1
0

[HUDI-1138] Add timeline-server-based marker file strategy for improving marker-related latency (#3233)

- Can be enabled for cloud stores like S3. Not supported for hdfs yet, due to partial write failures.
This commit is contained in:
Y Ethan Guo
2021-08-11 08:48:13 -07:00
committed by GitHub
parent 29332498af
commit 4783176554
52 changed files with 2144 additions and 353 deletions

View File

@@ -58,9 +58,9 @@ import org.apache.hudi.table.BulkInsertPartitioner;
import org.apache.hudi.table.HoodieFlinkTable;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.HoodieTimelineArchiveLog;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.action.HoodieWriteMetadata;
import org.apache.hudi.table.action.compact.FlinkCompactHelpers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hudi.table.upgrade.FlinkUpgradeDowngrade;
import org.apache.hudi.util.FlinkClientUtil;
@@ -316,7 +316,7 @@ public class HoodieFlinkWriteClient<T extends HoodieRecordPayload> extends
Option<Map<String, String>> extraMetadata) {
try {
// Delete the marker directory for the instant.
new MarkerFiles(createTable(config, hadoopConf), instantTime)
WriteMarkersFactory.get(config.getMarkersType(), createTable(config, hadoopConf), instantTime)
.quietDeleteMarkerDir(context, config.getMarkersDeleteParallelism());
// We cannot have unbounded commit files. Archive commits if we have to archive
HoodieTimelineArchiveLog archiveLog = new HoodieTimelineArchiveLog(config, table);

View File

@@ -24,7 +24,8 @@ import org.apache.hudi.common.model.HoodieRecord;
import org.apache.hudi.common.model.HoodieRecordPayload;
import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hadoop.fs.Path;
import org.slf4j.Logger;
@@ -49,6 +50,7 @@ public class FlinkAppendHandle<T extends HoodieRecordPayload, I, K, O>
private static final Logger LOG = LoggerFactory.getLogger(FlinkAppendHandle.class);
private boolean isClosed = false;
private final WriteMarkers writeMarkers;
public FlinkAppendHandle(
HoodieWriteConfig config,
@@ -59,6 +61,7 @@ public class FlinkAppendHandle<T extends HoodieRecordPayload, I, K, O>
Iterator<HoodieRecord<T>> recordItr,
TaskContextSupplier taskContextSupplier) {
super(config, instantTime, hoodieTable, partitionPath, fileId, recordItr, taskContextSupplier);
this.writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), hoodieTable, instantTime);
}
@Override
@@ -66,10 +69,9 @@ public class FlinkAppendHandle<T extends HoodieRecordPayload, I, K, O>
// In some rare cases, the task was pulled up again with same write file name,
// for e.g, reuse the small log files from last commit instant.
// Just skip the marker file creation if it already exists, the new data would append to
// Just skip the marker creation if it already exists, the new data would append to
// the file directly.
MarkerFiles markerFiles = new MarkerFiles(hoodieTable, instantTime);
markerFiles.createIfNotExists(partitionPath, dataFileName, getIOType());
writeMarkers.createIfNotExists(partitionPath, dataFileName, getIOType());
}
@Override

View File

@@ -27,7 +27,8 @@ import org.apache.hudi.common.util.Option;
import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieException;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.avro.Schema;
import org.apache.hadoop.fs.Path;
@@ -102,8 +103,8 @@ public class FlinkCreateHandle<T extends HoodieRecordPayload, I, K, O>
@Override
protected void createMarkerFile(String partitionPath, String dataFileName) {
MarkerFiles markerFiles = new MarkerFiles(hoodieTable, instantTime);
markerFiles.createIfNotExists(partitionPath, dataFileName, getIOType());
WriteMarkers writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), hoodieTable, instantTime);
writeMarkers.createIfNotExists(partitionPath, dataFileName, getIOType());
}
@Override

View File

@@ -29,7 +29,8 @@ import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieException;
import org.apache.hudi.exception.HoodieIOException;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hadoop.fs.Path;
import org.apache.log4j.LogManager;
@@ -104,8 +105,8 @@ public class FlinkMergeAndReplaceHandle<T extends HoodieRecordPayload, I, K, O>
@Override
protected void createMarkerFile(String partitionPath, String dataFileName) {
MarkerFiles markerFiles = new MarkerFiles(hoodieTable, instantTime);
markerFiles.createIfNotExists(partitionPath, dataFileName, getIOType());
WriteMarkers writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), hoodieTable, instantTime);
writeMarkers.createIfNotExists(partitionPath, dataFileName, getIOType());
}
@Override

