StarRocks分布式元数据源码解析

1. 支持元数据表

https://github.com/StarRocks/starrocks/pull/44276/files

核心类：LogicalIcebergMetadataTable，Iceberg元数据表，将元数据的各个字段做成表的列，后期可以通过sql操作从元数据获取字段，这个表的组成字段是DataFile相关的字段

java 复制代码

public static LogicalIcebergMetadataTable create(String catalogName, String originDb, String originTable) {
    return new LogicalIcebergMetadataTable(catalogName,
            ConnectorTableId.CONNECTOR_ID_GENERATOR.getNextId().asInt(),
            ICEBERG_LOGICAL_METADATA_TABLE_NAME,
            Table.TableType.METADATA,
            builder()
                    .columns(PLACEHOLDER_COLUMNS)
                    .column("content", ScalarType.createType(PrimitiveType.INT))
                    .column("file_path", ScalarType.createVarcharType())
                    .column("file_format", ScalarType.createVarcharType())
                    .column("spec_id", ScalarType.createType(PrimitiveType.INT))
                    .column("partition_data", ScalarType.createType(PrimitiveType.VARBINARY))
                    .column("record_count", ScalarType.createType(PrimitiveType.BIGINT))
                    .column("file_size_in_bytes", ScalarType.createType(PrimitiveType.BIGINT))
                    .column("split_offsets", ARRAY_BIGINT)
                    .column("sort_id", ScalarType.createType(PrimitiveType.INT))
                    .column("equality_ids", ARRAY_INT)
                    .column("file_sequence_number", ScalarType.createType(PrimitiveType.BIGINT))
                    .column("data_sequence_number", ScalarType.createType(PrimitiveType.BIGINT))
                    .column("column_stats", ScalarType.createType(PrimitiveType.VARBINARY))
                    .column("key_metadata", ScalarType.createType(PrimitiveType.VARBINARY))
                    .build(),
            originDb,
            originTable,
            MetadataTableType.LOGICAL_ICEBERG_METADATA);
}

2. Iceberg表扫描

https://github.com/StarRocks/starrocks/pull/44313

核心类：StarRocksIcebergTableScan，扫描Iceberg表的实现类，基于Iceberg的上层接口实现，类似Iceberg默认提供的DataTableScan，doPlanFiles中定义实际的元数据文件扫描逻辑

这一块应当属于数据上层扫描逻辑

java 复制代码

protected CloseableIterable<FileScanTask> doPlanFiles() {
    List<ManifestFile> dataManifests = findMatchingDataManifests(snapshot());
    List<ManifestFile> deleteManifests = findMatchingDeleteManifests(snapshot());

    boolean mayHaveEqualityDeletes = !deleteManifests.isEmpty() && mayHaveEqualityDeletes(snapshot());
    boolean loadColumnStats = mayHaveEqualityDeletes || shouldReturnColumnStats();

    if (shouldPlanLocally(dataManifests, loadColumnStats)) {
        return planFileTasksLocally(dataManifests, deleteManifests);
    } else {
        return planFileTasksRemotely(dataManifests, deleteManifests);
    }
}

3. Iceberg元数据信息接口

[Feature] Introduce meta spec interface by stephen-shelby · Pull Request #44527 · StarRocks/starrocks · GitHub

核心类：IcebergMetaSpec，Iceberg元数据描述，核心是RemoteMetaSplit的一个List，代表了元数据文件的列表，基于这个做分布式解析

这一块应当属于元数据文件的切片逻辑

java 复制代码

public List<RemoteMetaSplit> getSplits() {
    return splits;
}

4. Iceberg元数据扫描节点

https://github.com/StarRocks/starrocks/pull/44581

核心类：IcebergMetadataScanNode，Iceberg元数据的扫描节点，袭乘自PlanNode类，主要是把上节的RemoteMetaSplit放到StarRocks的执行结构当中

这一块属于Iceberg逻辑向StarRocks逻辑的中间转换层

java 复制代码

private void addSplitScanRangeLocations(RemoteMetaSplit split) {
    TScanRangeLocations scanRangeLocations = new TScanRangeLocations();

    THdfsScanRange hdfsScanRange = new THdfsScanRange();
    hdfsScanRange.setUse_iceberg_jni_metadata_reader(true);

    hdfsScanRange.setSerialized_split(split.getSerializeSplit());
    hdfsScanRange.setFile_length(split.length());
    hdfsScanRange.setLength(split.length());

