前言

经过前两个实验的铺垫，终于到了执行 SQL 语句的时候了。这篇博客将会介绍 SQL 执行计划实验的实现过程，下面进入正题。

总体架构

一条 SQL 语句的处理流程可以归纳为：

1. SQL 被 Parser 解析为抽象语法树 AST
2. Binber 将 AST转换为 Bustub 可以理解的更高级的 AST
3. Tree rewriter 将语法树转换为逻辑执行计划
4. Optimizer 生成最终要执行的物理执行计划
5. 执行引擎执行物理执行计划，返回查询结果

物理执行计划定义了具体的执行方式，比如逻辑计划中的 Join 可以被替换为 Nest loop join、 Hash join 或者 Index join。由于 Fall 2020 版本的代码没有 Parse 和 Optimizer，所以测试用例中都是手动构造的物理执行计划。

系统目录

数据库会维护一个内部目录，以跟踪有关数据库的元数据。目录中可以存放数据表的信息、索引信息和统计数据。Bustub 中使用 Catalog 类表示系统目录，内部存放 table_oid_t 到 TableMetadata 的映射表以及 index_oid_t 到 IndexInfo 的映射表。

TableMetadata 描述了一张表的信息，包括表名、Schema、表 id 和表的指针。代码如下所示：

struct TableMetadata {
  TableMetadata(Schema schema, std::string name, std::unique_ptr<TableHeap> &&table, table_oid_t oid)
      : schema_(std::move(schema)), name_(std::move(name)), table_(std::move(table)), oid_(oid) {}
  Schema schema_;
  std::string name_;
  std::unique_ptr<TableHeap> table_;
  table_oid_t oid_;
};

TableHeap 代表了一张表，实现了 tuple 的增删改查操作。它的内部存放了第一个表页 TablePage 的 id，由于每个 TablePage 都会存放前一个和下一个表页的 id，这样就将表组织为双向链表，可以通过 TableIterator 进行迭代。

TablePage 使用分槽页结构（slotted page），tuple 从后往前插入，每个 tuple 由一个 RID 标识。

class RID {
 public:
  RID() = default;

  /**
   * Creates a new Record Identifier for the given page identifier and slot number.
   */
  RID(page_id_t page_id, uint32_t slot_num) : page_id_(page_id), slot_num_(slot_num) {}

  explicit RID(int64_t rid) : page_id_(static_cast<page_id_t>(rid >> 32)), slot_num_(static_cast<uint32_t>(rid)) {}

  inline int64_t Get() const { return (static_cast<int64_t>(page_id_)) << 32 | slot_num_; }

  inline page_id_t GetPageId() const { return page_id_; }

  inline uint32_t GetSlotNum() const { return slot_num_; }

  bool operator==(const RID &other) const { return page_id_ == other.page_id_ && slot_num_ == other.slot_num_; }

 private:
  page_id_t page_id_{INVALID_PAGE_ID};
  uint32_t slot_num_{0};  // logical offset from 0, 1...
};

Catalog 中有三个与表相关的方法：CreateTable、GetTable(const std::string &table_name) 和 GetTable(table_oid_t table_oid)，第一个方法用于创建一个新的表，后面两个方法用于获取表元数据：

/**
* Create a new table and return its metadata.
* @param txn the transaction in which the table is being created
* @param table_name the name of the new table
* @param schema the schema of the new table
* @return a pointer to the metadata of the new table
*/
TableMetadata *CreateTable(Transaction *txn, const std::string &table_name, const Schema &schema) {
  BUSTUB_ASSERT(names_.count(table_name) == 0, "Table names should be unique!");
  auto tid = next_table_oid_++;

  auto table_heap = std::make_unique<TableHeap>(bpm_, lock_manager_, log_manager_, txn);
  tables_[tid] = std::make_unique<TableMetadata>(schema, table_name, std::move(table_heap), tid);
  names_[table_name] = tid;

  return tables_[tid].get();
}

/** @return table metadata by name */
TableMetadata *GetTable(const std::string &table_name) {
  auto it = names_.find(table_name);
  if (it == names_.end()) {
    throw std::out_of_range("Table is not found");
  }

  return tables_[it->second].get();
}

/** @return table metadata by oid */
TableMetadata *GetTable(table_oid_t table_oid) {
  auto it = tables_.find(table_oid);
  if (it == tables_.end()) {
    throw std::out_of_range("Table is not found");
  }

