else { + putRef(recordBuffer, offset + PARENT, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_KEYS, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_MULTI_VALUE_KEYS, BlobRef.ABSENT); + putRef(recordBuffer, offset + SPARSE_KEY_TYPES, BlobRef.ABSENT); + } + recordBuffer.putLong(offset + RAW_UNCOMPRESSED_SIZE, + metadata.getRawForwardIndexUncompressedValueSizeInBytes()); + recordBuffer.putLong(offset + DICTIONARY_UNCOMPRESSED_SIZE, + metadata.getDictionaryEncodedUncompressedValueSizeInBytes()); + putRef(recordBuffer, offset + COMPRESSION_TYPE, + blobs.add(metadata.getRawForwardIndexChunkCompressionType() != null + ? metadata.getRawForwardIndexChunkCompressionType().name() : null)); + List<IndexSize> sizes = indexSizes.getOrDefault(column, List.of()); + recordBuffer.putInt(offset + INDEX_COUNT, sizes.size()); + putRef(recordBuffer, offset + INDEXES, sizes.isEmpty() ? BlobRef.ABSENT : blobs.add(encodeIndexSizes(sizes))); + } + byte[] blobBytes = blobs.toByteArray(); + long totalSize = (long) HEADER_SIZE + records.length + blobBytes.length; + if (totalSize > Integer.MAX_VALUE) { + throw new IOException("Mapped metadata exceeds 2 GiB: " + totalSize); + } + byte[] bytes = new byte[(int) totalSize]; + ByteBuffer header = ByteBuffer.wrap(bytes).order(ByteOrder.BIG_ENDIAN); + header.putInt(0, MAGIC); + header.putInt(4, VERSION); + header.putLong(8, totalSize); + header.putLong(16, propertiesFingerprint._length); + header.putInt(24, propertiesFingerprint._crc); + header.putLong(28, indexMapFingerprint._length); + header.putInt(36, indexMapFingerprint._crc); + header.putInt(44, segmentConfig._offset); + header.putInt(48, segmentConfig._length); + header.putInt(52, columns.size()); + header.putInt(56, HEADER_SIZE); + header.putInt(60, blobBase); + System.arraycopy(records, 0, bytes, HEADER_SIZE, records.length); + System.arraycopy(blobBytes, 0, bytes, blobBase, blobBytes.length); + header.putInt(40, sidecarCrc(bytes)); + return bytes; + } + + private static ObjectNode fieldSpecJson(FieldSpec fieldSpec) { + ObjectNode json = fieldSpec.toJsonObject(); + // TimeFieldSpec omits the discriminator, including when nested in a ComplexFieldSpec. + json.put("fieldType", fieldSpec.getFieldType().toString()); + if (fieldSpec instanceof ComplexFieldSpec) { + ObjectNode children = (ObjectNode) json.get("childFieldSpecs"); + for (Map.Entry<String, FieldSpec> entry : ((ComplexFieldSpec) fieldSpec).getChildFieldSpecs().entrySet()) { + children.set(entry.getKey(), fieldSpecJson(entry.getValue())); + } + } + return json; + } + + private static byte[] encodeSegmentConfiguration(PropertiesConfiguration config) + throws IOException { + ByteArrayOutputStream bytes = new ByteArrayOutputStream(); + try (DataOutputStream output = new DataOutputStream(bytes)) { + List<String> keys = CommonsConfigurationUtils.getKeys(config).stream() + .filter(key -> !key.startsWith(V1Constants.MetadataKeys.Column.COLUMN_PROPS_KEY_PREFIX)) + .sorted().collect(Collectors.toList()); + output.writeInt(keys.size()); + for (String key : keys) { + writeString(output, key); + Object value = config.getProperty(key); + if (value instanceof List<?>) { + List<?> values = (List<?>) value; + output.writeInt(values.size()); + for (Object entry : values) { + writeString(output, String.valueOf(entry)); + } + } else { + output.writeInt(1); + writeString(output, String.valueOf(value)); + } + } + } + return bytes.toByteArray(); + } + + private static byte[] encodePartitionFunction(String column, PropertiesConfiguration metadata) + throws IOException { + String prefix = V1Constants.MetadataKeys.Column.COLUMN_PROPS_KEY_PREFIX + column + "."; + String name = metadata.getString(prefix + V1Constants.MetadataKeys.Column.PARTITION_FUNCTION); + int numPartitions = metadata.getInt(prefix + V1Constants.MetadataKeys.Column.NUM_PARTITIONS); + Configuration functionConfig = metadata.subset(prefix + V1Constants.MetadataKeys.Column.PARTITION_FUNCTION_CONFIG); + Map<String, String> config = new HashMap<>(); + functionConfig.forEach((key, value) -> config.put(key, + value instanceof List ? ((List<?