Skip to content

[CBO] Make optimizer get only columns stats that are used in request #7550

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

Merged
merged 20 commits into from
Aug 16, 2024
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
2 changes: 1 addition & 1 deletion ydb/core/kqp/host/kqp_host.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -1762,7 +1762,7 @@ class TKqpHost : public IKqpHost {

void Init(EKikimrQueryType queryType) {
TransformCtx = MakeIntrusive<TKqlTransformContext>(Config, SessionCtx->QueryPtr(), SessionCtx->TablesPtr());
KqpRunner = CreateKqpRunner(Gateway, Cluster, TypesCtx, SessionCtx, TransformCtx, *FuncRegistry);
KqpRunner = CreateKqpRunner(Gateway, Cluster, TypesCtx, SessionCtx, TransformCtx, *FuncRegistry, ActorSystem);

ExprCtx->NodesAllocationLimit = SessionCtx->Config()._KqpExprNodesAllocationLimit.Get().GetRef();
ExprCtx->StringsAllocationLimit = SessionCtx->Config()._KqpExprStringsAllocationLimit.Get().GetRef();
Expand Down
2 changes: 1 addition & 1 deletion ydb/core/kqp/host/kqp_host_impl.h
Original file line number Diff line number Diff line change
Expand Up @@ -265,7 +265,7 @@ class IKqpRunner : public TThrRefBase {

TIntrusivePtr<IKqpRunner> CreateKqpRunner(TIntrusivePtr<IKqpGateway> gateway, const TString& cluster,
const TIntrusivePtr<NYql::TTypeAnnotationContext>& typesCtx, const TIntrusivePtr<NYql::TKikimrSessionContext>& sessionCtx,
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry);
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry, TActorSystem* actorSystem);

TAutoPtr<NYql::IGraphTransformer> CreateKqpExplainPreparedTransformer(TIntrusivePtr<IKqpGateway> gateway,
const TString& cluster, TIntrusivePtr<TKqlTransformContext> transformCtx, const NMiniKQL::IFunctionRegistry* funcRegistry,
Expand Down
14 changes: 10 additions & 4 deletions ydb/core/kqp/host/kqp_runner.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
#include <ydb/core/kqp/opt/kqp_opt.h>
#include <ydb/core/kqp/opt/logical/kqp_opt_log.h>
#include <ydb/core/kqp/opt/kqp_statistics_transformer.h>
#include <ydb/core/kqp/opt/kqp_column_statistics_requester.h>
#include <ydb/core/kqp/opt/kqp_constant_folding_transformer.h>
#include <ydb/core/kqp/opt/logical/kqp_opt_cbo.h>

Expand Down Expand Up @@ -137,7 +138,8 @@ class TKqpRunner : public IKqpRunner {
public:
TKqpRunner(TIntrusivePtr<IKqpGateway> gateway, const TString& cluster,
const TIntrusivePtr<TTypeAnnotationContext>& typesCtx, const TIntrusivePtr<TKikimrSessionContext>& sessionCtx,
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry)
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry,
TActorSystem* actorSystem)
: Gateway(gateway)
, Cluster(cluster)
, TypesCtx(*typesCtx)
Expand All @@ -149,6 +151,7 @@ class TKqpRunner : public IKqpRunner {
sessionCtx->TablesPtr(), sessionCtx->GetUserRequestContext()))
, BuildQueryCtx(MakeIntrusive<TKqpBuildQueryContext>())
, Pctx(TKqpProviderContext(*OptimizeCtx, Config->CostBasedOptimizationLevel.Get().GetOrElse(Config->DefaultCostBasedOptimizationLevel)))
, ActorSystem(actorSystem)
{
CreateGraphTransformer(typesCtx, sessionCtx, funcRegistry);
}
Expand Down Expand Up @@ -297,6 +300,7 @@ class TKqpRunner : public IKqpRunner {
.AddPostTypeAnnotation(/* forSubgraph */ true)
.AddCommonOptimization()
.Add(CreateKqpConstantFoldingTransformer(OptimizeCtx, *typesCtx, Config), "ConstantFolding")
.Add(CreateKqpColumnStatisticsRequester(Config, *typesCtx, SessionCtx->Tables(), Cluster, ActorSystem), "ColumnStatisticsRequester")
.Add(CreateKqpStatisticsTransformer(OptimizeCtx, *typesCtx, Config, Pctx), "Statistics")
.Add(CreateKqpLogOptTransformer(OptimizeCtx, *typesCtx, Config), "LogicalOptimize")
.Add(CreateLogicalDataProposalsInspector(*typesCtx), "ProvidersLogicalOptimize")
Expand Down Expand Up @@ -324,7 +328,7 @@ class TKqpRunner : public IKqpRunner {
Config),
"BuildPhysicalTxs")
.Build(false));

