// 添加 Writer 算子，有并行度
SingleOutputStreamOperator<WriteResult> writerStream =
appendWriter(distributeStream, flinkRowType, equalityFieldIds);
// 添加 Commit 算子，并行度固定为 1
SingleOutputStreamOperator<Void> committerStream = appendCommitter(writerStream);
// 添加 sink
return appendDummySink(committerStream);
}
### appendWriter 方法
private SingleOutputStreamOperator appendWriter(
DataStream input, RowType flinkRowType, List equalityFieldIds) {
....
if (flinkWriteConf.upsertMode()) {
if (!table.spec().isUnpartitioned()) {
// 在 upser 模式下检查分区建必须在 equalityFieldIds 中
for (PartitionField partitionField : table.spec().fields()) {
Preconditions.checkState(
equalityFieldIds.contains(partitionField.sourceId()),
"In UPSERT mode, partition field '%s' should be included in equality fields: '%s'",
partitionField,
equalityFieldColumns);
}
}
}
// 创建 streamWriter
IcebergStreamWriter<RowData> streamWriter =
createStreamWriter(table, flinkWriteConf, flinkRowType, equalityFieldIds);
// 设置并行度 如果没有指定则和输入流的并行度一样
int parallelism =
flinkWriteConf.writeParallelism() == null
? input.getParallelism()
: flinkWriteConf.writeParallelism();
....
return writerStream;
}
### createStreamWriter 方法
static IcebergStreamWriter createStreamWriter(
Table table,
FlinkWriteConf flinkWriteConf,
RowType flinkRowType,
List equalityFieldIds) {
...
Table serializableTable = SerializableTable.copyOf(table);
FileFormat format = flinkWriteConf.dataFileFormat();
// 创建 TaskWriterFactory 根据 表的 Schema 创建对应的 Writer
TaskWriterFactory<RowData> taskWriterFactory =
new RowDataTaskWriterFactory(
serializableTable,
flinkRowType,
flinkWriteConf.targetDataFileSize(),
format,
writeProperties(table, format, flinkWriteConf),
equalityFieldIds,
flinkWriteConf.upsertMode());
// 新建 IcebergStreamWriter
return new IcebergStreamWriter<>(table.name(), taskWriterFactory);
}
### IcebergStreamWriter 类
该类为一个 Flink 内部的 OneInputStreamOperator 类，拥有 Flink 算子相关特性
class IcebergStreamWriter extends AbstractStreamOperator
implements OneInputStreamOperator<T, WriteResult>, BoundedOneInput {
...
@Override
public void open() {
...
// 初始化相关监控类
this.writerMetrics = new IcebergStreamWriterMetrics(super.metrics, fullTableName);
// 初始化 taskWriterFactory 用于创建 writer
this.taskWriterFactory.initialize(subTaskId, attemptId);
// 创建 writer
// 主要分成四类
// 根据 Iceberg 表是否有分区和开启Upsert模式
// UnpartitionedWriter ： 无分区 insert only
// RowDataPartitionedFanoutWriter ： 分区 insert only
// UnpartitionedDeltaWriter ：无分区 Upsert
// PartitionedDeltaWriter ：有分区 Upsert
this.writer = taskWriterFactory.create();
}
@Override
public void processElement(StreamRecord element) throws Exception {
// 处理数据写入
writer.write(element.getValue());
}
...
// 将本次写入数据文件下发至 Commit 进行统一提交
private void flush() throws IOException {
if (writer == null) {
return;
}
long startNano = System.nanoTime();
WriteResult result = writer.complete();
writerMetrics.updateFlushResult(result);
output.collect(new StreamRecord<>(result));
writerMetrics.flushDuration(TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - startNano));
writer = null;
}
}
### IcebergFilesCommitter 类
class IcebergFilesCommitter extends AbstractStreamOperator
implements OneInputStreamOperator<WriteResult, Void>, BoundedOneInput {
...