  return it->second.get();
}

创建索引

Catalog 使用 CreateIndex() 方法创建索引，创建的时候需要将表中的数据转换为键值对插入索引中：

/**
  * Create a new index, populate existing data of the table and return its metadata.
  * @param txn the transaction in which the table is being created
  * @param index_name the name of the new index
  * @param table_name the name of the table
  * @param schema the schema of the table
  * @param key_schema the schema of the key
  * @param key_attrs key attributes
  * @param keysize size of the key
  * @return a pointer to the metadata of the new table
  */
template <class KeyType, class ValueType, class KeyComparator>
IndexInfo *CreateIndex(Transaction *txn, const std::string &index_name, const std::string &table_name,
                        const Schema &schema, const Schema &key_schema, const std::vector<uint32_t> &key_attrs,
                        size_t keysize) {
  BUSTUB_ASSERT(index_names_.count(index_name) == 0, "Index names should be unique!");
  auto id = next_index_oid_++;

  auto meta = new IndexMetadata(index_name, table_name, &schema, key_attrs);
  auto index = std::make_unique<BPLUSTREE_INDEX_TYPE>(meta, bpm_);

  // 初始化索引
  auto table = GetTable(table_name)->table_.get();
  for (auto it = table->Begin(txn); it != table->End(); ++it) {
    index->InsertEntry(it->KeyFromTuple(schema, key_schema, key_attrs), it->GetRid(), txn);
  }

  indexes_[id] = std::make_unique<IndexInfo>(key_schema, index_name, std::move(index), id, table_name, keysize);
  index_names_[table_name][index_name] = id;
  return indexes_[id].get();
}

数据库中有多个表，一个表可以拥有多个索引，但是每个索引对应一个全局唯一的 index_oid_t：

IndexInfo *GetIndex(const std::string &index_name, const std::string &table_name) {
  auto it = index_names_.find(table_name);
  if (it == index_names_.end()) {
    throw std::out_of_range("Table is not found");
  }

  auto iit = it->second.find(index_name);
  if (iit == it->second.end()) {
    throw std::out_of_range("Index is not found");
  }

  return indexes_[iit->second].get();
}

IndexInfo *GetIndex(index_oid_t index_oid) {
  auto it = indexes_.find(index_oid);
  if (it == indexes_.end()) {
    throw std::out_of_range("Index is not found");
  }

  return it->second.get();
}

std::vector<IndexInfo *> GetTableIndexes(const std::string &table_name) {
  auto it = index_names_.find(table_name);
  if (it == index_names_.end()) {
    return {};
  };

  std::vector<IndexInfo *> indexes;
  for (auto &[name, id] : it->second) {
    indexes.push_back(GetIndex(id));
  }

  return indexes;
}

执行器

如下图的右下角所示，执行计划由一系列算子组合而成，每个算子可以拥有自己的子算子，数据从子算子流向父算子，最终从根节点输出执行结果。执行计划有三种执行模型：

• 迭代模型：每个算子都会实现 Next() 方法，父算子调用子算子的 Next() 方法获取一条记录，外部通过不断调用根节点的 Next() 方法直至没有更多数据输出。这种方法的优点就是一次只产生一条 Tuple，内存占用小

• 物化模型：每个算子一次性返回所有记录

• 向量模型：迭代模型和物化模型的折中版本，一次返回一批数据

本次实验使用迭代模型，伪代码如下图所示：

Bustub 使用执行引擎 ExecutionEngine 执行物理计划，这个类的代码很简洁，只有一个 Execute() 方法。可以看到这个方法会先将执行计划转换为对应的执行器 executor，使用 Init() 初始化后循环调用 executor 的 Next() 方法获取查询结果：

class ExecutionEngine {
 public:
  ExecutionEngine(BufferPoolManager *bpm, TransactionManager *txn_mgr, Catalog *catalog)
      : bpm_(bpm), txn_mgr_(txn_mgr), catalog_(catalog) {}

  DISALLOW_COPY_AND_MOVE(ExecutionEngine);

  bool Execute(const AbstractPlanNode *plan, std::vector<Tuple> *result_set, Transaction *txn,
               ExecutorContext *exec_ctx) {
    // construct executor
    auto executor = ExecutorFactory::CreateExecutor(exec_ctx, plan);

    // prepare
    executor->Init();

