概述

在RocksDB 3.0中加入了Column Family特性，加入这个特性之后，每一个KV对都会关联一个Column Family,其中默认的Column Family是 “default”. Column Family主要是提供给RocksDB一个逻辑的分区.从实现上来看不同的Column Family共享WAL，而都有自己的Memtable和SST.这就意味着我们可以很 快速已经方便的设置不同的属性给不同的Column Family以及快速删除对应的Column Family.

主要API

首先是创建Column Family，这里注意我们可以通过两种方式来创建Column Family，一种是在Open DB的时候通过传递需要创建的Column Family，一种是当DB创建并打开之后， 通过直接的CreateColumnFamily来创建Column Family.

  1. DB::Open(const DBOptions& db_options, const std::string& name, const std::vector<ColumnFamilyDescriptor>& column_families, std::vector<ColumnFamilyHandle*>* handles, DB** dbptr);
  2. DB::CreateColumnFamily(const ColumnFamilyOptions& options, const std::string& column_family_name, ColumnFamilyHandle** handle);

这里可以看到不管是哪一种方式最终都会返回一个ColumnFamilyHandle给调用者来使用.

然后就是删除Column Family的方式，这里很简单就是传递之前创建的ColumnFamilyHandle给RocksDB，然后用以删除.

  1. DropColumnFamily(ColumnFamilyHandle* column_family);

实现

所有的Column Family都是通过一个叫做ColumnFamilySet的结构来管理的，而每一个Column Family都是一个ColumnFamilyData.

先来看ColumnFamilySet,这里可以看到它有两个数据结构来管理Column Family,分别是map(column_family_data_)以及一个双向链表(dummy_cfd_). 其中map用来保存Column Family名字和对应的id以及ColumnFamilyData的映射. 这里要注意在RocksDB内部是将没一个ColumnFamily的名字表示为一个uint32类型的ID(max_column_family_).也就是这个ID是一个简单的递增的数值.

  1. class ColumnFamilySet {
  2.  public:
  3.   // ColumnFamilySet supports iteration
  4.    public:
  5. .................................
  7.   ColumnFamilyData* CreateColumnFamily(const std::string& name, uint32_t id,
  8.                                        Version* dummy_version,
  9.                                        const ColumnFamilyOptions& options);
  10.   iterator begin() { return iterator(dummy_cfd_->next_); }
  11.   iterator end() { return iterator(dummy_cfd_); }
  12. ...............................
  13.  private:
  14.   friend class ColumnFamilyData;
  15.   // helper function that gets called from cfd destructor
  16.   // REQUIRES: DB mutex held
  17.   void RemoveColumnFamily(ColumnFamilyData* cfd);
  19.   // column_families_ and column_family_data_ need to be protected:
  20.   // * when mutating both conditions have to be satisfied:
  21.   // 1. DB mutex locked
  22.   // 2. thread currently in single-threaded write thread
  23.   // * when reading, at least one condition needs to be satisfied:
  24.   // 1. DB mutex locked
  25.   // 2. accessed from a single-threaded write thread
  26.   std::unordered_map<std::string, uint32_t> column_families_;
  27.   std::unordered_map<uint32_t, ColumnFamilyData*> column_family_data_;
  29.   uint32_t max_column_family_;
  30.   ColumnFamilyData* dummy_cfd_;
  31.   // We don't hold the refcount here, since default column family always exists
  32.   // We are also not responsible for cleaning up default_cfd_cache_. This is
  33.   // just a cache that makes common case (accessing default column family)
  34.   // faster
  35.   ColumnFamilyData* default_cfd_cache_;
  37. ..................................
  38. };

然后来看ColumnFamilyData，这个数据结构就是用来表示一个ColumnFamily,保存了对应的信息,我们可以看到有ID/name以及当前ColumnFamily对应的所有的version(dummy_versions_). 其中这里的next_/prev_就是在ColumnFamilySet中用来表示所有ColumnFamily的双向链表.