View File

@@ -28,7 +28,8 @@ import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieException;
import org.apache.hudi.exception.HoodieIOException;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hadoop.fs.Path;
import org.apache.log4j.LogManager;
@@ -116,8 +117,8 @@ public class FlinkMergeHandle<T extends HoodieRecordPayload, I, K, O>
@Override
protected void createMarkerFile(String partitionPath, String dataFileName) {
MarkerFiles markerFiles = new MarkerFiles(hoodieTable, instantTime);
markerFiles.createIfNotExists(partitionPath, dataFileName, getIOType());
WriteMarkers writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), hoodieTable, instantTime);
writeMarkers.createIfNotExists(partitionPath, dataFileName, getIOType());
}
@Override

View File

@@ -31,7 +31,8 @@ import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieIOException;
import org.apache.hudi.exception.HoodieInsertException;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.flink.table.data.RowData;
import org.apache.flink.table.types.logical.RowType;
@@ -188,8 +189,8 @@ public class HoodieRowDataCreateHandle implements Serializable {
* @param partitionPath Partition path
*/
private void createMarkerFile(String partitionPath, String dataFileName) {
MarkerFiles markerFiles = new MarkerFiles(table, instantTime);
markerFiles.create(partitionPath, dataFileName, IOType.CREATE);
WriteMarkers writeMarkers = WriteMarkersFactory.get(writeConfig.getMarkersType(), table, instantTime);
writeMarkers.create(partitionPath, dataFileName, IOType.CREATE);
}
private String getWriteToken() {

View File

@@ -32,9 +32,11 @@ import org.apache.hudi.common.table.timeline.HoodieInstant;
import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieRollbackException;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hadoop.fs.FileStatus;
import java.util.ArrayList;
import java.io.IOException;
import java.util.List;
@@ -52,17 +54,17 @@ public class FlinkMarkerBasedRollbackStrategy<T extends HoodieRecordPayload> ext
@Override
public List<HoodieRollbackStat> execute(HoodieInstant instantToRollback) {
try {
MarkerFiles markerFiles = new MarkerFiles(table, instantToRollback.getTimestamp());
List<HoodieRollbackStat> rollbackStats = context.map(markerFiles.allMarkerFilePaths(), markerFilePath -> {
WriteMarkers writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), table, instantToRollback.getTimestamp());
List<HoodieRollbackStat> rollbackStats = context.map(new ArrayList<>(writeMarkers.allMarkerFilePaths()), markerFilePath -> {
String typeStr = markerFilePath.substring(markerFilePath.lastIndexOf(".") + 1);
IOType type = IOType.valueOf(typeStr);
switch (type) {
case MERGE:
return undoMerge(MarkerFiles.stripMarkerSuffix(markerFilePath));
return undoMerge(WriteMarkers.stripMarkerSuffix(markerFilePath));
case APPEND:
return undoAppend(MarkerFiles.stripMarkerSuffix(markerFilePath), instantToRollback);
return undoAppend(WriteMarkers.stripMarkerSuffix(markerFilePath), instantToRollback);
case CREATE:
return undoCreate(MarkerFiles.stripMarkerSuffix(markerFilePath));
return undoCreate(WriteMarkers.stripMarkerSuffix(markerFilePath));
default:
throw new HoodieRollbackException("Unknown marker type, during rollback of " + instantToRollback);
}

View File

@@ -24,7 +24,8 @@ import org.apache.hudi.common.table.timeline.HoodieInstant;
import org.apache.hudi.common.table.timeline.HoodieTimeline;
import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.table.HoodieFlinkTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import java.util.List;
import java.util.stream.Collectors;
@@ -41,9 +42,9 @@ public class OneToZeroDowngradeHandler implements DowngradeHandler {
HoodieTimeline inflightTimeline = table.getMetaClient().getCommitsTimeline().filterPendingExcludingCompaction();
List<HoodieInstant> commits = inflightTimeline.getReverseOrderedInstants().collect(Collectors.toList());
for (HoodieInstant commitInstant : commits) {
// delete existing marker files
MarkerFiles markerFiles = new MarkerFiles(table, commitInstant.getTimestamp());
markerFiles.quietDeleteMarkerDir(context, config.getMarkersDeleteParallelism());
// delete existing markers
WriteMarkers writeMarkers = WriteMarkersFactory.get(config.getMarkersType(), table, commitInstant.getTimestamp());
writeMarkers.quietDeleteMarkerDir(context, config.getMarkersDeleteParallelism());
}
}
}