    // execute
    try {
      Tuple tuple;
      RID rid;
      while (executor->Next(&tuple, &rid)) {
        if (result_set != nullptr) {
          result_set->push_back(tuple);
        }
      }
    } catch (Exception &e) {
      // TODO(student): handle exceptions
    }

    return true;
  }

 private:
  [[maybe_unused]] BufferPoolManager *bpm_;
  [[maybe_unused]] TransactionManager *txn_mgr_;
  [[maybe_unused]] Catalog *catalog_;
};

SeqScanExecutor 用于进行全表扫描操作，内部带有 SeqScanPlan 执行计划：

/**
 * SeqScanExecutor executes a sequential scan over a table.
 */
class SeqScanExecutor : public AbstractExecutor {
 public:
  /**
   * Creates a new sequential scan executor.
   * @param exec_ctx the executor context
   * @param plan the sequential scan plan to be executed
   */
  SeqScanExecutor(ExecutorContext *exec_ctx, const SeqScanPlanNode *plan);

  void Init() override;

  bool Next(Tuple *tuple, RID *rid) override;

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); }

 private:
  /** The sequential scan plan node to be executed. */
  const SeqScanPlanNode *plan_;
  TableMetadata *table_metadata_;
  TableIterator it_;
};

SeqScanPlan 声明如下，Schema *output 指明了输出列，table_oid 代表被扫描的表，而 AbstractExpression *predicate 代表谓词算子：

/**
 * SeqScanPlanNode identifies a table that should be scanned with an optional predicate.
 */
class SeqScanPlanNode : public AbstractPlanNode {
 public:
  /**
   * Creates a new sequential scan plan node.
   * @param output the output format of this scan plan node
   * @param predicate the predicate to scan with, tuples are returned if predicate(tuple) = true or predicate = nullptr
   * @param table_oid the identifier of table to be scanned
   */
  SeqScanPlanNode(const Schema *output, const AbstractExpression *predicate, table_oid_t table_oid)
      : AbstractPlanNode(output, {}), predicate_{predicate}, table_oid_(table_oid) {}

  PlanType GetType() const override { return PlanType::SeqScan; }

  /** @return the predicate to test tuples against; tuples should only be returned if they evaluate to true */
  const AbstractExpression *GetPredicate() const { return predicate_; }

  /** @return the identifier of the table that should be scanned */
  table_oid_t GetTableOid() const { return table_oid_; }

 private:
  /** The predicate that all returned tuples must satisfy. */
  const AbstractExpression *predicate_;
  /** The table whose tuples should be scanned. */
  table_oid_t table_oid_;
};

举个栗子，SELECT name, age FROM t_student WHERE age > 16 的 age > 16 部分就是 predicate ，实际数据类型为 ComparisonExpression ，而 predicate 又由 ColumnValueExpression（代表 age 列的值） 和 ConstantValueExpression（代表 16）组成。

要实现全表扫描只需在 Next 函数中判断迭代器所指的 tuple 是否满足查询条件并递增迭代器，如果满足条件就返回该 tuple，不满足就接着迭代。

SeqScanExecutor::SeqScanExecutor(ExecutorContext *exec_ctx, const SeqScanPlanNode *plan)
    : AbstractExecutor(exec_ctx),
      plan_(plan), table_metadata_(exec_ctx->GetCatalog()->GetTable(plan->GetTableOid())) {}

void SeqScanExecutor::Init() { it_ = table_metadata_->table_->Begin(exec_ctx_->GetTransaction()); }

bool SeqScanExecutor::Next(Tuple *tuple, RID *rid) {
  auto predicate = plan_->GetPredicate();

  while (it_ != table_metadata_->table_->End()) {
    *tuple = *it_++;
    *rid = tuple->GetRid();

    if (!predicate || predicate->Evaluate(tuple, &table_metadata_->schema_).GetAs<bool>()) {
      // 只保留输出列
      std::vector<Value> values;
      for (auto &col : GetOutputSchema()->GetColumns()) {
        values.push_back(col.GetExpr()->Evaluate(tuple, &table_metadata_->schema_));
      }

      *tuple = {values, GetOutputSchema()};
      return true;
    }
  }

  return false;
}

测试用例中通过下述代码手动构造出 SELECT colA, colB FROM test_1 WHERE colA < 500 的全表扫描执行计划并执行：

// Construct query plan
TableMetadata *table_info = GetExecutorContext()->GetCatalog()->GetTable("test_1");
Schema &schema = table_info->schema_;
auto *colA = MakeColumnValueExpression(schema, 0, "colA");
auto *colB = MakeColumnValueExpression(schema, 0, "colB");
auto *const500 = MakeConstantValueExpression(ValueFactory::GetIntegerValue(500));
auto *predicate = MakeComparisonExpression(colA, const500, ComparisonType::LessThan);
auto *out_schema = MakeOutputSchema({{"colA", colA}, {"colB", colB}});
SeqScanPlanNode plan{out_schema, predicate, table_info->oid_};

