use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rex.RexLocalRef in project beam by apache.
the class BeamIOPushDownRule method isProjectRenameOnlyProgram.
/**
* Determine whether a program only performs renames and/or projects. RexProgram#isTrivial is not
* sufficient in this case, because number of projects does not need to be the same as inputs.
* Calc should NOT be dropped in the following cases:<br>
* 1. Projected fields are manipulated (ex: 'select field1+10').<br>
* 2. When the same field projected more than once.<br>
* 3. When an IO does not supports field reordering and projects fields in a different (from
* schema) order.
*
* @param program A program to check.
* @param projectReorderingSupported Whether project push-down supports field reordering.
* @return True when program performs only projects (w/o any modifications), false otherwise.
*/
@VisibleForTesting
boolean isProjectRenameOnlyProgram(RexProgram program, boolean projectReorderingSupported) {
int fieldCount = program.getInputRowType().getFieldCount();
Set<Integer> projectIndex = new HashSet<>();
int previousIndex = -1;
for (RexLocalRef ref : program.getProjectList()) {
int index = ref.getIndex();
if (// Projected values are InputRefs.
index >= fieldCount || // Each field projected once.
!projectIndex.add(ref.getIndex()) || (!projectReorderingSupported && index <= previousIndex)) {
// In the same order.
return false;
}
previousIndex = index;
}
return true;
}
use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rex.RexLocalRef in project beam by apache.
the class CalcRelSplitter method execute.
// ~ Methods ----------------------------------------------------------------
public RelNode execute() {
// expressions to the left.
assert program.isValid(Litmus.THROW, null);
final List<RexNode> exprList = program.getExprList();
final RexNode[] exprs = exprList.toArray(new RexNode[0]);
assert !RexUtil.containComplexExprs(exprList);
// Figure out what level each expression belongs to.
int[] exprLevels = new int[exprs.length];
// The type of a level is given by
// relTypes[levelTypeOrdinals[level]].
int[] levelTypeOrdinals = new int[exprs.length];
int levelCount = chooseLevels(exprs, -1, exprLevels, levelTypeOrdinals);
// For each expression, figure out which is the highest level where it
// is used.
int[] exprMaxUsingLevelOrdinals = new HighestUsageFinder(exprs, exprLevels).getMaxUsingLevelOrdinals();
// If expressions are used as outputs, mark them as higher than that.
final List<RexLocalRef> projectRefList = program.getProjectList();
final RexLocalRef conditionRef = program.getCondition();
for (RexLocalRef projectRef : projectRefList) {
exprMaxUsingLevelOrdinals[projectRef.getIndex()] = levelCount;
}
if (conditionRef != null) {
exprMaxUsingLevelOrdinals[conditionRef.getIndex()] = levelCount;
}
// Print out what we've got.
if (RULE_LOGGER.isTraceEnabled()) {
traceLevelExpressions(exprs, exprLevels, levelTypeOrdinals, levelCount);
}
// Now build the calcs.
RelNode rel = child;
final int inputFieldCount = program.getInputRowType().getFieldCount();
int[] inputExprOrdinals = identityArray(inputFieldCount);
boolean doneCondition = false;
for (int level = 0; level < levelCount; level++) {
final int[] projectExprOrdinals;
final RelDataType outputRowType;
if (level == (levelCount - 1)) {
outputRowType = program.getOutputRowType();
projectExprOrdinals = new int[projectRefList.size()];
for (int i = 0; i < projectExprOrdinals.length; i++) {
projectExprOrdinals[i] = projectRefList.get(i).getIndex();
}
} else {
outputRowType = null;
// Project the expressions which are computed at this level or
// before, and will be used at later levels.
