[Enhancement] support common expr reuse in complex case-when expr in scan predicates (#62779)

Signed-off-by: silverbullet233 <3675229+silverbullet233@users.noreply.github.com>
This commit is contained in:
eyes_on_me 2025-09-08 13:51:04 +08:00 committed by GitHub
parent 668fb46cad
commit 726fa67c3f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 63 additions and 6 deletions

View File

@ -18,11 +18,14 @@ import com.starrocks.catalog.Column;
import com.starrocks.sql.optimizer.Utils;
import com.starrocks.sql.optimizer.operator.OperatorType;
import com.starrocks.sql.optimizer.operator.scalar.BinaryPredicateOperator;
import com.starrocks.sql.optimizer.operator.scalar.CallOperator;
import com.starrocks.sql.optimizer.operator.scalar.CaseWhenOperator;
import com.starrocks.sql.optimizer.operator.scalar.ColumnRefOperator;
import com.starrocks.sql.optimizer.operator.scalar.CompoundPredicateOperator;
import com.starrocks.sql.optimizer.operator.scalar.ConstantOperator;
import com.starrocks.sql.optimizer.operator.scalar.LambdaFunctionOperator;
import com.starrocks.sql.optimizer.operator.scalar.ScalarOperator;
import com.starrocks.sql.optimizer.operator.scalar.ScalarOperatorUtil;
import com.starrocks.sql.optimizer.operator.scalar.ScalarOperatorVisitor;
import java.util.LinkedList;
@ -102,7 +105,7 @@ public class TableScanPredicateExtractor {
}
// CanFullyPushDownVisitor is used to check whether a predicate can be pushed down into ScanNode.
// currently, we only allow single-column predicates that not contain lambda expressions to be pushed down.
// currently, we only allow single-column predicates that not contain lambda expressions and complex case-when to be pushed down.
private class CanFullyPushDownVisitor extends ScalarOperatorVisitor<Boolean, CanFullyPushDownVisitorContext> {
private final Map<ColumnRefOperator, Column> columnRefOperatorColumnMap;
@ -172,5 +175,17 @@ public class TableScanPredicateExtractor {
public Boolean visitLambdaFunctionOperator(LambdaFunctionOperator op, CanFullyPushDownVisitorContext context) {
return false;
}
@Override
public Boolean visitCaseWhenOperator(CaseWhenOperator op, CanFullyPushDownVisitorContext context) {
if (!visit(op, context)) {
return false;
}
boolean containsCallOperator = op.getAllConditionClause().stream()
.anyMatch(scalarOperator ->
ScalarOperatorUtil.getStream(scalarOperator)
.anyMatch(operator -> operator instanceof CallOperator));
return !containsCallOperator;
}
}
}

View File

@ -315,6 +315,39 @@ public class ScanPredicateExprReuseTest extends PlanTestBase {
" Pruned type: 7 <-> [MAP<INT,INT>]\n" +
" ColumnAccessPath: [/v1/OFFSET, /v5/b/a/OFFSET, /v6/OFFSET]");
}
{
String sql = "select * from tarray where " +
"(case when array_length(v3) < 8 then 0 when array_length(v3) < 16 then 1 end) >= 1";
String plan = getFragmentPlan(sql);
assertContains(plan, " 2:SELECT\n" +
" | predicates: CASE WHEN 4: array_length < 8 THEN 0 WHEN 4: array_length < 16 THEN 1 END >= 1\n" +
" | \n" +
" 1:Project\n" +
" | <slot 1> : 1: v1\n" +
" | <slot 2> : 2: v2\n" +
" | <slot 3> : 3: v3\n" +
" | <slot 4> : array_length(3: v3)");
}
{
String sql = "select * from tarray where (case array_length(v3) when 1 then v1 when 2 then v2 end) >= 10";
String plan = getFragmentPlan(sql);
assertContains(plan, " 2:SELECT\n" +
" | predicates: CASE 4: array_length WHEN 1 THEN 1: v1 WHEN 2 THEN 2: v2 END >= 10\n" +
" | \n" +
" 1:Project\n" +
" | <slot 1> : 1: v1\n" +
" | <slot 2> : 2: v2\n" +
" | <slot 3> : 3: v3\n" +
" | <slot 4> : array_length(3: v3)");
}
{
String sql = "select * from t0 where (case when v1 > 1 then v1 + 10 when v1 > 2 then v1 + 20 else v1 end) = 2";
String plan = getFragmentPlan(sql);
assertContains(plan, " 0:OlapScanNode\n" +
" TABLE: t0\n" +
" PREAGGREGATION: ON\n" +
" PREDICATES: CASE WHEN 1: v1 > 1 THEN 1: v1 + 10 WHEN 1: v1 > 2 THEN 1: v1 + 20 ELSE 1: v1 END = 2");
}
}
@Test

View File

@ -613,11 +613,8 @@ class SelectStmtWithCaseWhenTest {
} else {
Assertions.assertTrue(patterns.stream().anyMatch(plan::contains), joiner.toString());
}
starRocksAssert.getCtx().getSessionVariable().setEnablePredicateExprReuse(true);
// String code = "{\n\"" + sql.replace("\n", " ") + "\",\n" + Arrays.stream(plan.split("\n"))
// .filter(c -> c.contains("PREDICATES:")).map(c -> c.replace("PREDICATES:", "").trim())
// .collect(Collectors.joining(", ", "\"", "\"")) + "},";
// System.out.println(code);
}
@Test
@ -639,6 +636,18 @@ class SelectStmtWithCaseWhenTest {
"FROM cte01\n" +
"WHERE len_bucket IS NOT NULL;";
String plan = UtFrameUtils.getVerboseFragmentPlan(starRocksAssert.getCtx(), sql);
Assert.assertTrue(plan.contains("CASE WHEN array_length"));
Assert.assertTrue(plan.contains(" 2:SELECT\n" +
" | predicates: 3: case IS NOT NULL\n" +
" | cardinality: 1\n" +
" | \n" +
" 1:Project\n" +
" | output columns:\n" +
" | 1 <-> [1: id, VARCHAR, false]\n" +
" | 3 <-> CASE WHEN [5: array_length, INT, true] < 2 THEN 'bucket1' " +
"WHEN ([5: array_length, INT, true] >= 2) AND ([5: array_length, INT, true] < 4) THEN 'bucket2' ELSE NULL END\n" +
" | 4 <-> [5: array_length, INT, true]\n" +
" | common expressions:\n" +
" | 5 <-> array_length[([2: col_arr, ARRAY<VARCHAR(100)>, true]); " +
"args: INVALID_TYPE; result: INT; args nullable: true; result nullable: true]"));
}
}