// Execute
std::vector<Tuple> result_set;
GetExecutionEngine()->Execute(&plan, &result_set, GetTxn(), GetExecutorContext());

上一节中实现了 B+ 树索引，使用索引可以大大减小查询范围，大大加快查询速度。由于 IndexScanExecutor 不是模板类，所以这里使用的 KeyType 为 GenericKey<8>，KeyComparator 为 GenericComparator<8>：

#define B_PLUS_TREE_INDEX_ITERATOR_TYPE IndexIterator<GenericKey<8>, RID, GenericComparator<8>>
#define B_PLUS_TREE_INDEX_TYPE BPlusTreeIndex<GenericKey<8>, RID, GenericComparator<8>>

class IndexScanExecutor : public AbstractExecutor {
 public:
  /**
   * Creates a new index scan executor.
   * @param exec_ctx the executor context
   * @param plan the index scan plan to be executed
   */
  IndexScanExecutor(ExecutorContext *exec_ctx, const IndexScanPlanNode *plan);

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); };

  void Init() override;

  bool Next(Tuple *tuple, RID *rid) override;

 private:
  /** The index scan plan node to be executed. */
  const IndexScanPlanNode *plan_;
  IndexInfo *index_info_;
  B_PLUS_TREE_INDEX_TYPE *index_;
  TableMetadata *table_metadata_;
  B_PLUS_TREE_INDEX_ITERATOR_TYPE it_;
};

索引扫描的代码和全表扫描几乎一样，只是迭代器换成了 B+ 树的迭代器：

IndexScanExecutor::IndexScanExecutor(ExecutorContext *exec_ctx, const IndexScanPlanNode *plan)
    : AbstractExecutor(exec_ctx),
      plan_(plan),
      index_info_(exec_ctx->GetCatalog()->GetIndex(plan->GetIndexOid())),
      index_(dynamic_cast<B_PLUS_TREE_INDEX_TYPE *>(index_info_->index_.get())),
      table_metadata_(exec_ctx->GetCatalog()->GetTable(index_info_->table_name_)) {}

void IndexScanExecutor::Init() { it_ = index_->GetBeginIterator(); }

bool IndexScanExecutor::Next(Tuple *tuple, RID *rid) {
  auto predicate = plan_->GetPredicate();

  while (it_ != index_->GetEndIterator()) {
    *rid = (*it_).second;
    table_metadata_->table_->GetTuple(*rid, tuple, exec_ctx_->GetTransaction());
    ++it_;

    if (!predicate || predicate->Evaluate(tuple, &table_metadata_->schema_).GetAs<bool>()) {
      // 只保留输出列
      std::vector<Value> values;
      for (auto &col : GetOutputSchema()->GetColumns()) {
        values.push_back(col.GetExpr()->Evaluate(tuple, &table_metadata_->schema_));
      }

      *tuple = {values, GetOutputSchema()};
      return true;
    }
  }

  return false;
}

插入操作分为两种：

• raw inserts：插入数据直接来自插入执行器本身，比如 INSERT INTO tbl_user VALUES (1, 15), (2, 16)
• not-raw inserts：插入的数据来自子执行器，比如 INSERT INTO tbl_user1 SELECT * FROM tbl_user2

可以使用插入计划的 IsRawInsert() 判断插入操作的类型，这个函数根据子查询器列表是否为空进行判断：

/** @return true if we embed insert values directly into the plan, false if we have a child plan providing tuples */
bool IsRawInsert() const { return GetChildren().empty(); }

如果是 raw inserts，我们直接根据插入执行器中的数据构造 tuple 并插入表中，否则调用子执行器的 Next 函数获取数据并插入表中。因为表中可能建了索引，所以插入数据之后需要更新索引：

class InsertExecutor : public AbstractExecutor {
 public:
  /**
   * Creates a new insert executor.
   * @param exec_ctx the executor context
   * @param plan the insert plan to be executed
   * @param child_executor the child executor to obtain insert values from, can be nullptr
   */
  InsertExecutor(ExecutorContext *exec_ctx, const InsertPlanNode *plan,
                 std::unique_ptr<AbstractExecutor> &&child_executor);