List<Integer> projectExprOrdinalList = new ArrayList<>();
for (int i = 0; i < exprs.length; i++) {
RexNode expr = exprs[i];
if (expr instanceof RexLiteral) {
// Don't project literals. They are always created in
// the level where they are used.
exprLevels[i] = -1;
continue;
}
if ((exprLevels[i] <= level) && (exprMaxUsingLevelOrdinals[i] > level)) {
projectExprOrdinalList.add(i);
}
}
projectExprOrdinals = Ints.toArray(projectExprOrdinalList);
}
final RelType relType = relTypes[levelTypeOrdinals[level]];
// Can we do the condition this level?
int conditionExprOrdinal = -1;
if ((conditionRef != null) && !doneCondition) {
conditionExprOrdinal = conditionRef.getIndex();
if ((exprLevels[conditionExprOrdinal] > level) || !relType.supportsCondition()) {
// stand down -- we're not ready to do the condition yet
conditionExprOrdinal = -1;
} else {
doneCondition = true;
}
}
RexProgram program1 = createProgramForLevel(level, levelCount, rel.getRowType(), exprs, exprLevels, inputExprOrdinals, projectExprOrdinals, conditionExprOrdinal, outputRowType);
rel = relType.makeRel(cluster, traits, relBuilder, rel, program1);
rel = handle(rel);
// The outputs of this level will be the inputs to the next level.
inputExprOrdinals = projectExprOrdinals;
}
Preconditions.checkArgument(doneCondition || (conditionRef == null), "unhandled condition");
return rel;
}
use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rex.RexLocalRef in project beam by apache.
the class CalcRelSplitter method computeTopologicalOrdering.
/**
* Computes the order in which to visit expressions, so that we decide the level of an expression
* only after the levels of lower expressions have been decided.
*
* <p>First, we need to ensure that an expression is visited after all of its inputs.
*
* <p>Further, if the expression is a member of a cohort, we need to visit it after the inputs of
* all other expressions in that cohort. With this condition, expressions in the same cohort will
* very likely end up in the same level.
*
* <p>Note that if there are no cohorts, the expressions from the {@link RexProgram} are already
* in a suitable order. We perform the topological sort just to ensure that the code path is
* well-trodden.
*
* @param exprs Expressions
* @param cohorts List of cohorts, each of which is a set of expr ordinals
* @return Expression ordinals in topological order
*/
private static List<Integer> computeTopologicalOrdering(RexNode[] exprs, List<Set<Integer>> cohorts) {
final DirectedGraph<Integer, DefaultEdge> graph = DefaultDirectedGraph.create();
for (int i = 0; i < exprs.length; i++) {
graph.addVertex(i);
}
for (int i = 0; i < exprs.length; i++) {
final RexNode expr = exprs[i];
final Set<Integer> cohort = findCohort(cohorts, i);
final Set<Integer> targets;
if (cohort == null) {
targets = Collections.singleton(i);
} else {
targets = cohort;
}
expr.accept(new RexVisitorImpl<Void>(true) {
@Override
public Void visitLocalRef(RexLocalRef localRef) {
for (Integer target : targets) {
graph.addEdge(localRef.getIndex(), target);
}
return null;
}
});
}
TopologicalOrderIterator<Integer, DefaultEdge> iter = new TopologicalOrderIterator<>(graph);
final List<Integer> permutation = new ArrayList<>();
while (iter.hasNext()) {
permutation.add(iter.next());
}
return permutation;
}
use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rex.RexLocalRef in project beam by apache.
the class AbstractBeamCalcRel method estimateFilterSelectivity.
private static double estimateFilterSelectivity(RelNode child, RexProgram program, RelMetadataQuery mq) {
// Similar to calcite, if the calc node is representing filter operation we estimate the filter
// selectivity based on the number of equality conditions, number of inequality conditions, ....
RexLocalRef programCondition = program.getCondition();
RexNode condition;
if (programCondition == null) {
condition = null;
} else {
condition = program.expandLocalRef(programCondition);
}
// Currently this gets the selectivity based on Calcite's Selectivity Handler (RelMdSelectivity)
return mq.getSelectivity(child, condition);
}
use of org.apache.beam.vendor.calcite.v1_28_0.org.apache.calcite.rex.RexLocalRef in project beam by apache.
the class CalcRelSplitter method createProgramForLevel.