>) value).stream().map(String::valueOf).collect(Collectors.joining(",")) + : value.toString())); + ByteArrayOutputStream bytes = new ByteArrayOutputStream(); + try (DataOutputStream output = new DataOutputStream(bytes)) { + writeString(output, name); + output.writeInt(numPartitions); + output.writeInt(config.size()); + for (String key : config.keySet().stream().sorted().collect(Collectors.toList())) { + writeString(output, key); + writeString(output, config.get(key)); + } + } + return bytes.toByteArray(); + } + + private static byte[] encodePartitions(Set<Integer> partitions) + throws IOException { + ByteArrayOutputStream bytes = new ByteArrayOutputStream(); + try (DataOutputStream output = new DataOutputStream(bytes)) { + output.writeInt(partitions.size()); + for (int partition : partitions.stream().sorted().collect(Collectors.toList())) { + output.writeInt(partition); + } + } + return bytes.toByteArray(); + } + + private static byte[] encodeIndexSizes(List<IndexSize> sizes) + throws IOException { + ByteArrayOutputStream bytes = new ByteArrayOutputStream(); + try (DataOutputStream output = new DataOutputStream(bytes)) { + for (IndexSize size : sizes) { + writeString(output, size._type); + output.writeLong(size._size); + } + } + return bytes.toByteArray(); + } + + private static Map<String, List<IndexSize>> readIndexSizes(File indexDir) + throws ConfigurationException { + File indexMap = findIndexMap(indexDir); + if (indexMap == null) { + return Map.of(); + } + // Match SegmentMetadataImpl: index sizes come only from the local v3 index_map. + PropertiesConfiguration config = CommonsConfigurationUtils.fromFile(indexMap); + Map<String, List<IndexSize>> result = new HashMap<>(); + for (String key : CommonsConfigurationUtils.getKeys(config)) { + try { + String[] parsed = ColumnIndexUtils.parseIndexMapKeys(key, indexDir.getPath()); + if (ColumnIndexUtils.MAP_KEY_NAME_SIZE.equals(parsed[2])) { + long size = config.getLong(key); + if (size < 0 || size > (1L << 48) - 1) { + continue; + } + IndexType<?, ?, ?> type = IndexService.getInstance().get(parsed[1]); Review Comment: Minor: index sizes depend on which index types `IndexService` has registered when the sidecar is built, and the cache key doesn't include that. An entry with a type unknown at build time is dropped here (the `catch` below), and `validateIndexSizes()` only rejects unknown types that are actually stored. If the type becomes available later (for example, rolling back across a release that added an index type and then forward again, or an index plugin that failed to load once), the old sidecar still validates. Mapped `getIndexSizeFor()` then keeps returning `UNAVAILABLE` for that index until `metadata.properties` or `index_map` changes, while eager metadata picks it up on the next load. The impact is limited to the size and compression-stats endpoints. Two possible fixes: fold the registered index ids into the header fingerprint, or flag a sidecar that dropped entries so `open()` rebuilds it. ########## pinot-segment-spi/src/main/java/org/apache/pinot/segment/spi/index/metadata/MappedSegmentMetadataStore.java: ########## @@ -0,0 +1,904 @@ +/** + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, + * software distributed under the License is distributed on an + * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY + * KIND, either express or implied. See the License for the + * specific language governing permissions and limitations + * under the License. + */ +package org.apache.pinot.segment.spi.index.metadata; + +import com.fasterxml.jackson.databind.node.ObjectNode; +import