auto physicalBuildQueryTransformer = TTransformationPipeline(typesCtx)
.AddServiceTransformers()
.Add(Log("PhysicalBuildQuery"), "LogPhysicalBuildQuery")
Expand Down Expand Up @@ -399,15 +403,17 @@ class TKqpRunner : public IKqpRunner {
TKqpProviderContext Pctx;

TAutoPtr<IGraphTransformer> Transformer;

TActorSystem* ActorSystem;
};

} // namespace

TIntrusivePtr<IKqpRunner> CreateKqpRunner(TIntrusivePtr<IKqpGateway> gateway, const TString& cluster,
const TIntrusivePtr<TTypeAnnotationContext>& typesCtx, const TIntrusivePtr<TKikimrSessionContext>& sessionCtx,
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry)
const TIntrusivePtr<TKqlTransformContext>& transformCtx, const NMiniKQL::IFunctionRegistry& funcRegistry, TActorSystem* actorSystem)
{
return new TKqpRunner(gateway, cluster, typesCtx, sessionCtx, transformCtx, funcRegistry);
return new TKqpRunner(gateway, cluster, typesCtx, sessionCtx, transformCtx, funcRegistry, actorSystem);
}

} // namespace NKqp
Expand Down
253 changes: 253 additions & 0 deletions ydb/core/kqp/opt/kqp_column_statistics_requester.cpp
Original file line number Diff line number Diff line change
@@ -0,0 +1,253 @@
#include "kqp_column_statistics_requester.h"

#include <ydb/library/yql/core/yql_expr_optimize.h>
#include <ydb/core/statistics/service/service.h>
#include <ydb/core/statistics/events.h>
#include <ydb/core/kqp/gateway/actors/kqp_ic_gateway_actors.h>
#include <ydb/library/yql/core/yql_statistics.h>
#include <ydb/library/yql/providers/dq/common/yql_dq_settings.h>
#include <ydb/library/yql/dq/opt/dq_opt_stat.h>
#include <ydb/library/yql/utils/log/log.h>

namespace NKikimr::NKqp {

using namespace NThreading;
using namespace NYql;

void TKqpColumnStatisticsRequester::PropagateTableToLambdaArgument(const TExprNode::TPtr& input) {
if (input->ChildrenSize() < 2) {
return;
}

auto callableInput = input->ChildRef(0);


for (size_t i = 1; i < input->ChildrenSize(); ++i) {
auto maybeLambda = TExprBase(input->ChildRef(i));
if (!maybeLambda.Maybe<TCoLambda>()) {
continue;
}

auto lambda = maybeLambda.Cast<TCoLambda>();
if (!lambda.Args().Size()){
continue;
}

if (callableInput->IsList()){
for (size_t j = 0; j < callableInput->ChildrenSize(); ++j){
KqpTableByExprNode[lambda.Args().Arg(j).Ptr()] = KqpTableByExprNode[callableInput->Child(j)];
}
} else {
KqpTableByExprNode[lambda.Args().Arg(0).Ptr()] = KqpTableByExprNode[callableInput.Get()];
}
}
}

