Skip to content

[fix](local shuffle) Require hash input for distinct finalize agg without group keys #66570

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
924060929 wants to merge 1 commit into
base: master
Choose a base branch
from
+302 −20
Open

[fix](local shuffle) Require hash input for distinct finalize agg without group keys #66570

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
65 changes: 45 additions & 20 deletions fe/fe-core/src/main/java/org/apache/doris/planner/AggregationNode.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -335,7 +335,13 @@ public Pair<PlanNode, LocalExchangeType> enforceAndDeriveLocalExchange(
// early return also catches FIRST_MERGE, dropping the HASH requirement and
// causing wrong-result (e.g. PASSTHROUGH over serial child breaks the
// group-by-key invariant — DORIS-25413).
if (!hasKeys) {
if (!hasPartitionRequirement()) {

@github-actions github-actions Bot Aug 10, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P1] Do not create a zero-key HASH exchange for direct multi-distinct calls

This predicate also matches a directly called scalar multi_distinct_count(col). That function has isDistinct=false, so SplitAggWithoutDistinct builds a two-phase LOCAL aggregate with no partitionExpressions; over a non-hash scan both the child distribution list and grouping list are empty. With the default LE setting this branch now reaches requireHash(), and getLocalExchangeDistributeExprs() supplies zero expressions. BE's hash partitioner initializes every row to hash 0, so the resulting FE-planned HASH exchange sends the whole input to channel 0, collapsing the partial aggregate to one task per BE and concentrating its distinct state in one task. BE's own AggSinkOperatorX sees empty _partition_exprs on this path and uses the base requirement instead. Please base the LOCAL decision on a non-empty effective key (while retaining the keyed finalize fix), and add a direct-function plan case that rejects a zero-expression HASH exchange.

// No group keys and no DISTINCT aggregates (e.g. COUNT(*)): the input
// distribution is irrelevant. Any agg with group or distinct keys keeps
// its partition requirement — a finalize agg emits per-instance scalar
// values (sum0(multi_distinct_count(...)) above) that the parent sums,
// so same-key rows must stay in a single instance. This mirrors BE's
// `_partition_exprs` (grouping exprs, or distinct/distribute exprs).
requireChild = needsFinalize
? LocalExchangeTypeRequire.noRequire()
: baseClassRequire(connectContext);
Expand All @@ -348,13 +354,16 @@ public Pair<PlanNode, LocalExchangeType> enforceAndDeriveLocalExchange(
// FIRST_MERGE (correctness) or finalize+colocate → HASH.
requireChild = parentRequire.autoRequireHash();
} else if (hasPartitionExprs(parentRequire)) {
// FE-only heuristic: finalize non-colocate with parent hash requirement
// → inherit parent's specific hash type.
// finalize non-colocate with a parent hash requirement → inherit the
// parent's specific hash type.
requireChild = parentRequire.autoRequireHash();
} else {
// FE-only heuristic: finalize non-colocate without parent hash → skip
// LE (child Exchange already provides hash distribution).
requireChild = LocalExchangeTypeRequire.noRequire();
// finalize non-colocate without a parent hash requirement: the input
// must still be key-aligned (group/distinct key), so require HASH
// explicitly instead of trusting the child's distribution. When the
// child already provides hash distribution, satisfy() passes and no
// LE is inserted, so this is safe and free in the common case.
requireChild = LocalExchangeTypeRequire.requireHash();
}
}

Expand All @@ -371,6 +380,32 @@ private LocalExchangeTypeRequire baseClassRequire(ConnectContext connectContext)
: LocalExchangeTypeRequire.noRequire();
}

/**
* Whether this agg needs key-aligned (hash-partitioned) input from its child.
* Mirrors BE AggSinkOperatorX::update_operator's `_partition_exprs`: partition
* exprs are the grouping exprs, or the distinct/distribute exprs when the agg
* has DISTINCT functions. A finalize agg that emits per-instance scalar values
* (e.g. the sum0(multi_distinct_count(...)) above) is only correct when
* same-key rows stay in a single instance — overlapping keys across instances
* get summed multiple times by the parent.
*/
private boolean hasPartitionRequirement() {
return !aggInfo.getGroupingExprs().isEmpty() || hasDistinctAggregate();
}

private boolean hasDistinctAggregate() {
// Multi-distinct aggregates are detected by function name. Nereids rewrites
// count/sum/group_concat(distinct ...) into dedicated MultiDistinct* functions
// constructed with distinct=false, so by this legacy FunctionCallExpr layer
// isDistinct() is already false and the function name is the only signal.
return aggInfo.getAggregateExprs().stream()
.map(FunctionCallExpr::getFnName)
.filter(name -> name != null)
.map(name -> name.getFunction())
.filter(name -> name != null)
.anyMatch(name -> name.startsWith("multi_distinct_"));
}

@Override
protected List<Expr> getSemanticPartitionExprs() {
return aggInfo.getGroupingExprs();
Expand All @@ -386,18 +421,7 @@ protected List<Expr> getLocalExchangeDistributeExprs(int childIndex, boolean fol
// chain scatters same-group rows across N instances, leaving partial_preagg essentially a
// no-op and breaking row-arrival order at downstream merge-finalize (e.g. group_concat).
List<Expr> childDist = getChildDistributeExprList(childIndex);
// Multi-distinct aggregates are detected by function name. Nereids rewrites
// count/sum(distinct ...) into dedicated MultiDistinct* functions constructed with
// distinct=false and a "multi_distinct_" name, so by this legacy FunctionCallExpr layer
// isDistinct() is already false and the function name is the only remaining signal —
// there is no structural flag to test here.
boolean hasDistinct = aggInfo.getAggregateExprs().stream()
.map(FunctionCallExpr::getFnName)
.filter(name -> name != null)
.map(name -> name.getFunction())
.filter(name -> name != null)
.anyMatch(name -> name.startsWith("multi_distinct_"));
if (childDist != null && !childDist.isEmpty() && (followedByShuffled || hasDistinct)) {
if (childDist != null && !childDist.isEmpty() && (followedByShuffled || hasDistinctAggregate())) {
return childDist;
}
return Lists.newArrayList(aggInfo.getGroupingExprs());
Expand All @@ -406,13 +430,14 @@ protected List<Expr> getLocalExchangeDistributeExprs(int childIndex, boolean fol
@Override
public boolean requiresShuffleForCorrectness() {
// Mirrors BE's AggSinkOperatorX::is_shuffled_operator() exactly:
// finalize agg with group keys needs hash-distributed input for correctness.
// finalize agg with partition exprs (group keys or DISTINCT aggregates)
// needs hash-distributed input for correctness.
// GLOBAL dedup (!needsFinalize) is intentionally NOT included here — if a
// GLOBAL dedup exists, a finalize agg always sits above it (e.g. DISTINCT_GLOBAL
// above DISTINCT_LOCAL/GLOBAL_DEDUP), and the finalize agg propagates the flag
// down via inheritedShuffled. A solo finalize agg satisfies hash distribution
// through its own child requirement.
return needsFinalize && !aggInfo.getGroupingExprs().isEmpty();
return needsFinalize && hasPartitionRequirement();
}

private boolean canUseDistinctStreamingAgg(SessionVariable sessionVariable) {
Expand Down
Loading
Loading