com.google.common.collect.Interner; +import com.google.common.collect.Interners; +import java.io.ByteArrayInputStream; +import java.io.ByteArrayOutputStream; +import java.io.Closeable; +import java.io.DataInputStream; +import java.io.DataOutputStream; +import java.io.File; +import java.io.IOException; +import java.math.BigDecimal; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; +import java.nio.charset.StandardCharsets; +import java.nio.file.AtomicMoveNotSupportedException; +import java.nio.file.Files; +import java.nio.file.Path; +import java.nio.file.StandardCopyOption; +import java.util.ArrayList; +import java.util.Arrays; +import java.util.Comparator; +import java.util.HashMap; +import java.util.List; +import java.util.Map; +import java.util.Set; +import java.util.stream.Collectors; +import java.util.zip.CRC32C; +import javax.annotation.Nullable; +import org.apache.commons.configuration2.Configuration; +import org.apache.commons.configuration2.PropertiesConfiguration; +import org.apache.commons.configuration2.ex.ConfigurationException; +import org.apache.pinot.segment.spi.ColumnMetadata; +import org.apache.pinot.segment.spi.V1Constants; +import org.apache.pinot.segment.spi.creator.SegmentVersion; +import org.apache.pinot.segment.spi.index.IndexService; +import org.apache.pinot.segment.spi.index.IndexType; +import org.apache.pinot.segment.spi.memory.PinotDataBuffer; +import org.apache.pinot.segment.spi.partition.PartitionFunction; +import org.apache.pinot.segment.spi.store.ColumnIndexUtils; +import org.apache.pinot.segment.spi.store.SegmentDirectoryPaths; +import org.apache.pinot.segment.spi.utils.SegmentMetadataUtils; +import org.apache.pinot.spi.config.table.FieldConfig.EncodingType; +import org.apache.pinot.spi.data.ComplexFieldSpec; +import org.apache.pinot.spi.data.FieldSpec; +import org.apache.pinot.spi.env.CommonsConfigurationUtils; +import org.apache.pinot.spi.utils.ByteArray; +import org.apache.pinot.spi.utils.ColumnNameInterner; +import org.apache.pinot.spi.utils.JsonUtils; + +/// Versioned, file-backed metadata for a sealed segment. The source `metadata.properties` remains the canonical +/// representation for old readers and preprocessing. +/// +/// The big-endian file starts with a 64-byte header containing the magic, format version, total length, source +/// `metadata.properties` and `index_map` length/CRC fingerprints, payload CRC, segment-property blob reference, +/// column count, and record-table bounds. Sorted columns occupy fixed 180-byte records, followed by variable blobs; +/// each blob reference is an absolute 32-bit offset and length. Numeric column statistics are read directly from the +/// mapping. The column layout retains names and reusable [FieldSpec] references on heap. +/// +/// A stale or corrupt sidecar is rebuilt from the canonical source and replaced locally, using atomic rename where +/// supported. Change `VERSION` whenever the encoding or the values derived from canonical metadata change; old +/// versions are rebuilt rather than migrated. +final class MappedSegmentMetadataStore implements Closeable { + static final String FILE_NAME = "segment_metadata.bin"; + private static final String TEMP_FILE_PREFIX = ".segment_metadata."; + private static final String TEMP_FILE_SUFFIX = ".tmp"; + + static boolean isSidecarFileName(String fileName) { + return FILE_NAME.equals(fileName) + || (fileName.startsWith(TEMP_FILE_PREFIX) && fileName.endsWith(TEMP_FILE_SUFFIX)); + } + + private static final int MAGIC = 0x504d4431; // PMD1 + private static final int VERSION = 4; + private static final int HEADER_SIZE = 64; + private static final int RECORD_SIZE = 180; + + static final int NAME = 0; + static final int FIELD_SPEC = 8; + static final int TOTAL_DOCS = 16; + static final int CARDINALITY = 20; + static final int FLAGS = 24; + static