IGraphTransformer::TStatus TKqpColumnStatisticsRequester::DoTransform(TExprNode::TPtr input, TExprNode::TPtr& output, TExprContext& ctx) {
Y_UNUSED(ctx);

output = input;
auto optLvl = Config->CostBasedOptimizationLevel.Get().GetOrElse(TDqSettings::TDefault::CostBasedOptimizationLevel);
auto enableColumnStats = Config->FeatureFlags.GetEnableColumnStatistics();
if (!(optLvl > 0 && enableColumnStats)) {
return IGraphTransformer::TStatus::Ok;
}

VisitExprLambdasLast(
input,
[&](const TExprNode::TPtr& input) {
BeforeLambdas(input) || BeforeLambdasUnmatched(input);

if (input->IsCallable()) {
PropagateTableToLambdaArgument(input);
}

return true;
},
[&](const TExprNode::TPtr& input) {
return AfterLambdas(input) || AfterLambdasUnmatched(input);
}
);

if (ColumnsByTableName.empty()) {
return IGraphTransformer::TStatus::Ok;
}

struct TTableMeta {
TString TableName;
THashMap<ui32, TString> ColumnNameByTag;
};
THashMap<TPathId, TTableMeta> tableMetaByPathId;

// TODO: Add other statistics, not only COUNT_MIN_SKETCH.
auto getStatisticsRequest = MakeHolder<NStat::TEvStatistics::TEvGetStatistics>();
getStatisticsRequest->StatType = NKikimr::NStat::EStatType::COUNT_MIN_SKETCH;

for (const auto& [table, columns]: ColumnsByTableName) {
auto tableMeta = Tables.GetTable(Cluster, table).Metadata;
auto& columnsMeta = tableMeta->Columns;

auto pathId = TPathId(tableMeta->PathId.OwnerId(), tableMeta->PathId.TableId());
for (const auto& column: columns) {
if (TypesCtx.ColumnStatisticsByTableName.contains(table) && TypesCtx.ColumnStatisticsByTableName[table]->Data.contains(column)) {
continue;
}

if (!columns.contains(column)) {
YQL_CLOG(DEBUG, ProviderKikimr) << "Table: " + table + " doesn't contain " + column + " to request for column statistics";
}

NKikimr::NStat::TRequest req;
req.ColumnTag = columnsMeta[column].Id;
req.PathId = pathId;
getStatisticsRequest->StatRequests.push_back(req);

tableMetaByPathId[pathId].TableName = table;
tableMetaByPathId[pathId].ColumnNameByTag[req.ColumnTag.value()] = column;
}
}

if (getStatisticsRequest->StatRequests.empty()) {
return IGraphTransformer::TStatus::Ok;
}

using TRequest = NStat::TEvStatistics::TEvGetStatistics;
using TResponse = NStat::TEvStatistics::TEvGetStatisticsResult;
struct TResult : public NYql::IKikimrGateway::TGenericResult {
THashMap<TString, TOptimizerStatistics::TColumnStatMap> columnStatisticsByTableName;
};

auto promise = NewPromise<TResult>();
auto callback = [tableMetaByPathId = std::move(tableMetaByPathId)]
(TPromise<TResult> promise, NStat::TEvStatistics::TEvGetStatisticsResult&& response) mutable {
if (!response.Success) {
promise.SetValue(NYql::NCommon::ResultFromError<TResult>("can't get column statistics!"));
}

THashMap<TString, TOptimizerStatistics::TColumnStatMap> columnStatisticsByTableName;

for (auto&& stat: response.StatResponses) {
auto meta = tableMetaByPathId[stat.Req.PathId];
auto columnName = meta.ColumnNameByTag[stat.Req.ColumnTag.value()];
auto& columnStatistics = columnStatisticsByTableName[meta.TableName].Data[columnName];
columnStatistics.CountMinSketch = std::move(stat.CountMinSketch.CountMin);
}

promise.SetValue(TResult{.columnStatisticsByTableName = std::move(columnStatisticsByTableName)});
};
auto statServiceId = NStat::MakeStatServiceID(ActorSystem->NodeId);
IActor* requestHandler =
new TActorRequestHandler<TRequest, TResponse, TResult>(statServiceId, getStatisticsRequest.Release(), promise, callback);
auto actorId = ActorSystem
->Register(requestHandler, TMailboxType::HTSwap, ActorSystem->AppData<TAppData>()->UserPoolId);
Y_UNUSED(actorId);