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); };

  void Init() override;

  // Note that Insert does not make use of the tuple pointer being passed in.
  // We return false if the insert failed for any reason, and return true if all inserts succeeded.
  bool Next([[maybe_unused]] Tuple *tuple, RID *rid) override;

  void InsertTuple(Tuple *tuple, RID *rid);

 private:
  /** The insert plan node to be executed. */
  const InsertPlanNode *plan_;
  std::unique_ptr<AbstractExecutor> child_executor_;
  TableMetadata *table_metadata_;
  std::vector<IndexInfo *> index_infos_;
  uint32_t index_{0};
};

InsertExecutor::InsertExecutor(ExecutorContext *exec_ctx, const InsertPlanNode *plan,
                               std::unique_ptr<AbstractExecutor> &&child_executor)
    : AbstractExecutor(exec_ctx),
      plan_(plan),
      child_executor_(std::move(child_executor)),
      table_metadata_(exec_ctx->GetCatalog()->GetTable(plan->TableOid())),
      index_infos_(exec_ctx->GetCatalog()->GetTableIndexes(table_metadata_->name_)) {}

void InsertExecutor::Init() {
  if (!plan_->IsRawInsert()) {
    child_executor_->Init();
  }
}

bool InsertExecutor::Next([[maybe_unused]] Tuple *tuple, RID *rid) {
  if (plan_->IsRawInsert()) {
    if (index_ >= plan_->RawValues().size()) {
      return false;
    }

    *tuple = {plan_->RawValuesAt(index_++), &table_metadata_->schema_};
    InsertTuple(tuple, rid);
    return true;
  } else {
    auto has_data = child_executor_->Next(tuple, rid);
    if (has_data) {
      InsertTuple(tuple, rid);
    }
    return has_data;
  }
}

void InsertExecutor::InsertTuple(Tuple *tuple, RID *rid) {
  // 更新数据表
  table_metadata_->table_->InsertTuple(*tuple, rid, exec_ctx_->GetTransaction());

  // 更新索引
  for (auto &index_info : index_infos_) {
    index_info->index_->InsertEntry(
        tuple->KeyFromTuple(table_metadata_->schema_, index_info->key_schema_, index_info->index_->GetKeyAttrs()), *rid,
        exec_ctx_->GetTransaction());
  }
}

UpdateExecutor 从子执行器获取需要更新的 tuple，并调用 GenerateUpdatedTuple 生成更新之后的 tuple，同样也要更新索引。

class UpdateExecutor : public AbstractExecutor {
  friend class UpdatePlanNode;

 public:
  UpdateExecutor(ExecutorContext *exec_ctx, const UpdatePlanNode *plan,
                 std::unique_ptr<AbstractExecutor> &&child_executor);

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); };

  void Init() override;

  bool Next([[maybe_unused]] Tuple *tuple, RID *rid) override;

  /* Given an old tuple, creates a new updated tuple based on the updateinfo given in the plan */
  Tuple GenerateUpdatedTuple(const Tuple &old_tup);

 private:
  const UpdatePlanNode *plan_;
  const TableMetadata *table_info_;
  std::unique_ptr<AbstractExecutor> child_executor_;
  std::vector<IndexInfo *> index_infos_;
};

bool UpdateExecutor::Next([[maybe_unused]] Tuple *tuple, RID *rid) {
  if (!child_executor_->Next(tuple, rid)) {
    return false;
  }

  // 更新数据表
  auto new_tuple = GenerateUpdatedTuple(*tuple);
  table_info_->table_->UpdateTuple(new_tuple, *rid, exec_ctx_->GetTransaction());

  // 更新索引
  for (auto &index_info : index_infos_) {
    // 删除旧的 tuple
    index_info->index_->DeleteEntry(
        tuple->KeyFromTuple(table_info_->schema_, index_info->key_schema_, index_info->index_->GetKeyAttrs()), *rid,
        exec_ctx_->GetTransaction());