@Override
public void initializeState(StateInitializationContext context) throws Exception {
// 最大连续空提交
// 在间断指定次数 Checkpoint 都没有数据后才真正触发 Commit，生成 Snapshot。
// 减少空快照生成
maxContinuousEmptyCommits =
PropertyUtil.propertyAsInt(table.properties(), MAX\_CONTINUOUS\_EMPTY\_COMMITS, 10);
// 创建 文件输出 OutputFileFactory
this.manifestOutputFileFactory =
FlinkManifestUtil.createOutputFileFactory(
table, flinkJobId, operatorUniqueId, subTaskId, attemptId);
if (context.isRestored()) {
...
// 从状态中恢复未提交的数据文件
NavigableMap<Long, byte[]> uncommittedDataFiles =
Maps.newTreeMap(checkpointsState.get().iterator().next())
.tailMap(maxCommittedCheckpointId, false);
if (!uncommittedDataFiles.isEmpty()) {
// Committed all uncommitted data files from the old flink job to iceberg table.
long maxUncommittedCheckpointId = uncommittedDataFiles.lastKey();
// 如果存在未提交的文件 进行提交
commitUpToCheckpoint(
uncommittedDataFiles, restoredFlinkJobId, operatorUniqueId, maxUncommittedCheckpointId);
}
}
}
@Override
public void snapshotState(StateSnapshotContext context) throws Exception {
// 将 checkpointId 对应的写入完成的DATA-FILE生成清单文件并放入 dataFilesPerCheckpoint
dataFilesPerCheckpoint.put(checkpointId, writeToManifest(checkpointId));
// Reset the snapshot state to the latest state.
checkpointsState.clear();
// 存入状态
checkpointsState.add(dataFilesPerCheckpoint);
jobIdState.clear();
jobIdState.add(flinkJobId);
// Clear the local buffer for current checkpoint.
writeResultsOfCurrentCkpt.clear();
committerMetrics.checkpointDuration(
TimeUnit.NANOSECONDS.toMillis(System.nanoTime() - startNano));
}
@Override
public void notifyCheckpointComplete(long checkpointId) throws Exception {
...
if (checkpointId > maxCommittedCheckpointId) {
LOG.info("Checkpoint {} completed. Attempting commit.", checkpointId);
// 完成 checkpoint 对数据进行 COMMIT
commitUpToCheckpoint(dataFilesPerCheckpoint, flinkJobId, operatorUniqueId, checkpointId);
this.maxCommittedCheckpointId = checkpointId;
} else {
"Skipping committing checkpoint {}. {} is already committed.",
checkpointId,
maxCommittedCheckpointId);
}
}
private void commitUpToCheckpoint(
NavigableMap<Long, byte[]> deltaManifestsMap,
String newFlinkJobId,
String operatorId,
long checkpointId)
throws IOException {
NavigableMap<Long, byte[]> pendingMap =
// 获取等待提交的数据文件
deltaManifestsMap.headMap(checkpointId, true);
List manifests = Lists.newArrayList();
NavigableMap<Long, WriteResult> pendingResults = Maps.newTreeMap();
for (Map.Entry<Long, byte[]> e : pendingMap.entrySet()) {
// 数据文件为空则跳过
if (Arrays.equals(EMPTY_MANIFEST_DATA, e.getValue())) {
continue;
}
DeltaManifests deltaManifests =
SimpleVersionedSerialization.readVersionAndDeSerialize(
DeltaManifestsSerializer.INSTANCE, e.getValue());
pendingResults.put(
e.getKey(),
FlinkManifestUtil.readCompletedFiles(deltaManifests, table.io(), table.specs()));
manifests.addAll(deltaManifests.manifests());
}
// 获取当前待提交文件的 数据条数及数据文件大小
CommitSummary summary = new CommitSummary(pendingResults);
// 提交数据
commitPendingResult(pendingResults, summary, newFlinkJobId, operatorId, checkpointId);
committerMetrics.updateCommitSummary(summary);
pendingMap.clear();
// 清除已提交数据
deleteCommittedManifests(manifests, newFlinkJobId, checkpointId);
}
private void commitPendingResult(
NavigableMap<Long, WriteResult> pendingResults,
CommitSummary summary,
String newFlinkJobId,
String operatorId,
long checkpointId) {
...