  1. class ColumnFamilyData {
  2.  public:
  3.   ~ColumnFamilyData();
  5.   // thread-safe
  6.   uint32_t GetID() const { return id_; }
  7.   // thread-safe
  8.   const std::string& GetName() const { return name_; }
  10.   // Ref() can only be called from a context where the caller can guarantee
  11.   // that ColumnFamilyData is alive (while holding a non-zero ref already,
  12.   // holding a DB mutex, or as the leader in a write batch group).
  13.   void Ref() { refs_.fetch_add(1, std::memory_order_relaxed); }
  15.   // Unref decreases the reference count, but does not handle deletion
  16.   // when the count goes to 0.  If this method returns true then the
  17.   // caller should delete the instance immediately, or later, by calling
  18.   // FreeDeadColumnFamilies().  Unref() can only be called while holding
  19.   // a DB mutex, or during single-threaded recovery.
  20.   bool Unref() {
  21.     int old_refs = refs_.fetch_sub(1, std::memory_order_relaxed);
  22.     assert(old_refs > 0);
  23.     return old_refs == 1;
  24.   }
  25. ..............................
  27.  private:
  28.   friend class ColumnFamilySet;
  29.   ColumnFamilyData(uint32_t id, const std::string& name,
  30.                    Version* dummy_versions, Cache* table_cache,
  31.                    WriteBufferManager* write_buffer_manager,
  32.                    const ColumnFamilyOptions& options,
  33.                    const ImmutableDBOptions& db_options,
  34.                    const EnvOptions& env_options,
  35.                    ColumnFamilySet* column_family_set);
  37.   uint32_t id_;
  38.   const std::string name_;
  39.   Version* dummy_versions_;  // Head of circular doubly-linked list of versions.
  40.   Version* current_;         // == dummy_versions->prev_
  41. ......................................................
  43.   // Thread's local copy of SuperVersion pointer
  44.   // This needs to be destructed before mutex_
  45.   std::unique_ptr<ThreadLocalPtr> local_sv_;
  47.   // pointers for a circular linked list. we use it to support iterations over
  48.   // all column families that are alive (note: dropped column families can also
  49.   // be alive as long as client holds a reference)
  50.   ColumnFamilyData* next_;
  51.   ColumnFamilyData* prev_;
  52. ...................................
  54.   ColumnFamilySet* column_family_set_;
  55. ..................................
  56. };

然后就是返回给调用者的ColumnFamilyHandleImpl结构，这个结构主要是封装了ColumnFamilyData.

  1. // ColumnFamilyHandleImpl is the class that clients use to access different
  2. // column families. It has non-trivial destructor, which gets called when client
  3. // is done using the column family
  4. class ColumnFamilyHandleImpl : public ColumnFamilyHandle {
  5.  public:
  6.   // create while holding the mutex
  7.   ColumnFamilyHandleImpl(
  8.       ColumnFamilyData* cfd, DBImpl* db, InstrumentedMutex* mutex);
  9.   // destroy without mutex
  10.   virtual ~ColumnFamilyHandleImpl();
  11.   virtual ColumnFamilyData* cfd() const { return cfd_; }
  12. ......................................
  14.  private:
  15.   ColumnFamilyData* cfd_;
  16.   DBImpl* db_;
  17.   InstrumentedMutex* mutex_;
  18. };

接下来我们就来从ColumnFamily的创建以及删除来分析ColumnFamily的实现.我们从DBImpl::CreateColumnFamilyImpl开始.在这个函数 中首先就是通过调用GetNextColumnFamilyID来得到当前创建的ColumnFamily对应的ID(自增).然后再调用LogAndApply来对ColumnFamily 进行对应的操作.最后再返回封装好的ColumnFamilyHandle给调用者.

  1. Status DBImpl::CreateColumnFamilyImpl(const ColumnFamilyOptions& cf_options,
  2.                                       const std::string& column_family_name,
  3.                                       ColumnFamilyHandle** handle) {
  4. .......................................
  6.   {
  7. ...................................
  8.     VersionEdit edit;
  9.     edit.AddColumnFamily(column_family_name);
  10.     uint32_t new_id = versions_->GetColumnFamilySet()->GetNextColumnFamilyID();
  11.     edit.SetColumnFamily(new_id);
  12.     edit.SetLogNumber(logfile_number_);
  13.     edit.SetComparatorName(cf_options.comparator->Name());
  15.     // LogAndApply will both write the creation in MANIFEST and create
  16.     // ColumnFamilyData object
  17.     {  // write thread
  18.       WriteThread::Writer w;
  19.       write_thread_.EnterUnbatched(&w, &mutex_);
  20.       // LogAndApply will both write the creation in MANIFEST and create
  21.       // ColumnFamilyData object
  22.       s = versions_->LogAndApply(nullptr, MutableCFOptions(cf_options), &edit,
  23.                                  &mutex_, directories_.GetDbDir(), false,
  24.                                  &cf_options);
  25.       write_thread_.ExitUnbatched(&w);
  26.     }
  27.     if (s.ok()) {
  28. ........................................
  29.       *handle = new ColumnFamilyHandleImpl(cfd, this, &mutex_);
  30.       ROCKS_LOG_INFO(immutable_db_options_.info_log,
  31.                      "Created column family [%s] (ID %u)",
  32.                      column_family_name.c_str(), (unsigned)cfd->GetID());
  33.     }
  34. .............................................
  35.   }  // InstrumentedMutexLock l(&mutex_)
  37. .................................
  38.   return s;
  39. }

