Skip to content
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.

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

Restrict size of automatically broadcasted tables #11667

Merged
merged 3 commits into from
Oct 11, 2018
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,7 @@
import javax.inject.Inject;
sopel39 marked this conversation as resolved.
Show resolved Hide resolved

import java.util.List;
import java.util.Optional;
import java.util.OptionalInt;
import java.util.stream.Stream;

Expand All @@ -53,6 +54,7 @@ public final class SystemSessionProperties
{
public static final String OPTIMIZE_HASH_GENERATION = "optimize_hash_generation";
public static final String JOIN_DISTRIBUTION_TYPE = "join_distribution_type";
public static final String JOIN_MAX_BROADCAST_TABLE_SIZE = "join_max_broadcast_table_size";
public static final String DISTRIBUTED_JOIN = "distributed_join";
public static final String DISTRIBUTED_INDEX_JOIN = "distributed_index_join";
public static final String HASH_PARTITION_COUNT = "hash_partition_count";
Expand Down Expand Up @@ -152,6 +154,15 @@ public SystemSessionProperties(
false,
value -> JoinDistributionType.valueOf(((String) value).toUpperCase()),
JoinDistributionType::name),
new PropertyMetadata<>(
JOIN_MAX_BROADCAST_TABLE_SIZE,
"Maximum estimated size of a table that can be broadcast for JOIN.",
VARCHAR,
DataSize.class,
featuresConfig.getJoinMaxBroadcastTableSize(),
true,
value -> DataSize.valueOf((String) value),
DataSize::toString),
booleanProperty(
DISTRIBUTED_INDEX_JOIN,
"Distribute index joins on join keys instead of executing inline",
Expand Down Expand Up @@ -534,6 +545,11 @@ public static JoinDistributionType getJoinDistributionType(Session session)
return session.getSystemProperty(JOIN_DISTRIBUTION_TYPE, JoinDistributionType.class);
}

public static Optional<DataSize> getJoinMaxBroadcastTableSize(Session session)
{
return Optional.ofNullable(session.getSystemProperty(JOIN_MAX_BROADCAST_TABLE_SIZE, DataSize.class));
}

public static boolean isDistributedIndexJoinEnabled(Session session)
{
return session.getSystemProperty(DISTRIBUTED_INDEX_JOIN, Boolean.class);
Expand Down Expand Up @@ -811,6 +827,7 @@ public static OptionalInt getMaxDriversPerTask(Session session)
}
return OptionalInt.of(value);
}

private static int validateValueIsPowerOfTwo(Object value, String property)
{
int intValue = ((Number) requireNonNull(value, "value is null")).intValue();
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -65,6 +65,7 @@ public class FeaturesConfig
private double networkCostWeight = 15;
private boolean distributedIndexJoinsEnabled;
private JoinDistributionType joinDistributionType = PARTITIONED;
private DataSize joinMaxBroadcastTableSize;
private boolean colocatedJoinsEnabled;
private boolean groupedExecutionForAggregationEnabled;
private int concurrentLifespansPerTask;
Expand Down Expand Up @@ -290,6 +291,18 @@ public FeaturesConfig setJoinDistributionType(JoinDistributionType joinDistribut
return this;
}

public DataSize getJoinMaxBroadcastTableSize()
{
return joinMaxBroadcastTableSize;
}

@Config("join-max-broadcast-table-size")
public FeaturesConfig setJoinMaxBroadcastTableSize(DataSize joinMaxBroadcastTableSize)
{
this.joinMaxBroadcastTableSize = joinMaxBroadcastTableSize;
return this;
}

public boolean isGroupedExecutionForAggregationEnabled()
{
return groupedExecutionForAggregationEnabled;
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -17,18 +17,22 @@
import com.facebook.presto.cost.CostComparator;
import com.facebook.presto.cost.CostProvider;
import com.facebook.presto.cost.PlanNodeCostEstimate;
import com.facebook.presto.cost.PlanNodeStatsEstimate;
import com.facebook.presto.matching.Captures;
import com.facebook.presto.matching.Pattern;
import com.facebook.presto.sql.analyzer.FeaturesConfig.JoinDistributionType;
import com.facebook.presto.sql.planner.iterative.Rule;
import com.facebook.presto.sql.planner.plan.JoinNode;
import com.facebook.presto.sql.planner.plan.PlanNode;
import com.google.common.collect.Ordering;
import io.airlift.units.DataSize;

import java.util.ArrayList;
import java.util.List;
import java.util.Optional;

import static com.facebook.presto.SystemSessionProperties.getJoinDistributionType;
import static com.facebook.presto.SystemSessionProperties.getJoinMaxBroadcastTableSize;
import static com.facebook.presto.sql.analyzer.FeaturesConfig.JoinDistributionType.AUTOMATIC;
import static com.facebook.presto.sql.planner.optimizations.QueryCardinalityUtil.isAtMostScalar;
import static com.facebook.presto.sql.planner.plan.JoinNode.DistributionType.PARTITIONED;
Expand Down Expand Up @@ -68,26 +72,31 @@ public Result apply(JoinNode joinNode, Captures captures, Context context)
return Result.ofPlanNode(getSyntacticOrderJoin(joinNode, context, joinDistributionType));
}

private PlanNode getCostBasedJoin(JoinNode joinNode, Context context)
public static boolean canReplicate(JoinNode joinNode, Context context)
{
CostProvider costProvider = context.getCostProvider();
List<PlanNodeWithCost> possibleJoinNodes = new ArrayList<>();

if (!mustPartition(joinNode)) {
possibleJoinNodes.add(getJoinNodeWithCost(costProvider, joinNode.withDistributionType(REPLICATED)));
}
if (!mustReplicate(joinNode, context)) {
possibleJoinNodes.add(getJoinNodeWithCost(costProvider, joinNode.withDistributionType(PARTITIONED)));
JoinDistributionType joinDistributionType = getJoinDistributionType(context.getSession());
if (!joinDistributionType.canReplicate()) {
return false;
}