auto res = promise.GetFuture().GetValueSync();
if (!res.Issues().Empty()) {
TStringStream ss;
res.Issues().PrintTo(ss);
YQL_CLOG(DEBUG, ProviderKikimr) << "Can't load columns statistics for request: " << ss.Str();
return IGraphTransformer::TStatus::Ok;
}

for (auto&& [tableName, columnStatistics]: res.columnStatisticsByTableName) {
TypesCtx.ColumnStatisticsByTableName.insert(
{std::move(tableName), new TOptimizerStatistics::TColumnStatMap(std::move(columnStatistics))}
);
}

return IGraphTransformer::TStatus::Ok;
}

bool TKqpColumnStatisticsRequester::BeforeLambdas(const TExprNode::TPtr& input) {
bool matched = true;

if (TKqpTable::Match(input.Get())) {
KqpTableByExprNode[input.Get()] = input.Get();
} else if (auto maybeStreamLookup = TExprBase(input).Maybe<TKqpCnStreamLookup>()) {
KqpTableByExprNode[input.Get()] = maybeStreamLookup.Cast().Table().Ptr();
} else {
matched = false;
}

return matched;
}

bool TKqpColumnStatisticsRequester::BeforeLambdasUnmatched(const TExprNode::TPtr& input) {
for (const auto& node: input->Children()) {
if (KqpTableByExprNode.contains(node)) {
KqpTableByExprNode[input.Get()] = KqpTableByExprNode[node];
return true;
}
}

return true;
}

bool TKqpColumnStatisticsRequester::AfterLambdas(const TExprNode::TPtr& input) {
bool matched = true;

if (
TCoFilterBase::Match(input.Get()) ||
TCoFlatMapBase::Match(input.Get()) && IsPredicateFlatMap(TExprBase(input).Cast<TCoFlatMapBase>().Lambda().Body().Ref())
) {
std::shared_ptr<TOptimizerStatistics> dummyStats = nullptr;
auto computer = NDq::TPredicateSelectivityComputer(dummyStats, true);

if (TCoFilterBase::Match(input.Get())) {
computer.Compute(TExprBase(input).Cast<TCoFilterBase>().Lambda().Body());
} else if (TCoFlatMapBase::Match(input.Get())) {
computer.Compute(TExprBase(input).Cast<TCoFlatMapBase>().Lambda().Body());
} else {
Y_ENSURE(false);
}

auto columnStatsUsedMembers = computer.GetColumnStatsUsedMembers();
for (const auto& item: columnStatsUsedMembers.Data) {
auto exprNode = TExprBase(item.Member).Ptr();
if (!KqpTableByExprNode.contains(exprNode) || KqpTableByExprNode[exprNode] == nullptr) {
continue;
}

auto table = TExprBase(KqpTableByExprNode[exprNode]).Cast<TKqpTable>().Path().StringValue();
auto column = item.Member.Name().StringValue();
size_t pointPos = column.find('.'); // table.column
if (pointPos != TString::npos) {
column = column.substr(pointPos + 1);
}

ColumnsByTableName[table].insert(std::move(column));
}
} else {
matched = false;
}

return matched;
}

bool TKqpColumnStatisticsRequester::AfterLambdasUnmatched(const TExprNode::TPtr& input) {
if (KqpTableByExprNode.contains(input.Get())) {
return true;
}

for (const auto& node: input->Children()) {
if (KqpTableByExprNode.contains(node)) {
KqpTableByExprNode[input.Get()] = KqpTableByExprNode[node];
return true;
}
}

return true;
}

TAutoPtr<IGraphTransformer> CreateKqpColumnStatisticsRequester(
const TKikimrConfiguration::TPtr& config,
TTypeAnnotationContext& typesCtx,
TKikimrTablesData& tables,
TString cluster,
TActorSystem* actorSystem
) {
return THolder<IGraphTransformer>(new TKqpColumnStatisticsRequester(config, typesCtx, tables, cluster, actorSystem));
}

} // end of NKikimr::NKqp
Loading
Loading