    // for distributed scheduler
    hdfsScanRange.setFull_path(split.path());
    hdfsScanRange.setOffset(0);

    TScanRange scanRange = new TScanRange();
    scanRange.setHdfs_scan_range(hdfsScanRange);
    scanRangeLocations.setScan_range(scanRange);

    TScanRangeLocation scanRangeLocation = new TScanRangeLocation(new TNetworkAddress("-1", -1));
    scanRangeLocations.addToLocations(scanRangeLocation);

    result.add(scanRangeLocations);
}

5. Iceberg元数据读取

https://github.com/StarRocks/starrocks/pull/44632

核心类：IcebergMetadataScanner，这个应该是Iceberg元数据的实际读取类，实现自StarRocks的ConnectorScanner

ConnectorScanner是StarRocks的设计的介于C++-based的BE和Java-based的大数据组件之间的JNI抽象中间层，可以直接复用Java SDK，规避了对BE代码的侵入以及使用C++访问大数据存储的诸多不便

这一块属于时实际元数据文件读取的Java侧代码

java 复制代码

public int getNext() throws IOException {
    try (ThreadContextClassLoader ignored = new ThreadContextClassLoader(classLoader)) {
        int numRows = 0;
        for (; numRows < getTableSize(); numRows++) {
            if (!reader.hasNext()) {
                break;
            }
            ContentFile<?> file = reader.next();
            for (int i = 0; i < requiredFields.length; i++) {
                Object fieldData = get(requiredFields[i], file);
                if (fieldData == null) {
                    appendData(i, null);
                } else {
                    ColumnValue fieldValue = new IcebergMetadataColumnValue(fieldData);
                    appendData(i, fieldValue);
                }
            }
        }
        return numRows;
    } catch (Exception e) {
        close();
        LOG.error("Failed to get the next off-heap table chunk of iceberg metadata.", e);
        throw new IOException("Failed to get the next off-heap table chunk of iceberg metadata.", e);
    }
}

这一块目前没有找到Java侧的上层调用，应该在C++中调用，如下，其构造类是在C++中的

cpp 复制代码

// ---------------iceberg metadata jni scanner------------------
std::unique_ptr<JniScanner> create_iceberg_metadata_jni_scanner(const JniScanner::CreateOptions& options) {
    const auto& scan_range = *(options.scan_range);
    ;

    const auto* hdfs_table = dynamic_cast<const IcebergMetadataTableDescriptor*>(options.hive_table);
    std::map<std::string, std::string> jni_scanner_params;

    jni_scanner_params["required_fields"] = hdfs_table->get_hive_column_names();
    jni_scanner_params["metadata_column_types"] = hdfs_table->get_hive_column_types();
    jni_scanner_params["serialized_predicate"] = options.scan_node->serialized_predicate;

    jni_scanner_params["serialized_table"] = options.scan_node->serialized_table;
    jni_scanner_params["split_info"] = scan_range.serialized_split;
    jni_scanner_params["load_column_stats"] = options.scan_node->load_column_stats ? "true" : "false";

    const std::string scanner_factory_class = "com/starrocks/connector/iceberg/IcebergMetadataScannerFactory";
    return std::make_unique<JniScanner>(scanner_factory_class, jni_scanner_params);
}

6. 元数据收集任务

https://github.com/StarRocks/starrocks/pull/44679/files

核心类：IcebergMetadataCollectJob，Iceberg元数据的收集类，实现自MetadataCollectJob，目前看就是通过执行SQL语句，从前文的LogicalIcebergMetadataTable表当中获取数据

这一块属于最终的元数据收集

java 复制代码

private static final String ICEBERG_METADATA_TEMPLATE = "SELECT content" + // INTEGER
        ", file_path" + // VARCHAR
        ", file_format" + // VARCHAR
        ", spec_id" + // INTEGER
        ", partition_data" + // BINARY
        ", record_count" + // BIGINT
        ", file_size_in_bytes" + // BIGINT
        ", split_offsets" + // ARRAY<BIGINT>
        ", sort_id" + // INTEGER
        ", equality_ids" + // ARRAY<INTEGER>
        ", file_sequence_number" + // BIGINT
        ", data_sequence_number " + // BIGINT
        ", column_stats " + // BINARY
        ", key_metadata " + // BINARY
        "FROM `$catalogName`.`$dbName`.`$tableName$logical_iceberg_metadata` " +
        "FOR VERSION AS OF $snapshotId " +
        "WHERE $predicate'";