JoinNode flipped = joinNode.flipChildren();
if (!mustPartition(flipped)) {
possibleJoinNodes.add(getJoinNodeWithCost(costProvider, flipped.withDistributionType(REPLICATED)));
}
if (!mustReplicate(flipped, context)) {
possibleJoinNodes.add(getJoinNodeWithCost(costProvider, flipped.withDistributionType(PARTITIONED)));
Optional<DataSize> joinMaxBroadcastTableSize = getJoinMaxBroadcastTableSize(context.getSession());
if (!joinMaxBroadcastTableSize.isPresent()) {
return true;
}

PlanNode buildSide = joinNode.getRight();
PlanNodeStatsEstimate buildSideStatsEstimate = context.getStatsProvider().getStats(buildSide);
double buildSideSizeInBytes = buildSideStatsEstimate.getOutputSizeInBytes(buildSide.getOutputSymbols(), context.getSymbolAllocator().getTypes());
return buildSideSizeInBytes <= joinMaxBroadcastTableSize.get().toBytes();
}

private PlanNode getCostBasedJoin(JoinNode joinNode, Context context)
{
List<PlanNodeWithCost> possibleJoinNodes = new ArrayList<>();

addJoinsWithDifferentDistributions(joinNode, possibleJoinNodes, context);
addJoinsWithDifferentDistributions(joinNode.flipChildren(), possibleJoinNodes, context);

if (possibleJoinNodes.stream().anyMatch(result -> result.getCost().hasUnknownComponents()) || possibleJoinNodes.isEmpty()) {
return getSyntacticOrderJoin(joinNode, context, AUTOMATIC);
}
Expand All @@ -97,6 +106,16 @@ private PlanNode getCostBasedJoin(JoinNode joinNode, Context context)
return planNodeOrderings.min(possibleJoinNodes).getPlanNode();
}

private void addJoinsWithDifferentDistributions(JoinNode joinNode, List<PlanNodeWithCost> possibleJoinNodes, Context context)
{
if (!mustPartition(joinNode) && canReplicate(joinNode, context)) {
possibleJoinNodes.add(getJoinNodeWithCost(context.getCostProvider(), joinNode.withDistributionType(REPLICATED)));
}
if (!mustReplicate(joinNode, context)) {
possibleJoinNodes.add(getJoinNodeWithCost(context.getCostProvider(), joinNode.withDistributionType(PARTITIONED)));
}
}

private PlanNode getSyntacticOrderJoin(JoinNode joinNode, Context context, JoinDistributionType joinDistributionType)
{
if (mustPartition(joinNode)) {
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@
import static com.facebook.presto.sql.planner.DeterminismEvaluator.isDeterministic;
import static com.facebook.presto.sql.planner.EqualityInference.createEqualityInference;
import static com.facebook.presto.sql.planner.EqualityInference.nonInferrableConjuncts;
import static com.facebook.presto.sql.planner.iterative.rule.DetermineJoinDistributionType.canReplicate;
import static com.facebook.presto.sql.planner.iterative.rule.ReorderJoins.JoinEnumerationResult.INFINITE_COST_RESULT;
import static com.facebook.presto.sql.planner.iterative.rule.ReorderJoins.JoinEnumerationResult.UNKNOWN_COST_RESULT;
import static com.facebook.presto.sql.planner.iterative.rule.ReorderJoins.MultiJoinNode.toMultiJoinNode;
Expand Down Expand Up @@ -374,7 +375,7 @@ private JoinEnumerationResult setJoinNodeProperties(JoinNode joinNode)
possibleJoinNodes.add(createJoinEnumerationResult(joinNode.withDistributionType(PARTITIONED)));
possibleJoinNodes.add(createJoinEnumerationResult(joinNode.flipChildren().withDistributionType(PARTITIONED)));
}
if (joinDistributionType.canReplicate()) {
if (canReplicate(joinNode, context)) {
possibleJoinNodes.add(createJoinEnumerationResult(joinNode.withDistributionType(REPLICATED)));
possibleJoinNodes.add(createJoinEnumerationResult(joinNode.flipChildren().withDistributionType(REPLICATED)));
}
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,7 @@ public void testDefaults()
.setNetworkCostWeight(15)
.setDistributedIndexJoinsEnabled(false)
.setJoinDistributionType(PARTITIONED)
.setJoinMaxBroadcastTableSize(null)
.setGroupedExecutionForAggregationEnabled(false)
.setConcurrentLifespansPerTask(0)
.setFastInequalityJoins(true)
Expand Down Expand Up @@ -125,6 +126,7 @@ public void testExplicitPropertyMappings()
.put("deprecated.legacy-char-to-varchar-coercion", "true")
.put("distributed-index-joins-enabled", "true")
.put("join-distribution-type", "BROADCAST")
.put("join-max-broadcast-table-size", "42GB")
.put("grouped-execution-for-aggregation-enabled", "true")
.put("concurrent-lifespans-per-task", "1")
.put("fast-inequality-joins", "false")
Expand Down Expand Up @@ -179,6 +181,7 @@ public void testExplicitPropertyMappings()
.setEnableStatsCalculator(false)
.setDistributedIndexJoinsEnabled(true)
.setJoinDistributionType(BROADCAST)
.setJoinMaxBroadcastTableSize(new DataSize(42, GIGABYTE))
.setGroupedExecutionForAggregationEnabled(true)
.setConcurrentLifespansPerTask(1)
.setFastInequalityJoins(false)
Expand Down
Loading