View File

@@ -32,10 +32,12 @@ import org.apache.hudi.config.HoodieWriteConfig;
import org.apache.hudi.exception.HoodieRollbackException;
import org.apache.hudi.table.HoodieFlinkTable;
import org.apache.hudi.table.HoodieTable;
import org.apache.hudi.table.MarkerFiles;
import org.apache.hudi.table.action.rollback.ListingBasedRollbackHelper;
import org.apache.hudi.table.action.rollback.ListingBasedRollbackRequest;
import org.apache.hudi.table.action.rollback.RollbackUtils;
import org.apache.hudi.table.marker.WriteMarkers;
import org.apache.hudi.table.marker.WriteMarkersFactory;
import org.apache.hudi.table.marker.MarkerType;
import org.apache.hadoop.fs.FileStatus;
import org.apache.hadoop.fs.Path;
@@ -60,35 +62,37 @@ public class ZeroToOneUpgradeHandler implements UpgradeHandler {
commits.remove(instantTime);
}
for (String commit : commits) {
// for every pending commit, delete old marker files and re-create marker files in new format
recreateMarkerFiles(commit, table, context, config.getMarkersDeleteParallelism());
// for every pending commit, delete old markers and re-create markers in new format
recreateMarkers(commit, table, context, config.getMarkersType(), config.getMarkersDeleteParallelism());
}
}
/**
* Recreate marker files in new format.
* Step1: Delete existing marker files
* Recreate markers in new format.
* Step1: Delete existing markers
* Step2: Collect all rollback file info.
* Step3: recreate marker files for all interested files.
* Step3: recreate markers for all interested files.
*
* @param commitInstantTime instant of interest for which marker files need to be recreated.
* @param commitInstantTime instant of interest for which markers need to be recreated.
* @param table instance of {@link HoodieFlinkTable} to use
* @param context instance of {@link HoodieEngineContext} to use
* @param markerType marker type to use
* @throws HoodieRollbackException on any exception during upgrade.
*/
private static void recreateMarkerFiles(final String commitInstantTime,
HoodieFlinkTable table,
HoodieEngineContext context,
int parallelism) throws HoodieRollbackException {
private static void recreateMarkers(final String commitInstantTime,
HoodieFlinkTable table,
HoodieEngineContext context,
MarkerType markerType,
int parallelism) throws HoodieRollbackException {
try {
// fetch hoodie instant
Option<HoodieInstant> commitInstantOpt = Option.fromJavaOptional(table.getActiveTimeline().getCommitsTimeline().getInstants()
.filter(instant -> HoodieActiveTimeline.EQUALS.test(instant.getTimestamp(), commitInstantTime))
.findFirst());
if (commitInstantOpt.isPresent()) {
// delete existing marker files
MarkerFiles markerFiles = new MarkerFiles(table, commitInstantTime);
markerFiles.quietDeleteMarkerDir(context, parallelism);
// delete existing markers
WriteMarkers writeMarkers = WriteMarkersFactory.get(markerType, table, commitInstantTime);
writeMarkers.quietDeleteMarkerDir(context, parallelism);
// generate rollback stats
List<ListingBasedRollbackRequest> rollbackRequests;
@@ -100,15 +104,15 @@ public class ZeroToOneUpgradeHandler implements UpgradeHandler {
List<HoodieRollbackStat> rollbackStats = new ListingBasedRollbackHelper(table.getMetaClient(), table.getConfig())
.collectRollbackStats(context, commitInstantOpt.get(), rollbackRequests);
// recreate marker files adhering to marker based rollback
// recreate markers adhering to marker based rollback
for (HoodieRollbackStat rollbackStat : rollbackStats) {
for (String path : rollbackStat.getSuccessDeleteFiles()) {
String dataFileName = path.substring(path.lastIndexOf("/") + 1);
// not feasible to differentiate MERGE from CREATE. hence creating with MERGE IOType for all base files.
markerFiles.create(rollbackStat.getPartitionPath(), dataFileName, IOType.MERGE);
writeMarkers.create(rollbackStat.getPartitionPath(), dataFileName, IOType.MERGE);
}
for (FileStatus fileStatus : rollbackStat.getCommandBlocksCount().keySet()) {
markerFiles.create(rollbackStat.getPartitionPath(), getFileNameForMarkerFromLogFile(fileStatus.getPath().toString(), table), IOType.APPEND);
writeMarkers.create(rollbackStat.getPartitionPath(), getFileNameForMarkerFromLogFile(fileStatus.getPath().toString(), table), IOType.APPEND);
}
}
}