7. 流程梳理

1. IcebergMetadataCollectJob的调用

IcebergMetadataCollectJob -> StarRocksIcebergTableScan.planFileTasksRemotely -> StarRocksIcebergTableScan.doPlanFiles -> 由Iceberg定义的TableScan流程触发

2. StarRocksIcebergTableScan的构建

StarRocksIcebergTableScan -> IcebergCatalog.getTableScan -> IcebergMetadata.collectTableStatisticsAndCacheIcebergSplit -> prepareMetadata()和triggerIcebergPlanFilesIfNeeded()

prepareMetadata()线路由PrepareCollectMetaTask任务触发，其执行逻辑中调用了prepareMetadata()接口。PrepareCollectMetaTask是OptimizerTask的子类，属于StarRocks优化器的一环，在Optimizer类执行优化的时候会。这一块属于CBO优化，默认是false，没找到设置成true的地方，目前应该没有启用

triggerIcebergPlanFilesIfNeeded()路线有几个调用的地方，主路线应该是getRemoteFileInfos()，其他两个看内容属于统计信息之类的信息收集

IcebergMetadata.getRemoteFileInfos -> IcebergScanNode.setupScanRangeLocations -> PlanFragmentBuilder.visitPhysicalIcebergScan -> PhysicalIcebergScanOperator

这一条调用链最终源头到PhysicalIcebergScanOperator，这个应当是IcebergScanNode经过SQL计划转换后的实际执行节点类

3. 元数据扫描

IcebergMetaSpec -> IcebergMetadata.getSerializedMetaSpec -> MetadataMgr.getSerializedMetaSpec -> IcebergMetadataScanNode.setupScanRangeLocations -> PlanFragmentBuilder.visitPhysicalIcebergMetadataScan -> PhysicalIcebergMetadataScanOperator

元数据扫描这一块源头最终走到PhysicalIcebergMetadataScanOperator，也就是IcebergMetadataScanNode对应的执行类

4. 元数据扫描和数据扫描的逻辑关系

目前整体流程在最上层就差PhysicalIcebergMetadataScanOperator和PhysicalIcebergScanOperator的逻辑关系，这个逻辑在StarRocks的SQL到执行计划的转换过程当中

往上追踪到BackendSelectorFactory，注意这里有两个扫描节点的分配策略：LocalFragmentAssignmentStrategy、RemoteFragmentAssignmentStrategy。根据类的说明，最左节点为scanNode的时候，使用LocalFragmentAssignmentStrategy，它首先将扫描范围分配给 worker，然后将分配给每个 worker 的扫描范围分派给片段实例

在LocalFragmentAssignmentStrategy的assignFragmentToWorker当中可以看到入参包含很多scanNode，追踪上层到CoordinatorPreprocessor，scanNode的来源是StarRocks的DAG图。这之后的源头就涉及到任务解析和DAG图的顺序构建，应当是先扫描元数据再扫描数据这样构建

java 复制代码

for (ExecutionFragment execFragment : executionDAG.getFragmentsInPostorder()) {
    fragmentAssignmentStrategyFactory.create(execFragment, workerProvider).assignFragmentToWorker(execFragment);
}

8. 代码解析

1. 元数据扫描

LogicalIcebergMetadataTable

首先从PhysicalIcebergMetadataScanOperator出发，访问者模式调用接口accept，走到PlanFragmentBuilder.visitPhysicalIcebergMetadataScan

这里首先跟LogicalIcebergMetadataTable关联了起来，这里PhysicalIcebergMetadataScanOperator里包含的表是LogicalIcebergMetadataTable表

LogicalIcebergMetadataTable的初始创建根据调用链追踪应当由CatalogMgr.createCatalog触发

java 复制代码

PhysicalIcebergMetadataScanOperator node = (PhysicalIcebergMetadataScanOperator) optExpression.getOp();

LogicalIcebergMetadataTable table = (LogicalIcebergMetadataTable) node.getTable();

IcebergMetadataScanNode

中间经历一些列的设置，之后构建了IcebergMetadataScanNode

java 复制代码

IcebergMetadataScanNode metadataScanNode =
        new IcebergMetadataScanNode(context.getNextNodeId(), tupleDescriptor,
                "IcebergMetadataScanNode", node.getTemporalClause());

