use of org.apache.flink.table.planner.codegen.agg.AggsHandlerCodeGenerator in project flink by apache.
the class StreamExecGroupAggregate method translateToPlanInternal.
@SuppressWarnings("unchecked")
@Override
protected Transformation<RowData> translateToPlanInternal(PlannerBase planner, ExecNodeConfig config) {
if (grouping.length > 0 && config.getStateRetentionTime() < 0) {
LOG.warn("No state retention interval configured for a query which accumulates state. " + "Please provide a query configuration with valid retention interval to prevent excessive " + "state size. You may specify a retention time of 0 to not clean up the state.");
}
final ExecEdge inputEdge = getInputEdges().get(0);
final Transformation<RowData> inputTransform = (Transformation<RowData>) inputEdge.translateToPlan(planner);
final RowType inputRowType = (RowType) inputEdge.getOutputType();
final AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), planner.getRelBuilder(), JavaScalaConversionUtil.toScala(inputRowType.getChildren()), // TODO: but other operators do not copy this input field.....
true).needAccumulate();
if (needRetraction) {
generator.needRetract();
}
final AggregateInfoList aggInfoList = AggregateUtil.transformToStreamAggregateInfoList(inputRowType, JavaScalaConversionUtil.toScala(Arrays.asList(aggCalls)), aggCallNeedRetractions, needRetraction, true, true);
final GeneratedAggsHandleFunction aggsHandler = generator.generateAggsHandler("GroupAggsHandler", aggInfoList);
final LogicalType[] accTypes = Arrays.stream(aggInfoList.getAccTypes()).map(LogicalTypeDataTypeConverter::fromDataTypeToLogicalType).toArray(LogicalType[]::new);
final LogicalType[] aggValueTypes = Arrays.stream(aggInfoList.getActualValueTypes()).map(LogicalTypeDataTypeConverter::fromDataTypeToLogicalType).toArray(LogicalType[]::new);
final GeneratedRecordEqualiser recordEqualiser = new EqualiserCodeGenerator(aggValueTypes).generateRecordEqualiser("GroupAggValueEqualiser");
final int inputCountIndex = aggInfoList.getIndexOfCountStar();
final boolean isMiniBatchEnabled = config.get(ExecutionConfigOptions.TABLE_EXEC_MINIBATCH_ENABLED);
final OneInputStreamOperator<RowData, RowData> operator;
if (isMiniBatchEnabled) {
MiniBatchGroupAggFunction aggFunction = new MiniBatchGroupAggFunction(aggsHandler, recordEqualiser, accTypes, inputRowType, inputCountIndex, generateUpdateBefore, config.getStateRetentionTime());
operator = new KeyedMapBundleOperator<>(aggFunction, AggregateUtil.createMiniBatchTrigger(config));
} else {
GroupAggFunction aggFunction = new GroupAggFunction(aggsHandler, recordEqualiser, accTypes, inputCountIndex, generateUpdateBefore, config.getStateRetentionTime());
operator = new KeyedProcessOperator<>(aggFunction);
}
// partitioned aggregation
final OneInputTransformation<RowData, RowData> transform = ExecNodeUtil.createOneInputTransformation(inputTransform, createTransformationMeta(GROUP_AGGREGATE_TRANSFORMATION, config), operator, InternalTypeInfo.of(getOutputType()), inputTransform.getParallelism());
// set KeyType and Selector for state
final RowDataKeySelector selector = KeySelectorUtil.getRowDataSelector(grouping, InternalTypeInfo.of(inputRowType));
transform.setStateKeySelector(selector);
transform.setStateKeyType(selector.getProducedType());
return transform;
}
use of org.apache.flink.table.planner.codegen.agg.AggsHandlerCodeGenerator in project flink by apache.
the class StreamExecGroupWindowAggregate method createAggsHandler.
