Skip to content

Commit

Permalink
feat(fuzzer): Support multiple joins in the join node "toSql" methods…
Browse files Browse the repository at this point in the history
… for reference query runners (facebookincubator#11801)

Summary:

Currently, the hash join and nested loop join "toSql" methods for all reference query runners only support a single join. This change extends it to support multiple joins, only needing the join node of the last join in the tree. It traverses up the tree and recursively builds the sql query.

Differential Revision: D66977480
  • Loading branch information
Daniel Hunte authored and facebook-github-bot committed Dec 30, 2024
1 parent 7cbfce0 commit c6ff398
Show file tree
Hide file tree
Showing 8 changed files with 433 additions and 312 deletions.
2 changes: 2 additions & 0 deletions velox/core/PlanNode.h
Original file line number Diff line number Diff line change
Expand Up @@ -324,6 +324,8 @@ class ValuesNode : public PlanNode {
const size_t repeatTimes_;
};

using ValuesNodePtr = std::shared_ptr<const ValuesNode>;

class ArrowStreamNode : public PlanNode {
public:
ArrowStreamNode(
Expand Down
150 changes: 17 additions & 133 deletions velox/exec/fuzzer/DuckQueryRunner.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,18 @@ DuckQueryRunner::aggregationFunctionDataSpecs() const {
return kAggregationFunctionDataSpecs;
}

std::multiset<std::vector<velox::variant>> DuckQueryRunner::execute(
const std::string& sql,
const core::PlanNodePtr& plan) {
DuckDbQueryRunner queryRunner;
std::unordered_map<std::string, std::vector<RowVectorPtr>> inputMap =
getAllTablesAndNames(plan);
for (const auto& [tableName, input] : inputMap) {
queryRunner.createTable(tableName, input);
}
return queryRunner.execute(sql, plan->outputType());
}

std::multiset<std::vector<velox::variant>> DuckQueryRunner::execute(
const std::string& sql,
const std::vector<RowVectorPtr>& input,
Expand Down Expand Up @@ -164,6 +176,11 @@ std::optional<std::string> DuckQueryRunner::toSql(
return toSql(joinNode);
}

if (const auto joinNode =
std::dynamic_pointer_cast<const core::ValuesNode>(plan)) {
return toSql(joinNode);
}

VELOX_NYI();
}

Expand Down Expand Up @@ -340,137 +357,4 @@ std::optional<std::string> DuckQueryRunner::toSql(

return sql.str();
}

std::optional<std::string> DuckQueryRunner::toSql(
const std::shared_ptr<const core::HashJoinNode>& joinNode) {
const auto& joinKeysToSql = [](auto keys) {
std::stringstream out;
for (auto i = 0; i < keys.size(); ++i) {
if (i > 0) {
out << ", ";
}
out << keys[i]->name();
}
return out.str();
};

const auto filterToSql = [](core::TypedExprPtr filter) {
auto call = std::dynamic_pointer_cast<const core::CallTypedExpr>(filter);
return toCallSql(call);
};

const auto& joinConditionAsSql = [&](auto joinNode) {
std::stringstream out;
for (auto i = 0; i < joinNode->leftKeys().size(); ++i) {
if (i > 0) {
out << " AND ";
}
out << joinNode->leftKeys()[i]->name() << " = "
<< joinNode->rightKeys()[i]->name();
}
if (joinNode->filter()) {
out << " AND " << filterToSql(joinNode->filter());
}
return out.str();
};

const auto& outputNames = joinNode->outputType()->names();

std::stringstream sql;
if (joinNode->isLeftSemiProjectJoin()) {
sql << "SELECT "
<< folly::join(", ", outputNames.begin(), --outputNames.end());
} else {
sql << "SELECT " << folly::join(", ", outputNames);
}