    // 插入新的 tuple
    index_info->index_->InsertEntry(
        new_tuple.KeyFromTuple(table_info_->schema_, index_info->key_schema_, index_info->index_->GetKeyAttrs()), *rid,
        exec_ctx_->GetTransaction());
  }

  return true;
}

DeleteExecutor 的数据来自于子执行器，删除之后需要更新索引。

DeleteExecutor::DeleteExecutor(ExecutorContext *exec_ctx, const DeletePlanNode *plan,
                               std::unique_ptr<AbstractExecutor> &&child_executor)
    : AbstractExecutor(exec_ctx),
      plan_(plan),
      child_executor_(std::move(child_executor)),
      table_metadata_(exec_ctx->GetCatalog()->GetTable(plan->TableOid())),
      index_infos_(exec_ctx->GetCatalog()->GetTableIndexes(table_metadata_->name_)) {}

void DeleteExecutor::Init() { child_executor_->Init(); }

bool DeleteExecutor::Next([[maybe_unused]] Tuple *tuple, RID *rid) {
  if (!child_executor_->Next(tuple, rid)) {
    return false;
  }

  table_metadata_->table_->MarkDelete(*rid, exec_ctx_->GetTransaction());

  // 更新索引
  for (auto &index_info : index_infos_) {
    index_info->index_->DeleteEntry(
        tuple->KeyFromTuple(table_metadata_->schema_, index_info->key_schema_, index_info->index_->GetKeyAttrs()), *rid,
        exec_ctx_->GetTransaction());
  }

  return true;
}

要实现连接操作，最简单粗暴的方法就是开个二重循环，外层循环是小表（指的是数据页较少），内层循环是大表，小表驱动大表。但是这种连接方法效率非常低，因为完全无法利用到缓存池（分块变成四重循环之后效果会好一些）：

假设一次磁盘 IO 的时间是 0.1ms，那么大表驱动小表耗时 1.3 小时，小表驱动大表耗时 1.1 小时，可见速度慢的感人。

循环嵌套连接执行器 NestLoopJoinExecutor 的声明如下，可以看到数据成员包括 left_executor_ 和 right_executor，前者代表外表执行器，后者代表内表的执行器：

class NestedLoopJoinExecutor : public AbstractExecutor {
 public:
  /**
   * Creates a new NestedLoop join executor.
   * @param exec_ctx the executor context
   * @param plan the NestedLoop join plan to be executed
   * @param left_executor the child executor that produces tuple for the left side of join
   * @param right_executor the child executor that produces tuple for the right side of join
   *
   */
  NestedLoopJoinExecutor(ExecutorContext *exec_ctx, const NestedLoopJoinPlanNode *plan,
                         std::unique_ptr<AbstractExecutor> &&left_executor,
                         std::unique_ptr<AbstractExecutor> &&right_executor);

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); };

  void Init() override;

  bool Next(Tuple *tuple, RID *rid) override;

 private:
  /** The NestedLoop plan node to be executed. */
  const NestedLoopJoinPlanNode *plan_;
  std::unique_ptr<AbstractExecutor> left_executor_;
  std::unique_ptr<AbstractExecutor> right_executor_;
  Tuple left_tuple_;
  bool is_done_;
};

由于一次只能返回一个 tuple，所以需要先保存外表的一个 tuple，然后循环调用内表执行器的 Next() 方法直至匹配，当内表遍历完一遍之后需要更新外表的 tuple。这个部分的代码写的比较奇怪，如果有 python 的 yield 关键字可能会好写很多：

void NestedLoopJoinExecutor::Init() {
  left_executor_->Init();
  right_executor_->Init();

  RID left_rid;
  is_done_ = !left_executor_->Next(&left_tuple_, &left_rid);
}

bool NestedLoopJoinExecutor::Next(Tuple *tuple, RID *rid) {
  Tuple right_tuple;
  RID right_rid, left_rid;
  auto predicate = plan_->Predicate();
  auto left_schema = left_executor_->GetOutputSchema();
  auto right_schema = right_executor_->GetOutputSchema();

  while (!is_done_) {
    while (right_executor_->Next(&right_tuple, &right_rid)) {
      if (!predicate || predicate->EvaluateJoin(&left_tuple_, left_schema, &right_tuple, right_schema).GetAs<bool>()) {
        // 拼接 tuple
        std::vector<Value> values;
        for (auto &col : GetOutputSchema()->GetColumns()) {
          values.push_back(col.GetExpr()->EvaluateJoin(&left_tuple_, left_schema, &right_tuple, right_schema));
        }