最终会在LogAndApply调用ColumnFamilySet的CreateColumnFamily函数(通过VersionSet::CreateColumnFamily),这个函数我们可看到主要做了下面三件事情

  1. 创建ColumnFamilyData对象
  2. 将新的创建好的CFD加入到双向链表

  3. 对应的Map数据结构更新数据

    1. // under a DB mutex AND write thread
    2. ColumnFamilyData* ColumnFamilySet::CreateColumnFamily(
    3.  const std::string& name, uint32_t id, Version* dummy_versions,
    4.  const ColumnFamilyOptions& options) {
    5.   assert(column_families_.find(name) == column_families_.end());
    6.   ColumnFamilyData* new_cfd = new ColumnFamilyData(
    7.    id, name, dummy_versions, table_cache_, write_buffer_manager_, options,
    8.    *db_options_, env_options_, this);
    9.   column_families_.insert({name, id});
    10.   column_family_data_.insert({id, new_cfd});
    11.   max_column_family_ = std::max(max_column_family_, id);
    12.   // add to linked list
    13.   new_cfd->next_ = dummy_cfd_;
    14.   auto prev = dummy_cfd_->prev_;
    15.   new_cfd->prev_ = prev;
    16.   prev->next_ = new_cfd;
    17.   dummy_cfd_->prev_ = new_cfd;
    18.   if (id == 0) {
    19.  default_cfd_cache_ = new_cfd;
    20.   }
    21.   return new_cfd;
    22. }

    然后来看如何删除ColumnFamily,这里所有的删除最终都会调用ColumnFamilySet::RemoveColumnFamily函数，这个函数是是从两个Map中删除对应的ColumnFamily. 这里或许我们要问了，为什么管理的双向链表不需要删除呢。这里原因是这样的，由于ColumnFamilyData是通过引用计数管理的，因此只有当所有的引用计数都清零之后， 才需要真正的函数ColumnFamilyData(也就是会从双向链表中删除数据).

    1. // under a DB mutex AND from a write thread
    2. void ColumnFamilySet::RemoveColumnFamily(ColumnFamilyData* cfd) {
    3.   auto cfd_iter = column_family_data_.find(cfd->GetID());
    4.   assert(cfd_iter != column_family_data_.end());
    5.   column_family_data_.erase(cfd_iter);
    6.   column_families_.erase(cfd->GetName());
    7. }

因此我们来看ColumnFamilyData的析构函数.可以看到析构函数中会从双向链表中删除对应的数据,以及处理对应的Version(corrent_).

  1. // DB mutex held
  2. ColumnFamilyData::~ColumnFamilyData() {
  3.   assert(refs_.load(std::memory_order_relaxed) == 0);
  4.   // remove from linked list
  5.   auto prev = prev_;
  6.   auto next = next_;
  7.   prev->next_ = next;
  8.   next->prev_ = prev;
  10.   if (!dropped_ && column_family_set_ != nullptr) {
  11.     // If it's dropped, it's already removed from column family set
  12.     // If column_family_set_ == nullptr, this is dummy CFD and not in
  13.     // ColumnFamilySet
  14.     column_family_set_->RemoveColumnFamily(this);
  15.   }
  17.   if (current_ != nullptr) {
  18.     current_->Unref();
  19.   }
  20. ..............................
  21. }

最后我们来看一下在磁盘上ColumnFamily是如何保存的，首先需要明确的是ColumnFamily是保存在MANIFEST文件中的，信息的保存比较简单(之前的文章有介绍), 和MANIFEST中其他的信息没什么区别，因此这里我们主要来看数据的读取以及初始化，这里所有的操作都是包含在VersionSet::Recover中，我们来看这个函数.

函数主要的逻辑就是读取MANIFEST然后来再来将磁盘上读取的ColumnFamily的信息初始化(初始化ColumnFamilySet结构),可以看到这里相当于将之前的create/drop 的操作全部回放一遍，也就是会调用CreateColumnFamily/DropColumnFamily来将磁盘的信息初始化到内存.