/**
* Creates a program containing the expressions for a given level.
*
* <p>The expression list of the program will consist of all entries in the expression list <code>
* allExprs[i]</code> for which the corresponding level ordinal <code>exprLevels[i]</code> is
* equal to <code>level</code>. Expressions are mapped according to <code>inputExprOrdinals</code>
* .
*
* @param level Level ordinal
* @param levelCount Number of levels
* @param inputRowType Input row type
* @param allExprs Array of all expressions
* @param exprLevels Array of the level ordinal of each expression
* @param inputExprOrdinals Ordinals in the expression list of input expressions. Input expression
* <code>i</code> will be found at position <code>inputExprOrdinals[i]</code>.
* @param projectExprOrdinals Ordinals of the expressions to be output this level.
* @param conditionExprOrdinal Ordinal of the expression to form the condition for this level, or
* -1 if there is no condition.
* @param outputRowType Output row type
* @return Relational expression
*/
private RexProgram createProgramForLevel(int level, int levelCount, RelDataType inputRowType, RexNode[] allExprs, int[] exprLevels, int[] inputExprOrdinals, final int[] projectExprOrdinals, int conditionExprOrdinal, @Nullable RelDataType outputRowType) {
// Build a list of expressions to form the calc.
List<RexNode> exprs = new ArrayList<>();
// exprInverseOrdinals describes where an expression in allExprs comes
// from -- from an input, from a calculated expression, or -1 if not
// available at this level.
int[] exprInverseOrdinals = new int[allExprs.length];
Arrays.fill(exprInverseOrdinals, -1);
int j = 0;
// and are used here.
for (int i = 0; i < inputExprOrdinals.length; i++) {
final int inputExprOrdinal = inputExprOrdinals[i];
exprs.add(new RexInputRef(i, allExprs[inputExprOrdinal].getType()));
exprInverseOrdinals[inputExprOrdinal] = j;
++j;
}
// Next populate the computed expressions.
final RexShuttle shuttle = new InputToCommonExprConverter(exprInverseOrdinals, exprLevels, level, inputExprOrdinals, allExprs);
for (int i = 0; i < allExprs.length; i++) {
if (exprLevels[i] == level || exprLevels[i] == -1 && level == (levelCount - 1) && allExprs[i] instanceof RexLiteral) {
RexNode expr = allExprs[i];
final RexNode translatedExpr = expr.accept(shuttle);
exprs.add(translatedExpr);
assert exprInverseOrdinals[i] == -1;
exprInverseOrdinals[i] = j;
++j;
}
}
// Form the projection and condition list. Project and condition
// ordinals are offsets into allExprs, so we need to map them into
// exprs.
final List<RexLocalRef> projectRefs = new ArrayList<>(projectExprOrdinals.length);
final List<String> fieldNames = new ArrayList<>(projectExprOrdinals.length);
for (int i = 0; i < projectExprOrdinals.length; i++) {
final int projectExprOrdinal = projectExprOrdinals[i];
final int index = exprInverseOrdinals[projectExprOrdinal];
assert index >= 0;
RexNode expr = allExprs[projectExprOrdinal];
projectRefs.add(new RexLocalRef(index, expr.getType()));
// Inherit meaningful field name if possible.
fieldNames.add(deriveFieldName(expr, i));
}
RexLocalRef conditionRef;
if (conditionExprOrdinal >= 0) {
final int index = exprInverseOrdinals[conditionExprOrdinal];
conditionRef = new RexLocalRef(index, allExprs[conditionExprOrdinal].getType());
} else {
conditionRef = null;
}
if (outputRowType == null) {
outputRowType = RexUtil.createStructType(typeFactory, projectRefs, fieldNames, null);
}
final RexProgram program = new RexProgram(inputRowType, exprs, projectRefs, conditionRef, outputRowType);
// call operands), since literals should be inlined.
return program;
}
Aggregations