database/mariadb-columnstore-engine
1
0
Fork 0
mirror of https://github.com/mariadb-corporation/mariadb-columnstore-engine.git synced 2025-11-02 06:13:16 +03:00
Code Activity

feat(rbo,rules): refactored statistics storage in gwi and implemented statistics based UNION rewrite.

Browse Source
This commit is contained in:
drrtuy
2025-07-17 15:37:20 +00:00
parent 67295b4320
commit 15be33fbc5
6 changed files with 122 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

91
dbcon/mysql/rulebased_optimizer.cpp
View File

@@ -191,9 +191,24 @@ execplan::ParseTree* filtersWithNewRangeAddedIfNeeded(execplan::SCSEP& csep,
return ptp;
}
execplan::SimpleColumn* findSuitableKeyColumn(execplan::CalpontSelectExecutionPlan& csep)
// Looking for a projected column that comes first in an available index and has EI statistics
// INV nullptr signifies that no suitable column was found
execplan::SimpleColumn* findSuitableKeyColumn(execplan::CalpontSelectExecutionPlan& csep, optimizer::RBOptimizerContext& ctx)
{
return dynamic_cast<execplan::SimpleColumn*>(csep.returnedCols().front().get());
for (auto& rc : csep.returnedCols())
{
auto* simpleColumn = dynamic_cast<execplan::SimpleColumn*>(rc.get());
if (simpleColumn)
{
std::cout << "Found simple column " << simpleColumn->columnName() << std::endl;
cal_impl_if::SchemaAndTableName schemaAndTableNam = {simpleColumn->tableName(), simpleColumn->columnName()};
auto columnStatistics = ctx.gwi.findStatisticsForATable(schemaAndTableNam);
return simpleColumn;
}
}
return nullptr;
}
execplan::CalpontSelectExecutionPlan::SelectList makeUnionFromTable(
@@ -201,49 +216,65 @@ execplan::CalpontSelectExecutionPlan::SelectList makeUnionFromTable(
{
execplan::CalpontSelectExecutionPlan::SelectList unionVec;
// unionVec.reserve(numberOfLegs);
execplan::SimpleColumn* keyColumn = findSuitableKeyColumn(csep);
std::cout << "looking for " << keyColumn->columnName() << " in ctx.gwi.columnStatisticsMap "
<< " with size " << ctx.gwi.columnStatisticsMap.size() << std::endl;
for (auto& [k, v] : ctx.gwi.columnStatisticsMap)
{
std::cout << "key " << k << " vector size " << v.size() << std::endl;
}
if (!keyColumn ||
ctx.gwi.columnStatisticsMap.find(keyColumn->columnName()) == ctx.gwi.columnStatisticsMap.end())
execplan::SimpleColumn* keyColumn = findSuitableKeyColumn(csep, ctx);
if (!keyColumn)
{
return unionVec;
}
auto columnStatistics = ctx.gwi.columnStatisticsMap[keyColumn->columnName()];
std::cout << "columnStatistics.size() " << columnStatistics.size() << std::endl;
std::cout << "looking for " << keyColumn->columnName() << " in ctx.gwi.tableStatisticsMap "
<< " with size " << ctx.gwi.tableStatisticsMap.size() << std::endl;
for (auto& [k, v] : ctx.gwi.tableStatisticsMap)
{
std::cout << "SchemaAndTableName " << k.schema << "." << k.table << " column map size " << v.size() << std::endl;
}
cal_impl_if::SchemaAndTableName schemaAndTableName = {keyColumn->schemaName(), keyColumn->tableName()};
auto tableColumnsStatisticsIt = ctx.gwi.tableStatisticsMap.find(schemaAndTableName);
if (tableColumnsStatisticsIt == ctx.gwi.tableStatisticsMap.end())
{
return unionVec;
}
auto columnStatisticsIt = tableColumnsStatisticsIt->second.find(keyColumn->columnName());
if (columnStatisticsIt == tableColumnsStatisticsIt->second.end())
{
return unionVec;
}
auto columnStatistics = columnStatisticsIt->second;
std::cout << "Histogram_json_hb histogram size " << columnStatistics.get_json_histogram().size() << std::endl;
// TODO char and other numerical types support
size_t numberOfUnionUnits = 2;
size_t numberOfBucketsPerUnionUnit = columnStatistics.size() / numberOfUnionUnits;
size_t numberOfUnionUnits = std::min(columnStatistics.get_json_histogram().size(), 16UL);
size_t numberOfBucketsPerUnionUnit = columnStatistics.get_json_histogram().size() / numberOfUnionUnits;
std::vector<std::pair<uint64_t, uint64_t>> bounds;
// TODO need to process tail if number of buckets is not divisible by number of union units
// TODO non-overlapping buckets if it is a problem at all
for (size_t i = 0; i < numberOfUnionUnits; ++i)
for (size_t i = 0; i < numberOfUnionUnits - 1; ++i)
{
auto bucket = columnStatistics.begin() + i * numberOfBucketsPerUnionUnit;
auto endBucket = columnStatistics.begin() + (i + 1) * numberOfBucketsPerUnionUnit;
// TODO find a median b/w the current bucket start and the previous bucket end
uint64_t currentLowerBound =
(bounds.empty() ? *(uint32_t*)bucket->start_value.data()
: std::min((uint64_t)*(uint32_t*)bucket->start_value.data(), bounds.back().second));
uint64_t currentUpperBound = currentLowerBound;
for (; bucket != endBucket; ++bucket)
{
uint64_t bucketLowerBound = *(uint32_t*)bucket->start_value.data();
std::cout << "bucket.start_value " << bucketLowerBound << std::endl;
currentUpperBound = bucketLowerBound + bucket->ndv;
}
auto bucket = columnStatistics.get_json_histogram().begin() + i * numberOfBucketsPerUnionUnit;
auto endBucket = columnStatistics.get_json_histogram().begin() + (i + 1) * numberOfBucketsPerUnionUnit;
uint64_t currentLowerBound = *(uint32_t*)bucket->start_value.data();
uint64_t currentUpperBound = *(uint32_t*)endBucket->start_value.data();
std::cout << "currentLowerBound " << currentLowerBound << " currentUpperBound " << currentUpperBound
<< std::endl;
bounds.push_back(std::make_pair(currentLowerBound, currentUpperBound));
}
// Add last range
auto lastBucket = columnStatistics.get_json_histogram().begin() + (numberOfUnionUnits - 1) * numberOfBucketsPerUnionUnit;
uint64_t currentLowerBound = *(uint32_t*)lastBucket->start_value.data();
uint64_t currentUpperBound = *(uint32_t*)columnStatistics.get_last_bucket_end_endp().data();
std::cout << "last currentLowerBound " << currentLowerBound << " last currentUpperBound " << currentUpperBound
<< std::endl;
bounds.push_back(std::make_pair(currentLowerBound, currentUpperBound));
for (auto& bound : bounds)
{
auto clonedCSEP = csep.cloneWORecursiveSelects();
@@ -275,8 +306,6 @@ void applyParallelCES(execplan::CalpontSelectExecutionPlan& csep, RBOptimizerCon
derivedSCEP->subType(execplan::CalpontSelectExecutionPlan::FROM_SUBS);
derivedSCEP->derivedTbAlias(tableAlias);
// TODO: hardcoded for now
// size_t parallelFactor = 2;
// Create a copy of the current leaf CSEP with additional filters to partition the key column
auto additionalUnionVec = makeUnionFromTable(csep, ctx);
derivedSCEP->unionVec().insert(derivedSCEP->unionVec().end(), additionalUnionVec.begin(),