yihua commented on code in PR #11947:
URL: https://github.com/apache/hudi/pull/11947#discussion_r1797491098
##########
hudi-utilities/src/main/java/org/apache/hudi/utilities/sources/HoodieIncrSource.java:
##########
@@ -212,35 +257,43 @@ public Pair<Option<Dataset<Row>>, String>
fetchNextBatch(Option<String> lastCkpt
.collect(Collectors.toMap(Pair::getLeft, Pair::getRight));
reader = reader.options(optionsMap);
}
+ boolean shouldFullScan =
+
queryContext.getActiveTimeline().isBeforeTimelineStartsByCompletionTime(startTime)
+ && missingCheckpointStrategy ==
MissingCheckpointStrategy.READ_UPTO_LATEST_COMMIT;
Dataset<Row> source;
- // Do Incr pull. Set end instant if available
- if (queryInfo.isIncremental()) {
- source = reader
- .options(readOpts)
- .option(QUERY_TYPE().key(), QUERY_TYPE_INCREMENTAL_OPT_VAL())
- .option(BEGIN_INSTANTTIME().key(), queryInfo.getStartInstant())
- .option(END_INSTANTTIME().key(), queryInfo.getEndInstant())
- .option(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- props.getString(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().defaultValue()))
- .option(INCREMENTAL_READ_HANDLE_HOLLOW_COMMIT().key(),
handlingMode.name())
- .load(srcPath);
- } else {
- // if checkpoint is missing from source table, and if strategy is set to
READ_UPTO_LATEST_COMMIT, we have to issue snapshot query
+ if (instantRange.isEmpty() || shouldFullScan) {
+ // snapshot query
Dataset<Row> snapshot = reader
.options(readOpts)
.option(DataSourceReadOptions.QUERY_TYPE().key(),
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL())
.load(srcPath);
if (snapshotLoadQuerySplitter.isPresent()) {
- queryInfo =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryInfo,
sourceProfileSupplier);
+ queryContext =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryContext,
sourceProfileSupplier);
+ // update endTime/next checkpoint
+ endTime =
queryContext.getEndInstant().orElse(queryContext.getLastInstant());
}
+ Set<String> validInstants = new HashSet<>(queryContext.getInstants());
+ snapshot =
queryContext.getPredicateFilter().map(snapshot::filter).orElse(snapshot);
source = snapshot
// add filtering so that only interested records are returned.
- .filter(String.format("%s > '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
- queryInfo.getStartInstant()))
+ // completion time comparison uses ( , ], but when comparing start
time we need to use [, ]
+ .filter(String.format("%s >= '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
+ queryContext.getStartInstant().get()))
.filter(String.format("%s <= '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
- queryInfo.getEndInstant()));
- source =
queryInfo.getPredicateFilter().map(source::filter).orElse(source);
+
queryContext.getEndInstant().orElse(queryContext.getLastInstant())))
+ .filter((Row row) ->
validInstants.contains(row.getAs(HoodieRecord.COMMIT_TIME_METADATA_FIELD)));
Review Comment:
Does this translate into `IN` predicate (`new
Column(HoodieRecord.COMMIT_TIME_METADATA_FIELD).isin(validInstants)`) which can
be pushed down? We should avoid scanning all rows for evaluating this
predicate.
##########
hudi-utilities/src/main/java/org/apache/hudi/utilities/sources/HoodieIncrSource.java:
##########
@@ -212,35 +257,43 @@ public Pair<Option<Dataset<Row>>, String>
fetchNextBatch(Option<String> lastCkpt
.collect(Collectors.toMap(Pair::getLeft, Pair::getRight));
reader = reader.options(optionsMap);
}
+ boolean shouldFullScan =
+
queryContext.getActiveTimeline().isBeforeTimelineStartsByCompletionTime(startTime)
+ && missingCheckpointStrategy ==
MissingCheckpointStrategy.READ_UPTO_LATEST_COMMIT;
Dataset<Row> source;
- // Do Incr pull. Set end instant if available
- if (queryInfo.isIncremental()) {
- source = reader
- .options(readOpts)
- .option(QUERY_TYPE().key(), QUERY_TYPE_INCREMENTAL_OPT_VAL())
- .option(BEGIN_INSTANTTIME().key(), queryInfo.getStartInstant())
- .option(END_INSTANTTIME().key(), queryInfo.getEndInstant())
- .option(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- props.getString(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().defaultValue()))
- .option(INCREMENTAL_READ_HANDLE_HOLLOW_COMMIT().key(),
handlingMode.name())
- .load(srcPath);
- } else {
- // if checkpoint is missing from source table, and if strategy is set to
READ_UPTO_LATEST_COMMIT, we have to issue snapshot query
+ if (instantRange.isEmpty() || shouldFullScan) {
+ // snapshot query
Dataset<Row> snapshot = reader
.options(readOpts)
.option(DataSourceReadOptions.QUERY_TYPE().key(),
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL())
.load(srcPath);
if (snapshotLoadQuerySplitter.isPresent()) {
- queryInfo =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryInfo,
sourceProfileSupplier);
+ queryContext =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryContext,
sourceProfileSupplier);
+ // update endTime/next checkpoint
+ endTime =
queryContext.getEndInstant().orElse(queryContext.getLastInstant());
}
+ Set<String> validInstants = new HashSet<>(queryContext.getInstants());
+ snapshot =
queryContext.getPredicateFilter().map(snapshot::filter).orElse(snapshot);
Review Comment:
It looks like the `predicateFilter` is never set?