private GeneratedClass<?> createAggsHandler(AggregateInfoList aggInfoList, ExecNodeConfig config, RelBuilder relBuilder, List<LogicalType> fieldTypes, ZoneId shiftTimeZone) {
final boolean needMerge;
final Class<?> windowClass;
if (window instanceof SlidingGroupWindow) {
ValueLiteralExpression size = ((SlidingGroupWindow) window).size();
needMerge = hasTimeIntervalType(size);
windowClass = hasRowIntervalType(size) ? CountWindow.class : TimeWindow.class;
} else if (window instanceof TumblingGroupWindow) {
needMerge = false;
ValueLiteralExpression size = ((TumblingGroupWindow) window).size();
windowClass = hasRowIntervalType(size) ? CountWindow.class : TimeWindow.class;
} else if (window instanceof SessionGroupWindow) {
needMerge = true;
windowClass = TimeWindow.class;
} else {
throw new TableException("Unsupported window: " + window.toString());
}
final AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), relBuilder, JavaScalaConversionUtil.toScala(fieldTypes), // copyInputField
false).needAccumulate();
if (needMerge) {
generator.needMerge(0, false, null);
}
if (needRetraction) {
generator.needRetract();
}
final List<WindowProperty> windowProperties = Arrays.asList(Arrays.stream(namedWindowProperties).map(NamedWindowProperty::getProperty).toArray(WindowProperty[]::new));
final boolean isTableAggregate = isTableAggregate(Arrays.asList(aggInfoList.getActualAggregateCalls()));
if (isTableAggregate) {
return generator.generateNamespaceTableAggsHandler("GroupingWindowTableAggsHandler", aggInfoList, JavaScalaConversionUtil.toScala(windowProperties), windowClass, shiftTimeZone);
} else {
return generator.generateNamespaceAggsHandler("GroupingWindowAggsHandler", aggInfoList, JavaScalaConversionUtil.toScala(windowProperties), windowClass, shiftTimeZone);
}
}
use of org.apache.flink.table.planner.codegen.agg.AggsHandlerCodeGenerator in project flink by apache.
the class StreamExecGroupTableAggregate method translateToPlanInternal.
@SuppressWarnings("unchecked")
@Override
protected Transformation<RowData> translateToPlanInternal(PlannerBase planner, ExecNodeConfig config) {
if (grouping.length > 0 && config.getStateRetentionTime() < 0) {
LOG.warn("No state retention interval configured for a query which accumulates state. " + "Please provide a query configuration with valid retention interval to prevent excessive " + "state size. You may specify a retention time of 0 to not clean up the state.");
}
final ExecEdge inputEdge = getInputEdges().get(0);
final Transformation<RowData> inputTransform = (Transformation<RowData>) inputEdge.translateToPlan(planner);
final RowType inputRowType = (RowType) inputEdge.getOutputType();
final AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), planner.getRelBuilder(), JavaScalaConversionUtil.toScala(inputRowType.getChildren()), // TODO: but other operators do not copy this input field.....
true).needAccumulate();
if (needRetraction) {
generator.needRetract();
}
final AggregateInfoList aggInfoList = AggregateUtil.transformToStreamAggregateInfoList(inputRowType, JavaScalaConversionUtil.toScala(Arrays.asList(aggCalls)), aggCallNeedRetractions, needRetraction, // isStateBackendDataViews
true, // needDistinctInfo
true);
final GeneratedTableAggsHandleFunction aggsHandler = generator.generateTableAggsHandler("GroupTableAggHandler", aggInfoList);
final LogicalType[] accTypes = Arrays.stream(aggInfoList.getAccTypes()).map(LogicalTypeDataTypeConverter::fromDataTypeToLogicalType).toArray(LogicalType[]::new);
final int inputCountIndex = aggInfoList.getIndexOfCountStar();
final GroupTableAggFunction aggFunction = new GroupTableAggFunction(aggsHandler, accTypes, inputCountIndex, generateUpdateBefore, config.getStateRetentionTime());
final OneInputStreamOperator<RowData, RowData> operator = new KeyedProcessOperator<>(aggFunction);
// partitioned aggregation
final OneInputTransformation<RowData, RowData> transform = ExecNodeUtil.createOneInputTransformation(inputTransform, createTransformationMeta(GROUP_TABLE_AGGREGATE_TRANSFORMATION, config), operator, InternalTypeInfo.of(getOutputType()), inputTransform.getParallelism());
// set KeyType and Selector for state
final RowDataKeySelector selector = KeySelectorUtil.getRowDataSelector(grouping, InternalTypeInfo.of(inputRowType));
transform.setStateKeySelector(selector);
transform.setStateKeyType(selector.getProducedType());
return transform;
}
use of org.apache.flink.table.planner.codegen.agg.AggsHandlerCodeGenerator in project flink by apache.
the class BatchExecOverAggregate method createOverWindowFrames.
private List<OverWindowFrame> createOverWindowFrames(FlinkRelBuilder relBuilder, ExecNodeConfig config, RowType inputType, SortSpec sortSpec, RowType inputTypeWithConstants) {
final List<OverWindowFrame> windowFrames = new ArrayList<>();
for (GroupSpec group : overSpec.getGroups()) {
OverWindowMode mode = inferGroupMode(group);
if (mode == OverWindowMode.OFFSET) {
for (AggregateCall aggCall : group.getAggCalls()) {
AggregateInfoList aggInfoList = AggregateUtil.transformToBatchAggregateInfoList(inputTypeWithConstants, JavaScalaConversionUtil.toScala(Collections.singletonList(aggCall)), new boolean[] { true }, /* needRetraction = true, See LeadLagAggFunction */
sortSpec.getFieldIndices());
AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), relBuilder, JavaScalaConversionUtil.toScala(inputType.getChildren()), // copyInputField
false);
// over agg code gen must pass the constants
GeneratedAggsHandleFunction genAggsHandler = generator.needAccumulate().needRetract().withConstants(JavaScalaConversionUtil.toScala(getConstants())).generateAggsHandler("BoundedOverAggregateHelper", aggInfoList);
// LEAD is behind the currentRow, so we need plus offset.