        *tuple = {values, GetOutputSchema()};
        return true;
      }
    }

    is_done_ = !left_executor_->Next(&left_tuple_, &left_rid);
    right_executor_->Init();
  }

  return false;
}

索引循环连接可以减少内表的扫描范围和磁盘 IO 次数，大大提升连接效率。假设走一次索引的 IO 次数为常数 \(C \ll m\)，那么总共只需 \(M+m \cdot C\) 次 IO：

嵌套循环执行器 NestIndexJoinExecutor 的声明如下，child_executor_ 是外表的执行器，内表的数据由索引提供，所以不需要内表的执行器：

class NestIndexJoinExecutor : public AbstractExecutor {
 public:
  NestIndexJoinExecutor(ExecutorContext *exec_ctx, const NestedIndexJoinPlanNode *plan,
                        std::unique_ptr<AbstractExecutor> &&child_executor);

  const Schema *GetOutputSchema() override { return plan_->OutputSchema(); }

  void Init() override;

  bool Next(Tuple *tuple, RID *rid) override;

 private:
  /** The nested index join plan node. */
  const NestedIndexJoinPlanNode *plan_;
  std::unique_ptr<AbstractExecutor> child_executor_;
  TableMetadata *inner_table_info_;
  IndexInfo *index_info_;
  Tuple left_tuple_;
  std::vector<RID> inner_result_;
};

在索引上寻找匹配值时需要将 left_tuple_ 转换为内表索引的 key：

bool NestIndexJoinExecutor::Next(Tuple *tuple, RID *rid) {
  Tuple right_tuple;
  RID left_rid, right_rid;

  auto left_schema = plan_->OuterTableSchema();
  auto right_schema = plan_->InnerTableSchema();

  while (true) {
    if (!inner_result_.empty()) {
      right_rid = inner_result_.back();
      inner_result_.pop_back();
      inner_table_info_->table_->GetTuple(right_rid, &right_tuple, exec_ctx_->GetTransaction());

      // 拼接 tuple
      std::vector<Value> values;
      for (auto &col : GetOutputSchema()->GetColumns()) {
        values.push_back(col.GetExpr()->EvaluateJoin(&left_tuple_, left_schema, &right_tuple, right_schema));
      }

      *tuple = {values, GetOutputSchema()};
      return true;
    }

    if (!child_executor_->Next(&left_tuple_, &left_rid)) {
      return false;
    }

    // 在内表的索引上寻找匹配值列表
    auto value = plan_->Predicate()->GetChildAt(0)->EvaluateJoin(&left_tuple_, left_schema, &right_tuple, right_schema);
    auto inner_key = Tuple({value}, index_info_->index_->GetKeySchema());
    index_info_->index_->ScanKey(inner_key, &inner_result_, exec_ctx_->GetTransaction());
  }

  return false;
}

由于 Fall2020 没有要求实现哈希索引，所以聚合执行器 AggregationExecutor 内部维护的是直接放在内存中的哈希表 SimpleAggregationHashTable 以及哈希表迭代器 aht_iterator_。将键值对插入哈希表的时候会立刻更新哈希表中保存的聚合结果，最终的查询结果也从该哈希表获取：

void AggregationExecutor::Init() {
  child_->Init();

  // 构造哈希表
  Tuple tuple;
  RID rid;
  while (child_->Next(&tuple, &rid)) {
    aht_.InsertCombine(MakeKey(&tuple), MakeVal(&tuple));
  }

  aht_iterator_ = aht_.Begin();
}

bool AggregationExecutor::Next(Tuple *tuple, RID *rid) {
  auto having = plan_->GetHaving();

  while (aht_iterator_ != aht_.End()) {
    auto group_bys = aht_iterator_.Key().group_bys_;
    auto aggregates = aht_iterator_.Val().aggregates_;
    ++aht_iterator_;

    if (!having || having->EvaluateAggregate(group_bys, aggregates).GetAs<bool>()) {
      std::vector<Value> values;
      for (auto &col : GetOutputSchema()->GetColumns()) {
        values.push_back(col.GetExpr()->EvaluateAggregate(group_bys, aggregates));
      }

      *tuple = {values, GetOutputSchema()};
      return true;
    }
  }

  return false;
}

测试

在终端输入：

cd build
cmake ..
make 

make executor_test
make grading_executor_test    # 从 grade scope 扒下来的测试代码

./test/executor_test
./test/grading_executor_test

测试结果如下，成功通过了所有测试用例：

后记

通过这次实验，可以加深对目录、查询计划、迭代模型和 tuple 页布局的理解，算是收获满满的一次实验了，以上~~