构建之后调用了setupScanRangeLocations，走到了IcebergMetadataScanNode的类逻辑，首先获取元数据文件的分片信息

java 复制代码

IcebergMetaSpec serializedMetaSpec = GlobalStateMgr.getCurrentState().getMetadataMgr()
        .getSerializedMetaSpec(catalogName, originDbName, originTableName, snapshotId, icebergPredicate).cast();

IcebergMetadata

这段逻辑跟IcebergMetadata关联了起来，调用其getSerializedMetaSpec接口，接口中就是获取Iceberg的元数据文件，中间经历了一定的过滤

java 复制代码

List<ManifestFile> dataManifests = snapshot.dataManifests(nativeTable.io());

List<ManifestFile> matchingDataManifests = filterManifests(dataManifests, nativeTable, predicate);
for (ManifestFile file : matchingDataManifests) {
    remoteMetaSplits.add(IcebergMetaSplit.from(file));
}

获取分片之后就是按StarRocks的扫描结构组装TScanRangeLocations，最终在实际执行时分布式分配解析

java 复制代码

private void addSplitScanRangeLocations(RemoteMetaSplit split) {
    TScanRangeLocations scanRangeLocations = new TScanRangeLocations();

    THdfsScanRange hdfsScanRange = new THdfsScanRange();
    hdfsScanRange.setUse_iceberg_jni_metadata_reader(true);

    hdfsScanRange.setSerialized_split(split.getSerializeSplit());
    hdfsScanRange.setFile_length(split.length());
    hdfsScanRange.setLength(split.length());

    // for distributed scheduler
    hdfsScanRange.setFull_path(split.path());
    hdfsScanRange.setOffset(0);

    TScanRange scanRange = new TScanRange();
    scanRange.setHdfs_scan_range(hdfsScanRange);
    scanRangeLocations.setScan_range(scanRange);

    TScanRangeLocation scanRangeLocation = new TScanRangeLocation(new TNetworkAddress("-1", -1));
    scanRangeLocations.addToLocations(scanRangeLocation);

    result.add(scanRangeLocations);
}

PlanFragment

visitPhysicalIcebergMetadataScan接口最终组装的是一个PlanFragment，这大体类似于Spark的stage，是物理执行计划的计划块

java 复制代码

PlanFragment fragment =
        new PlanFragment(context.getNextFragmentId(), metadataScanNode, DataPartition.RANDOM);
context.getFragments().add(fragment);
return fragment

IcebergMetadataScanner

IcebergMetadataScanner由于其调用逻辑来自于C++的代码，暂未梳理其逻辑，但是假定其执行了，可以看其效果，主要在getNext()接口中读取数据

可以看到其读取后的数据结构是ContentFile，是Iceberg中DataFile的上层父类

java 复制代码

ContentFile<?> file = reader.next();
for (int i = 0; i < requiredFields.length; i++) {
    Object fieldData = get(requiredFields[i], file);
    if (fieldData == null) {
        appendData(i, null);
    } else {
        ColumnValue fieldValue = new IcebergMetadataColumnValue(fieldData);
        appendData(i, fieldValue);
    }
}

主要在appendData接口当中，向表添加数据，可以看到这里设置了一个offHeapTable

offHeapTable是 StarRocks 中的一个特殊表类型，简单来说就是在堆外内存中建立一个表结构，将数据对应存储到堆外内存，之后可以以表形式去访问

java 复制代码

protected void appendData(int index, ColumnValue value) {
    offHeapTable.appendData(index, value);
}

2. 数据扫描中的元数据解析

首先同样到PlanFragmentBuilder.visitPhysicalIcebergScan，流程与visitPhysicalIcebergMetadataScan类似

首先是这里的表是数据表

java 复制代码

Table referenceTable = node.getTable();
context.getDescTbl().addReferencedTable(referenceTable);
TupleDescriptor tupleDescriptor = context.getDescTbl().createTupleDescriptor();
tupleDescriptor.setTable(referenceTable);

// set slot
prepareContextSlots(node, context, tupleDescriptor);

之后是IcebergScanNode

java 复制代码

IcebergScanNode icebergScanNode =
        new IcebergScanNode(context.getNextNodeId(), tupleDescriptor, "IcebergScanNode",
                equalityDeleteTupleDesc);

IcebergScanNode这里核心是调用setupScanRangeLocations

java 复制代码

icebergScanNode.setupScanRangeLocations(context.getDescTbl());