  1. while (reader.ReadRecord(&record, &scratch) && s.ok()) {
  2.       VersionEdit edit;
  3.       s = edit.DecodeFrom(record);
  4.       if (!s.ok()) {
  5.         break;
  6.       }
  8.       // Not found means that user didn't supply that column
  9.       // family option AND we encountered column family add
  10.       // record. Once we encounter column family drop record,
  11.       // we will delete the column family from
  12.       // column_families_not_found.
  13.       bool cf_in_not_found =
  14.           column_families_not_found.find(edit.column_family_) !=
  15.           column_families_not_found.end();
  16.       // in builders means that user supplied that column family
  17.       // option AND that we encountered column family add record
  18.       bool cf_in_builders =
  19.           builders.find(edit.column_family_) != builders.end();
  21.       // they can't both be true
  22.       assert(!(cf_in_not_found && cf_in_builders));
  24.       ColumnFamilyData* cfd = nullptr;
  26.       if (edit.is_column_family_add_) {
  27.         if (cf_in_builders || cf_in_not_found) {
  28.           s = Status::Corruption(
  29.               "Manifest adding the same column family twice");
  30.           break;
  31.         }
  32.         auto cf_options = cf_name_to_options.find(edit.column_family_name_);
  33.         if (cf_options == cf_name_to_options.end()) {
  34.           column_families_not_found.insert(
  35.               {edit.column_family_, edit.column_family_name_});
  36.         } else {
  37.           cfd = CreateColumnFamily(cf_options->second, &edit);
  38.           cfd->set_initialized();
  39.           builders.insert(
  40.               {edit.column_family_, new BaseReferencedVersionBuilder(cfd)});
  41.         }
  42.       } else if (edit.is_column_family_drop_) {
  43.         if (cf_in_builders) {
  44.           auto builder = builders.find(edit.column_family_);
  45.           assert(builder != builders.end());
  46.           delete builder->second;
  47.           builders.erase(builder);
  48.           cfd = column_family_set_->GetColumnFamily(edit.column_family_);
  49.           if (cfd->Unref()) {
  50.             delete cfd;
  51.             cfd = nullptr;
  52.           } else {
  53.             // who else can have reference to cfd!?
  54.             assert(false);
  55.           }
  56.         } else if (cf_in_not_found) {
  57.           column_families_not_found.erase(edit.column_family_);
  58.         } else {
  59.           s = Status::Corruption(
  60.               "Manifest - dropping non-existing column family");
  61.           break;
  62.         }
  63.       } else if (!cf_in_not_found) {
  64.         if (!cf_in_builders) {
  65.           s = Status::Corruption(
  66.               "Manifest record referencing unknown column family");
  67.           break;
  68.         }
  70.         cfd = column_family_set_->GetColumnFamily(edit.column_family_);
  71.         // this should never happen since cf_in_builders is true
  72.         assert(cfd != nullptr);
  74.         // if it is not column family add or column family drop,
  75.         // then it's a file add/delete, which should be forwarded
  76.         // to builder
  77.         auto builder = builders.find(edit.column_family_);
  78.         assert(builder != builders.end());
  79.         builder->second->version_builder()->Apply(&edit);
  80.       }
  82.       if (cfd != nullptr) {
  83.         if (edit.has_log_number_) {
  84.           if (cfd->GetLogNumber() > edit.log_number_) {
  85.             ROCKS_LOG_WARN(
  86.                 db_options_->info_log,
  87.                 "MANIFEST corruption detected, but ignored - Log numbers in "
  88.                 "records NOT monotonically increasing");
  89.           } else {
  90.             cfd->SetLogNumber(edit.log_number_);
  91.             have_log_number = true;
  92.           }
  93.         }
  94.         if (edit.has_comparator_ &&
  95.             edit.comparator_ != cfd->user_comparator()->Name()) {
  96.           s = Status::InvalidArgument(
  97.               cfd->user_comparator()->Name(),
  98.               "does not match existing comparator " + edit.comparator_);
  99.           break;
  100.         }
  101.       }
  103.       if (edit.has_prev_log_number_) {
  104.         previous_log_number = edit.prev_log_number_;
  105.         have_prev_log_number = true;
  106.       }
  108.       if (edit.has_next_file_number_) {
  109.         next_file = edit.next_file_number_;
  110.         have_next_file = true;
  111.       }
  113.       if (edit.has_max_column_family_) {
  114.         max_column_family = edit.max_column_family_;
  115.       }
  117.       if (edit.has_last_sequence_) {
  118.         last_sequence = edit.last_sequence_;
  119.         have_last_sequence = true;
  120.       }
  121.     }

原文：http://mysql.taobao.org/monthly/2018/06/09/