sunchao commented on a change in pull request #34659:
URL: https://github.com/apache/spark/pull/34659#discussion_r839059936
##########
File path:
sql/core/src/main/java/org/apache/spark/sql/execution/datasources/parquet/VectorizedRleValuesReader.java
##########
@@ -210,48 +213,358 @@ private void readBatchInternal(
} else if (rowId > rangeEnd) {
state.nextRange();
} else {
- // the range [rowId, rowId + n) overlaps with the current row range in
state
+ // The range [rowId, rowId + n) overlaps with the current row range in
state
long start = Math.max(rangeStart, rowId);
long end = Math.min(rangeEnd, rowId + n - 1);
- // skip the part [rowId, start)
+ // Skip the part [rowId, start)
int toSkip = (int) (start - rowId);
if (toSkip > 0) {
skipValues(toSkip, state, valueReader, updater);
rowId += toSkip;
leftInPage -= toSkip;
}
- // read the part [start, end]
+ // Read the part [start, end]
n = (int) (end - start + 1);
switch (mode) {
case RLE:
if (currentValue == state.maxDefinitionLevel) {
- updater.readValues(n, offset, values, valueReader);
- } else {
- nulls.putNulls(offset, n);
+ updater.readValues(n, state.valueOffset, values, valueReader);
+ state.valueOffset += n;
+ } else if (!state.isRequired && currentValue ==
state.maxDefinitionLevel - 1) {
+ // Only add null if this represents a null element, but not for
the case where a
+ // struct itself is null
+ nulls.putNulls(state.valueOffset, n);
+ state.valueOffset += n;
}
+ defLevels.putInts(state.levelOffset, n, currentValue);
break;
case PACKED:
for (int i = 0; i < n; ++i) {
- if (currentBuffer[currentBufferIdx++] ==
state.maxDefinitionLevel) {
- updater.readValue(offset + i, values, valueReader);
+ int value = currentBuffer[currentBufferIdx++];
+ if (value == state.maxDefinitionLevel) {
+ updater.readValue(state.valueOffset++, values, valueReader);
} else {
- nulls.putNull(offset + i);
+ // Only add null if this represents a null element, but not
for the case where a
+ // struct itself is null
+ nulls.putNull(state.valueOffset++);
}
+ defLevels.putInt(state.levelOffset + i, value);
}
break;
}
- offset += n;
+ state.levelOffset += n;
leftInBatch -= n;
rowId += n;
leftInPage -= n;
currentCount -= n;
+ defLevels.addElementsAppended(n);
+ }
+ }
+
+ state.rowsToReadInBatch = leftInBatch;
+ state.valuesToReadInPage = leftInPage;
+ state.rowId = rowId;
+ }
+
+ public void readBatchNested(
+ ParquetReadState state,
+ WritableColumnVector repLevels,
+ VectorizedRleValuesReader defLevelsReader,
+ WritableColumnVector defLevels,
+ WritableColumnVector values,
+ VectorizedValuesReader valueReader,
+ ParquetVectorUpdater updater) {
+ readBatchNestedInternal(state, repLevels, defLevelsReader, defLevels,
values, values, true,
+ valueReader, updater);
+ }
+
+ public void readIntegersNested(
+ ParquetReadState state,
+ WritableColumnVector repLevels,
+ VectorizedRleValuesReader defLevelsReader,
+ WritableColumnVector defLevels,
+ WritableColumnVector values,
+ WritableColumnVector nulls,
+ VectorizedValuesReader valueReader) {
+ readBatchNestedInternal(state, repLevels, defLevelsReader, defLevels,
values, nulls, false,
+ valueReader, new ParquetVectorUpdaterFactory.IntegerUpdater());
+ }
+
+ /**
+ * Keep reading repetition level values from the page until either: 1) we've
read enough
+ * top-level rows to fill the current batch, or 2) we've drained the data
page completely.
+ *
+ * @param valuesReused whether 'values' vector is reused for 'nulls'
+ */
+ public void readBatchNestedInternal(
+ ParquetReadState state,
+ WritableColumnVector repLevels,
+ VectorizedRleValuesReader defLevelsReader,
+ WritableColumnVector defLevels,
+ WritableColumnVector values,
+ WritableColumnVector nulls,
+ boolean valuesReused,
+ VectorizedValuesReader valueReader,
+ ParquetVectorUpdater updater) {
+
+ int leftInBatch = state.rowsToReadInBatch;
+ int leftInPage = state.valuesToReadInPage;
+ long rowId = state.rowId;
+
+ DefLevelProcessor defLevelProcessor = new
DefLevelProcessor(defLevelsReader, state, defLevels,
+ values, nulls, valuesReused, valueReader, updater);
+
+ while ((leftInBatch > 0 || !state.lastListCompleted) && leftInPage > 0) {
+ if (currentCount == 0 && !readNextGroup()) break;
+
+ // Values to read in the current RLE/PACKED block, must be <= what's
left in the page
+ int valuesLeftInBlock = Math.min(leftInPage, currentCount);
+
+ // The current row range start and end
+ long rangeStart = state.currentRangeStart();
+ long rangeEnd = state.currentRangeEnd();
+
+ switch (mode) {
+ case RLE:
+ // This RLE block is consist of top-level rows, so we'll need to
check
+ // if the rows should be skipped according to row indexes.
+ if (currentValue == 0) {
+ if (leftInBatch == 0) {
+ state.lastListCompleted = true;
+ } else {
+ // # of rows to read in the block, must be <= what's left in the
current batch
+ int n = Math.min(leftInBatch, valuesLeftInBlock);
+
+ if (rowId + n < rangeStart) {
+ // Need to skip all rows in [rowId, rowId + n)
+ defLevelProcessor.skipValues(n);
+ rowId += n;
+ currentCount -= n;
+ leftInPage -= n;
+ } else if (rowId > rangeEnd) {
+ // The current row index already beyond the current range:
move to the next range
+ // and repeat
+ state.nextRange();
+ } else {
+ // The range [rowId, rowId + n) overlaps with the current row
range
+ long start = Math.max(rangeStart, rowId);
+ long end = Math.min(rangeEnd, rowId + n - 1);
+
+ // Skip the rows in [rowId, start)
+ int toSkip = (int) (start - rowId);
+ if (toSkip > 0) {
+ defLevelProcessor.skipValues(toSkip);
+ rowId += toSkip;
+ currentCount -= toSkip;
+ leftInPage -= toSkip;
+ }
+
+ // Read the rows in [start, end]
+ n = (int) (end - start + 1);
+
+ if (n > 0) {
+ repLevels.appendInts(n, 0);
+ defLevelProcessor.readValues(n);
+ }
+
+ rowId += n;
+ currentCount -= n;
+ leftInBatch -= n;
+ leftInPage -= n;
+ }
+ }
+ } else {
+ // Not a top-level row: just read all the repetition levels in the
block if the row
+ // should be included according to row indexes, else skip the rows.
+ if (!state.shouldSkip) {
+ repLevels.appendInts(valuesLeftInBlock, currentValue);
+ }
+ state.numBatchedDefLevels += valuesLeftInBlock;
Review comment:
Yes, the definition levels and values are read in `DefLevelProcessor`,
in a batched fashion.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