最终同样封装成PlanFragment

java 复制代码

PlanFragment fragment =
        new PlanFragment(context.getNextFragmentId(), icebergScanNode, DataPartition.RANDOM);
context.getFragments().add(fragment);
return fragment;

IcebergScanNode

在setupScanRangeLocations当中，有一个操作是getRemoteFileInfos，这个就是获取数据文件信息，因此内部包含了元数据解析的部分

java 复制代码

List<RemoteFileInfo> splits = GlobalStateMgr.getCurrentState().getMetadataMgr().getRemoteFileInfos(
        catalogName, icebergTable, null, snapshotId, predicate, null, -1);

IcebergMetadata

getRemoteFileInfos是在IcebergMetadata当中，会调用triggerIcebergPlanFilesIfNeeded，看接口名字可以明确这是用来触发Iceberg的元数据解析的，最终走到了collectTableStatisticsAndCacheIcebergSplit

java 复制代码

private void triggerIcebergPlanFilesIfNeeded(IcebergFilter key, IcebergTable table, ScalarOperator predicate,
                                             long limit, Tracers tracers, ConnectContext connectContext) {
    if (!scannedTables.contains(key)) {
        tracers = tracers == null ? Tracers.get() : tracers;
        try (Timer ignored = Tracers.watchScope(tracers, EXTERNAL, "ICEBERG.processSplit." + key)) {
            collectTableStatisticsAndCacheIcebergSplit(table, predicate, limit, tracers, connectContext);
        }
    }
}

collectTableStatisticsAndCacheIcebergSplit当中获取了TableScan，这里的Scan就是StarRocksIcebergTableScan

java 复制代码

TableScan scan = icebergCatalog.getTableScan(nativeTbl, new StarRocksIcebergTableScanContext(
        catalogName, dbName, tableName, planMode(connectContext), connectContext))
        .useSnapshot(snapshotId)
        .metricsReporter(metricsReporter)
        .planWith(jobPlanningExecutor);

StarRocksIcebergTableScan

之后走scan.planFiles()，这个中间会基于Iceberg的逻辑进行调用

java 复制代码

CloseableIterable<FileScanTask> fileScanTaskIterable = TableScanUtil.splitFiles(
        scan.planFiles(), scan.targetSplitSize());

Icberg的逻辑中planFiles最终会调用TableScan的doPlanFiles，这里调用的就是StarRocksIcebergTableScan的实现接口，根据场景有本地和远程的调用方式

java 复制代码

if (shouldPlanLocally(dataManifests, loadColumnStats)) {
    return planFileTasksLocally(dataManifests, deleteManifests);
} else {
    return planFileTasksRemotely(dataManifests, deleteManifests);
}

Iceberg应当是使用的planFileTasksRemotely，内部会构建IcebergMetadataCollectJob

java 复制代码

MetadataCollectJob metadataCollectJob = new IcebergMetadataCollectJob(
        catalogName, dbName, tableName, TResultSinkType.METADATA_ICEBERG, snapshotId(), icebergSerializedPredicate);

metadataCollectJob.init(connectContext.getSessionVariable());

long currentTimestamp = System.currentTimeMillis();
String threadNamePrefix = String.format("%s-%s-%s-%d", catalogName, dbName, tableName, currentTimestamp);
executeInNewThread(threadNamePrefix + "-fetch_result", metadataCollectJob::asyncCollectMetadata);

MetadataExecutor执行

IcebergMetadataCollectJob的执行在MetadataExecutor当中，就是基本的SQL执行，这里是异步的

java 复制代码

public void asyncExecuteSQL(MetadataCollectJob job) {
    ConnectContext context = job.getContext();
    context.setThreadLocalInfo();
    String sql = job.getSql();
    ExecPlan execPlan;
    StatementBase parsedStmt;
    try {
        parsedStmt = SqlParser.parseOneWithStarRocksDialect(sql, context.getSessionVariable());
        execPlan = StatementPlanner.plan(parsedStmt, context, job.getSinkType());
    } catch (Exception e) {
        context.getState().setError(e.getMessage());
        return;
    }

    this.executor = new StmtExecutor(context, parsedStmt);
    context.setExecutor(executor);
    context.setQueryId(UUIDUtil.genUUID());
    context.getSessionVariable().setEnableMaterializedViewRewrite(false);

    LOG.info("Start to execute metadata collect job on {}.{}.{}", job.getCatalogName(), job.getDbName(), job.getTableName());
    executor.executeStmtWithResultQueue(context, execPlan, job.getResultQueue());
}