continuousEmptyCheckpoints = totalFiles == 0 ? continuousEmptyCheckpoints + 1 : 0;
// 数据文件不问 0 或者 连续最大空提交到达了配置的参数阈值触发提交
if (totalFiles != 0 || continuousEmptyCheckpoints % maxContinuousEmptyCommits == 0) {
if (replacePartitions) {
// replace 提交
// 使用 newReplacePartitions()
replacePartitions(pendingResults, summary, newFlinkJobId, operatorId, checkpointId);
} else {
// 普通提交
// 使用 newAppend()
commitDeltaTxn(pendingResults, summary, newFlinkJobId, operatorId, checkpointId);
}
continuousEmptyCheckpoints = 0;
...
}
...
private void replacePartitions(
NavigableMap<Long, WriteResult> pendingResults,
CommitSummary summary,
String newFlinkJobId,
String operatorId,
long checkpointId) {
Preconditions.checkState(
summary.deleteFilesCount() == 0, "Cannot overwrite partitions with delete files.");
// 使用 newReplacePartitions 提交
ReplacePartitions dynamicOverwrite = table.newReplacePartitions().scanManifestsWith(workerPool);
for (WriteResult result : pendingResults.values()) {
Preconditions.checkState(
result.referencedDataFiles().length == 0, "Should have no referenced data files.");
Arrays.stream(result.dataFiles()).forEach(dynamicOverwrite::addFile);
}
...
}
private void commitDeltaTxn(
NavigableMap<Long, WriteResult> pendingResults,
CommitSummary summary,
String newFlinkJobId,
String operatorId,
long checkpointId) {
if (summary.deleteFilesCount() == 0) {
// To be compatible with iceberg format V1.
AppendFiles appendFiles = table.newAppend().scanManifestsWith(workerPool);
for (WriteResult result : pendingResults.values()) {
Preconditions.checkState(
result.referencedDataFiles().length == 0,
"Should have no referenced data files for append.");
Arrays.stream(result.dataFiles()).forEach(appendFiles::appendFile);
}
commitOperation(appendFiles, summary, "append", newFlinkJobId, operatorId, checkpointId);
} else {
// To be compatible with iceberg format V2.
for (Map.Entry<Long, WriteResult> e : pendingResults.entrySet()) {
// We don't commit the merged result into a single transaction because for the sequential
// transaction txn1 and txn2, the equality-delete files of txn2 are required to be applied
// to data files from txn1. Committing the merged one will lead to the incorrect delete
// semantic.
WriteResult result = e.getValue();
RowDelta rowDelta = table.newRowDelta().scanManifestsWith(workerPool);
// 分别写入 DataFile 和 Delete File
Arrays.stream(result.dataFiles()).forEach(rowDelta::addRows);
Arrays.stream(result.deleteFiles()).forEach(rowDelta::addDeletes);
commitOperation(rowDelta, summary, "rowDelta", newFlinkJobId, operatorId, e.getKey());
}
}
}
private void commitOperation(
SnapshotUpdate<?> operation,
CommitSummary summary,
String description,
String newFlinkJobId,
String operatorId,
long checkpointId) {
...
// 提交操作
operation.commit(); // abort is automatically called if this fails.
...
committerMetrics.commitDuration(durationMs);
}
@Override
public void processElement(StreamRecord element) throws Exception{
final WriteResult value = element.getValue();
if ("DDL".equalsIgnoreCase(value.getType())) {
this.writeResultsOfCurrentCkpt.add(element.getValue());
doCommit();
} else {
this.writeResultsOfCurrentCkpt.add(element.getValue());
}
}
...
}