##########
hudi-utilities/src/main/java/org/apache/hudi/utilities/sources/HoodieIncrSource.java:
##########
@@ -212,35 +257,43 @@ public Pair<Option<Dataset<Row>>, String>
fetchNextBatch(Option<String> lastCkpt
.collect(Collectors.toMap(Pair::getLeft, Pair::getRight));
reader = reader.options(optionsMap);
}
+ boolean shouldFullScan =
+
queryContext.getActiveTimeline().isBeforeTimelineStartsByCompletionTime(startTime)
+ && missingCheckpointStrategy ==
MissingCheckpointStrategy.READ_UPTO_LATEST_COMMIT;
Dataset<Row> source;
- // Do Incr pull. Set end instant if available
- if (queryInfo.isIncremental()) {
- source = reader
- .options(readOpts)
- .option(QUERY_TYPE().key(), QUERY_TYPE_INCREMENTAL_OPT_VAL())
- .option(BEGIN_INSTANTTIME().key(), queryInfo.getStartInstant())
- .option(END_INSTANTTIME().key(), queryInfo.getEndInstant())
- .option(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- props.getString(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().defaultValue()))
- .option(INCREMENTAL_READ_HANDLE_HOLLOW_COMMIT().key(),
handlingMode.name())
- .load(srcPath);
- } else {
- // if checkpoint is missing from source table, and if strategy is set to
READ_UPTO_LATEST_COMMIT, we have to issue snapshot query
+ if (instantRange.isEmpty() || shouldFullScan) {
+ // snapshot query
Dataset<Row> snapshot = reader
.options(readOpts)
.option(DataSourceReadOptions.QUERY_TYPE().key(),
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL())
.load(srcPath);
if (snapshotLoadQuerySplitter.isPresent()) {
- queryInfo =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryInfo,
sourceProfileSupplier);
+ queryContext =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryContext,
sourceProfileSupplier);
+ // update endTime/next checkpoint
+ endTime =
queryContext.getEndInstant().orElse(queryContext.getLastInstant());
}
+ Set<String> validInstants = new HashSet<>(queryContext.getInstants());
+ snapshot =
queryContext.getPredicateFilter().map(snapshot::filter).orElse(snapshot);
source = snapshot
// add filtering so that only interested records are returned.
- .filter(String.format("%s > '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
- queryInfo.getStartInstant()))
+ // completion time comparison uses ( , ], but when comparing start
time we need to use [, ]
+ .filter(String.format("%s >= '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
+ queryContext.getStartInstant().get()))
.filter(String.format("%s <= '%s'",
HoodieRecord.COMMIT_TIME_METADATA_FIELD,
- queryInfo.getEndInstant()));
- source =
queryInfo.getPredicateFilter().map(source::filter).orElse(source);
+
queryContext.getEndInstant().orElse(queryContext.getLastInstant())))
Review Comment:
Could `queryContext.getStartInstant()` or `queryContext.getEndInstant()` be
null?
##########
hudi-utilities/src/main/java/org/apache/hudi/utilities/sources/HoodieIncrSource.java:
##########
@@ -212,35 +257,43 @@ public Pair<Option<Dataset<Row>>, String>
fetchNextBatch(Option<String> lastCkpt
.collect(Collectors.toMap(Pair::getLeft, Pair::getRight));
reader = reader.options(optionsMap);
}
+ boolean shouldFullScan =
+
queryContext.getActiveTimeline().isBeforeTimelineStartsByCompletionTime(startTime)
+ && missingCheckpointStrategy ==
MissingCheckpointStrategy.READ_UPTO_LATEST_COMMIT;
Dataset<Row> source;
- // Do Incr pull. Set end instant if available
- if (queryInfo.isIncremental()) {
- source = reader
- .options(readOpts)
- .option(QUERY_TYPE().key(), QUERY_TYPE_INCREMENTAL_OPT_VAL())
- .option(BEGIN_INSTANTTIME().key(), queryInfo.getStartInstant())
- .option(END_INSTANTTIME().key(), queryInfo.getEndInstant())
- .option(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- props.getString(INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().key(),
- INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN().defaultValue()))
- .option(INCREMENTAL_READ_HANDLE_HOLLOW_COMMIT().key(),
handlingMode.name())
- .load(srcPath);
- } else {
- // if checkpoint is missing from source table, and if strategy is set to
READ_UPTO_LATEST_COMMIT, we have to issue snapshot query
+ if (instantRange.isEmpty() || shouldFullScan) {
+ // snapshot query
Dataset<Row> snapshot = reader
.options(readOpts)
.option(DataSourceReadOptions.QUERY_TYPE().key(),
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL())
.load(srcPath);
if (snapshotLoadQuerySplitter.isPresent()) {
- queryInfo =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryInfo,
sourceProfileSupplier);
+ queryContext =
snapshotLoadQuerySplitter.get().getNextCheckpoint(snapshot, queryContext,
sourceProfileSupplier);
+ // update endTime/next checkpoint
+ endTime =
queryContext.getEndInstant().orElse(queryContext.getLastInstant());
Review Comment:
I think we need to consider that the checkpoint can be more than the
completion time itself, like `completionTime#key`.
--
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]