// LAG is in front of the currentRow, so we need minus offset.
long flag = aggCall.getAggregation().kind == SqlKind.LEAD ? 1L : -1L;
final Long offset;
final OffsetOverFrame.CalcOffsetFunc calcOffsetFunc;
// The second arg mean the offset arg index for leag/lag function, default is 1.
if (aggCall.getArgList().size() >= 2) {
int constantIndex = aggCall.getArgList().get(1) - overSpec.getOriginalInputFields();
if (constantIndex < 0) {
offset = null;
int rowIndex = aggCall.getArgList().get(1);
switch(inputType.getTypeAt(rowIndex).getTypeRoot()) {
case BIGINT:
calcOffsetFunc = row -> row.getLong(rowIndex) * flag;
break;
case INTEGER:
calcOffsetFunc = row -> (long) row.getInt(rowIndex) * flag;
break;
case SMALLINT:
calcOffsetFunc = row -> (long) row.getShort(rowIndex) * flag;
break;
default:
throw new RuntimeException("The column type must be in long/int/short.");
}
} else {
long constantOffset = getConstants().get(constantIndex).getValueAs(Long.class);
offset = constantOffset * flag;
calcOffsetFunc = null;
}
} else {
offset = flag;
calcOffsetFunc = null;
}
windowFrames.add(new OffsetOverFrame(genAggsHandler, offset, calcOffsetFunc));
}
} else {
AggregateInfoList aggInfoList = AggregateUtil.transformToBatchAggregateInfoList(// inputSchema.relDataType
inputTypeWithConstants, JavaScalaConversionUtil.toScala(group.getAggCalls()), // aggCallNeedRetractions
null, sortSpec.getFieldIndices());
AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), relBuilder, JavaScalaConversionUtil.toScala(inputType.getChildren()), // copyInputField
false);
// over agg code gen must pass the constants
GeneratedAggsHandleFunction genAggsHandler = generator.needAccumulate().withConstants(JavaScalaConversionUtil.toScala(getConstants())).generateAggsHandler("BoundedOverAggregateHelper", aggInfoList);
RowType valueType = generator.valueType();
final OverWindowFrame frame;
switch(mode) {
case RANGE:
if (isUnboundedWindow(group)) {
frame = new UnboundedOverWindowFrame(genAggsHandler, valueType);
} else if (isUnboundedPrecedingWindow(group)) {
GeneratedRecordComparator genBoundComparator = createBoundComparator(relBuilder, config, sortSpec, group.getUpperBound(), false, inputType);
frame = new RangeUnboundedPrecedingOverFrame(genAggsHandler, genBoundComparator);
} else if (isUnboundedFollowingWindow(group)) {
GeneratedRecordComparator genBoundComparator = createBoundComparator(relBuilder, config, sortSpec, group.getLowerBound(), true, inputType);
frame = new RangeUnboundedFollowingOverFrame(valueType, genAggsHandler, genBoundComparator);
} else if (isSlidingWindow(group)) {
GeneratedRecordComparator genLeftBoundComparator = createBoundComparator(relBuilder, config, sortSpec, group.getLowerBound(), true, inputType);
GeneratedRecordComparator genRightBoundComparator = createBoundComparator(relBuilder, config, sortSpec, group.getUpperBound(), false, inputType);
frame = new RangeSlidingOverFrame(inputType, valueType, genAggsHandler, genLeftBoundComparator, genRightBoundComparator);
} else {
throw new TableException("This should not happen.");
}
break;
case ROW:
if (isUnboundedWindow(group)) {
frame = new UnboundedOverWindowFrame(genAggsHandler, valueType);
} else if (isUnboundedPrecedingWindow(group)) {
frame = new RowUnboundedPrecedingOverFrame(genAggsHandler, OverAggregateUtil.getLongBoundary(overSpec, group.getUpperBound()));
} else if (isUnboundedFollowingWindow(group)) {
frame = new RowUnboundedFollowingOverFrame(valueType, genAggsHandler, OverAggregateUtil.getLongBoundary(overSpec, group.getLowerBound()));
} else if (isSlidingWindow(group)) {
frame = new RowSlidingOverFrame(inputType, valueType, genAggsHandler, OverAggregateUtil.getLongBoundary(overSpec, group.getLowerBound()), OverAggregateUtil.getLongBoundary(overSpec, group.getUpperBound()));
} else {
throw new TableException("This should not happen.");
}
break;
case INSENSITIVE:
frame = new InsensitiveOverFrame(genAggsHandler);
break;
default:
throw new TableException("This should not happen.");
}
windowFrames.add(frame);
}
}
return windowFrames;
}
use of org.apache.flink.table.planner.codegen.agg.AggsHandlerCodeGenerator in project flink by apache.