final int ENCODING = 28; + static final int SHORTEST = 32; + static final int LONGEST = 36; + static final int TOTAL_ENTRIES = 40; + static final int MAX_MV = 44; + static final int MAX_ROW = 48; + static final int BITS = 52; + static final int MIN_TAG = 56; + static final int MIN_RAW = 60; + static final int MIN_BLOB = 68; + static final int MAX_TAG = 76; + static final int MAX_RAW = 80; + static final int MAX_BLOB = 88; + static final int PARTITION_FUNCTION = 96; + static final int PARTITIONS = 104; + static final int PARENT = 112; + static final int SPARSE_KEYS = 120; + static final int SPARSE_MULTI_VALUE_KEYS = 128; + static final int SPARSE_KEY_TYPES = 136; + static final int RAW_UNCOMPRESSED_SIZE = 144; + static final int DICTIONARY_UNCOMPRESSED_SIZE = 152; + static final int COMPRESSION_TYPE = 160; + static final int INDEXES = 168; + static final int INDEX_COUNT = 176; + + static final int HAS_DICTIONARY = 1; + static final int SORTED = 1 << 1; + static final int NON_NULL = 1 << 2; + static final int INVALID_MIN_MAX = 1 << 3; + static final int ASCII = 1 << 4; + static final int AUTO_GENERATED = 1 << 5; + + private static final int NO_VALUE = 0; + private static final int INT_VALUE = 1; + private static final int LONG_VALUE = 2; + private static final int FLOAT_VALUE = 3; + private static final int DOUBLE_VALUE = 4; + private static final int DECIMAL_VALUE = 5; + private static final int STRING_VALUE = 6; + private static final int BYTES_VALUE = 7; + + private static final Interner<FieldSpec> FIELD_SPECS = Interners.newWeakInterner(); + private static final Interner<ColumnLayout> COLUMN_LAYOUTS = Interners.newWeakInterner(); + + private final PinotDataBuffer _buffer; + private final ColumnLayout _columns; + private final int _tableOffset; + private final int _segmentConfigOffset; + private final int _segmentConfigLength; + + private MappedSegmentMetadataStore(PinotDataBuffer buffer, ColumnLayout columns, int tableOffset, + int segmentConfigOffset, int segmentConfigLength) { + _buffer = buffer; + _columns = columns; + _tableOffset = tableOffset; + _segmentConfigOffset = segmentConfigOffset; + _segmentConfigLength = segmentConfigLength; + } + + static MappedSegmentMetadataStore open(File indexDir) + throws IOException, ConfigurationException { + File source = SegmentDirectoryPaths.findMetadataFile(indexDir); + if (source == null) { + throw new IOException("Missing metadata.properties in " + indexDir); + } + File sidecar = new File(indexDir, FILE_NAME); + boolean rebuilt = false; + for (int attempt = 0; attempt < 3; attempt++) { + Fingerprint propertiesFingerprint = Fingerprint.of(source); + Fingerprint indexMapFingerprint = Fingerprint.of(findIndexMap(indexDir)); + // The validated heap copy is transient; the store retains only the mapping and shared column layout. + byte[] bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + writeAtomically(sidecar, build(indexDir, propertiesFingerprint, indexMapFingerprint)); + rebuilt = true; + bytes = readValid(sidecar, propertiesFingerprint, indexMapFingerprint); + if (bytes == null) { + continue; + } + } + PinotDataBuffer buffer = PinotDataBuffer.mapFile(sidecar, true, 0, bytes.length, ByteOrder.BIG_ENDIAN, Review Comment: Should this honor the segment's `ReadMode`? Under `ReadMode.heap`, index buffers go through `PinotDataBuffer.loadFile()` (`SingleFileIndexDirectory`, `FilePerIndexDirectory`), so a table with `loadMode: HEAP` has no file mappings today. With the instance flag on, every segment of such a table gains one, which is the VMA headroom the description lists as a rollout gate. `SegmentLocalFSDirectory.loadMappedMetadata()` already has `_readMode`. Passing it through and using `loadFile()` in heap mode would keep the Java-heap savings without adding mappings. -- 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]