switch (joinNode->joinType()) {
case core::JoinType::kInner:
sql << " FROM t INNER JOIN u ON " << joinConditionAsSql(joinNode);
break;
case core::JoinType::kLeft:
sql << " FROM t LEFT JOIN u ON " << joinConditionAsSql(joinNode);
break;
case core::JoinType::kFull:
sql << " FROM t FULL OUTER JOIN u ON " << joinConditionAsSql(joinNode);
break;
case core::JoinType::kLeftSemiFilter:
// Multiple columns returned by a scalar subquery is not supported in
// DuckDB. A scalar subquery expression is a subquery that returns one
// result row from exactly one column for every input row.
if (joinNode->leftKeys().size() > 1) {
return std::nullopt;
}
sql << " FROM t WHERE " << joinKeysToSql(joinNode->leftKeys())
<< " IN (SELECT " << joinKeysToSql(joinNode->rightKeys())
<< " FROM u";
if (joinNode->filter()) {
sql << " WHERE " << filterToSql(joinNode->filter());
}
sql << ")";
break;
case core::JoinType::kLeftSemiProject:
if (joinNode->isNullAware()) {
sql << ", " << joinKeysToSql(joinNode->leftKeys()) << " IN (SELECT "
<< joinKeysToSql(joinNode->rightKeys()) << " FROM u";
if (joinNode->filter()) {
sql << " WHERE " << filterToSql(joinNode->filter());
}
sql << ") FROM t";
} else {
sql << ", EXISTS (SELECT * FROM u WHERE "
<< joinConditionAsSql(joinNode);
sql << ") FROM t";
}
break;
case core::JoinType::kAnti:
if (joinNode->isNullAware()) {
sql << " FROM t WHERE " << joinKeysToSql(joinNode->leftKeys())
<< " NOT IN (SELECT " << joinKeysToSql(joinNode->rightKeys())
<< " FROM u";
if (joinNode->filter()) {
sql << " WHERE " << filterToSql(joinNode->filter());
}
sql << ")";
} else {
sql << " FROM t WHERE NOT EXISTS (SELECT * FROM u WHERE "
<< joinConditionAsSql(joinNode);
sql << ")";
}
break;
default:
VELOX_UNREACHABLE(
"Unknown join type: {}", static_cast<int>(joinNode->joinType()));
}

return sql.str();
}

std::optional<std::string> DuckQueryRunner::toSql(
const std::shared_ptr<const core::NestedLoopJoinNode>& joinNode) {
std::stringstream sql;
sql << "SELECT " << folly::join(", ", joinNode->outputType()->names());

// Nested loop join without filter.
VELOX_CHECK(
joinNode->joinCondition() == nullptr,
"This code path should be called only for nested loop join without filter");
const std::string joinCondition{"(1 = 1)"};
switch (joinNode->joinType()) {
case core::JoinType::kInner:
sql << " FROM t INNER JOIN u ON " << joinCondition;
break;
case core::JoinType::kLeft:
sql << " FROM t LEFT JOIN u ON " << joinCondition;
break;
case core::JoinType::kFull:
sql << " FROM t FULL OUTER JOIN u ON " << joinCondition;
break;
default:
VELOX_UNREACHABLE(
"Unknown join type: {}", static_cast<int>(joinNode->joinType()));
}

return sql.str();
}
} // namespace facebook::velox::exec::test
13 changes: 7 additions & 6 deletions velox/exec/fuzzer/DuckQueryRunner.h
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,11 @@ class DuckQueryRunner : public ReferenceQueryRunner {
/// Assumes that source of AggregationNode or Window Node is 'tmp' table.
std::optional<std::string> toSql(const core::PlanNodePtr& plan) override;

/// Executes SQL query returned by the 'toSql' method based on the plan.
std::multiset<std::vector<velox::variant>> execute(
const std::string& sql,
const core::PlanNodePtr& plan) override;

/// Creates 'tmp' table with 'input' data and runs 'sql' query. Returns
/// results according to 'resultType' schema.
std::multiset<std::vector<velox::variant>> execute(
Expand All @@ -60,6 +65,8 @@ class DuckQueryRunner : public ReferenceQueryRunner {
const RowTypePtr& resultType) override;

private:
using ReferenceQueryRunner::toSql;

std::optional<std::string> toSql(
const std::shared_ptr<const core::AggregationNode>& aggregationNode);

Expand All @@ -72,12 +79,6 @@ class DuckQueryRunner : public ReferenceQueryRunner {
std::optional<std::string> toSql(
const std::shared_ptr<const core::RowNumberNode>& rowNumberNode);

std::optional<std::string> toSql(
const std::shared_ptr<const core::HashJoinNode>& joinNode);

std::optional<std::string> toSql(
const std::shared_ptr<const core::NestedLoopJoinNode>& joinNode);

std::unordered_set<std::string> aggregateFunctionNames_;
};

Expand Down
4 changes: 1 addition & 3 deletions velox/exec/fuzzer/JoinFuzzer.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -680,10 +680,8 @@ std::optional<MaterializedRowMultiset> JoinFuzzer::computeReferenceResults(
}

if (auto sql = referenceQueryRunner_->toSql(plan)) {
return referenceQueryRunner_->execute(
sql.value(), probeInput, buildInput, plan->outputType());
return referenceQueryRunner_->execute(*sql, plan);
}

LOG(INFO) << "Query not supported by the reference DB";
return std::nullopt;
}
Expand Down
Loading

0 comments on commit c6ff398

Please sign in to comment.