the class BatchExecOverAggregate method translateToPlanInternal.
@SuppressWarnings("unchecked")
@Override
protected Transformation<RowData> translateToPlanInternal(PlannerBase planner, ExecNodeConfig config) {
final ExecEdge inputEdge = getInputEdges().get(0);
final Transformation<RowData> inputTransform = (Transformation<RowData>) inputEdge.translateToPlan(planner);
final RowType inputType = (RowType) inputEdge.getOutputType();
// The generated sort is used for generating the comparator among partitions.
// So here not care the ASC or DESC for the grouping fields.
// TODO just replace comparator to equaliser
final int[] partitionFields = overSpec.getPartition().getFieldIndices();
final GeneratedRecordComparator genComparator = ComparatorCodeGenerator.gen(config.getTableConfig(), "SortComparator", inputType, SortUtil.getAscendingSortSpec(partitionFields));
// use aggInputType which considers constants as input instead of inputType
final RowType inputTypeWithConstants = getInputTypeWithConstants();
// Over operator could support different order-by keys with collation satisfied.
// Currently, this operator requires all order keys (combined with partition keys) are
// the same, but order-by keys may be different. Consider the following sql:
// select *, sum(b) over partition by a order by a, count(c) over partition by a from T
// So we can use any one from the groups. To keep the behavior with the rule, we use the
// last one.
final SortSpec sortSpec = overSpec.getGroups().get(overSpec.getGroups().size() - 1).getSort();
final TableStreamOperator<RowData> operator;
final long managedMemory;
if (!needBufferData()) {
// operator needn't cache data
final int numOfGroup = overSpec.getGroups().size();
final GeneratedAggsHandleFunction[] aggsHandlers = new GeneratedAggsHandleFunction[numOfGroup];
final boolean[] resetAccumulators = new boolean[numOfGroup];
for (int i = 0; i < numOfGroup; ++i) {
GroupSpec group = overSpec.getGroups().get(i);
AggregateInfoList aggInfoList = AggregateUtil.transformToBatchAggregateInfoList(inputTypeWithConstants, JavaScalaConversionUtil.toScala(group.getAggCalls()), // aggCallNeedRetractions
null, sortSpec.getFieldIndices());
AggsHandlerCodeGenerator generator = new AggsHandlerCodeGenerator(new CodeGeneratorContext(config.getTableConfig()), planner.getRelBuilder(), JavaScalaConversionUtil.toScala(inputType.getChildren()), // copyInputField
false);
// over agg code gen must pass the constants
aggsHandlers[i] = generator.needAccumulate().withConstants(JavaScalaConversionUtil.toScala(getConstants())).generateAggsHandler("BoundedOverAggregateHelper", aggInfoList);
OverWindowMode mode = inferGroupMode(group);
resetAccumulators[i] = mode == OverWindowMode.ROW && group.getLowerBound().isCurrentRow() && group.getUpperBound().isCurrentRow();
}
operator = new NonBufferOverWindowOperator(aggsHandlers, genComparator, resetAccumulators);
managedMemory = 0L;
} else {
List<OverWindowFrame> windowFrames = createOverWindowFrames(planner.getRelBuilder(), config, inputType, sortSpec, inputTypeWithConstants);
operator = new BufferDataOverWindowOperator(windowFrames.toArray(new OverWindowFrame[0]), genComparator, inputType.getChildren().stream().allMatch(BinaryRowData::isInFixedLengthPart));
managedMemory = config.get(ExecutionConfigOptions.TABLE_EXEC_RESOURCE_EXTERNAL_BUFFER_MEMORY).getBytes();
}
return ExecNodeUtil.createOneInputTransformation(inputTransform, createTransformationName(config), createTransformationDescription(config), SimpleOperatorFactory.of(operator), InternalTypeInfo.of(getOutputType()), inputTransform.getParallelism(), managedMemory);
}
Aggregations