xushiyan commented on code in PR #18040:
URL: https://github.com/apache/hudi/pull/18040#discussion_r2749138774
##########
hudi-flink-datasource/hudi-flink/src/main/java/org/apache/hudi/source/reader/HoodieRecordEmitter.java:
##########
@@ -18,19 +18,75 @@
package org.apache.hudi.source.reader;
+import org.apache.flink.api.common.eventtime.Watermark;
import org.apache.flink.api.connector.source.SourceOutput;
import org.apache.flink.connector.base.source.reader.RecordEmitter;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import org.apache.hudi.common.fs.FSUtils;
import org.apache.hudi.source.split.HoodieSourceSplit;
/**
* Default Hoodie record emitter.
- * @param <T>
+ *
+ * <p>This emitter handles watermark emission based on split information.
+ *
+ * @param <T> The type of records to emit
*/
public class HoodieRecordEmitter<T> implements
RecordEmitter<HoodieRecordWithPosition<T>, T, HoodieSourceSplit> {
+ private static final Logger LOG =
LoggerFactory.getLogger(HoodieRecordEmitter.class);
+ private HoodieSourceSplit lastSplit = null;
+ private long watermark = Long.MIN_VALUE;
@Override
public void emitRecord(HoodieRecordWithPosition<T> record, SourceOutput<T>
output, HoodieSourceSplit split) throws Exception {
+ if (lastSplit == null || !split.splitId().equals(lastSplit.splitId())) {
+ long newWatermark = extractWatermark(split);
+ if (newWatermark < watermark) {
+ LOG.warn(
+ "Received a new split with lower watermark. Previous watermark =
{}, current watermark = {}, previous split = {}, current split = {}",
+ watermark,
+ newWatermark,
+ lastSplit,
+ split);
+ } else {
+ watermark = newWatermark;
+ output.emitWatermark(new Watermark(watermark));
+ LOG.debug("Watermark = {} emitted based on split = {}", watermark,
split);
+ }
+
+ lastSplit = split;
+ }
+
output.collect(record.record());
split.updatePosition(record.fileOffset(), record.recordOffset());
}
+
+ private long extractWatermark(HoodieSourceSplit split) {
+ long maxInstantTime = Long.MIN_VALUE;
+
+ if (split.getBasePath().isPresent()) {
+ String basePath = split.getBasePath().get();
+ try {
+ long baseCommitTime = Long.parseLong(FSUtils.getCommitTime(basePath));
+ maxInstantTime = Math.max(baseCommitTime, maxInstantTime);
+ } catch (NumberFormatException e) {
+ LOG.warn("Failed to parse commit time from basePath: {}", basePath, e);
+ }
+ }
+
+ if (split.getLogPaths().isPresent()) {
+ for (String logPath : split.getLogPaths().get()) {
+ try {
+ long logCommitTime = Long.parseLong(FSUtils.getCommitTime(logPath));
+ maxInstantTime = Math.max(logCommitTime, maxInstantTime);
+ } catch (NumberFormatException e) {
+ LOG.warn("Failed to parse commit time from logPath: {}", logPath, e);
+ }
+ }
+ }
+
+ return maxInstantTime;
Review Comment:
we discussed and clarified that file-slice specific timestamp is more
accurate watermark. But it only aligns with commit time ordering. When table is
set to event time ordering or with custom merger, aligning with ordering
semantics would be tricky. Keeping this discussion open.
--
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]
