greenplum源码解析全局死锁检测GDD-1

背景

对于单机数据库而言[如postgres]，死锁检测机制相对于分布式数据库如greenplum来说相对简单，其本质是通过相关约束构造等待图，如果出现死锁情况，则采用拓扑排序重排序列规避死锁，若无解法则回滚等待环中的某一事务/或若干事务来打破死锁。但是对于share nothing 的MPP架构数据库而言，采用上述方法无法获取全局等待视图，死锁的攻克显得力不从心.另辟蹊径，greenplum设计的方法是：通过在 master 节点设置全局死锁检测进程，不断地轮询计算节点上的本地等待图graph,在master上汇总形成全局等待图，利用相应的规则和算法来打破死锁。接下来一起学习，探索源码：

关键数据结构

1 数据结构流转图

2 概念介绍

3 数据结构分析
3.1 GddCtx：全局死锁上下文结构体，该结构体记录了全局等待图的相关信息，包括所有节点[分布式事务]的出/入度、segid与节点间映射关系、segid与本地等待图间的映射关系信息

struct GddCtx
{GddStat        topstat;        /* overall in/out degrees of all verts on all graphs */GddMap       globals;        /* Map<vid, Stat>, global in/out degrees */GddMap     graphs;         /* Map<segid, Graph>, the local graphs */
};

3.2 GddStat：记录某个分布式事务节点的出入度信息

/** Accounting information of a vert.*/
struct GddStat
{int            id;             /* vert id */        // 节点号  ===》 分布式事务号int          indeg;          /* in degree */      // 该节点的入度int           outdeg;         /* out degree */     // 该节点的出度
};

3.3 GddEdge:该结构记录等待图中边信息：包括是否为硬边、出节点、入节点信息和存放的数据

/** A directed edge.*/
struct GddEdge
{bool       solid;          /* a solid edge? */GddVert      *from;          /* the from vert */GddVert      *to;            /* the to vert */void        *data;
};

3.4 GddVert：该结构为计算segment上某个节点的相关信息，包括该节点在全局视图中的出入度，出边和入边等信息

/** A vertex (vert) on a segment.*/
struct GddVert
{int            id;             /* vert id *//** Pointer to ctx->topstat, all the verts on all the globals maintain* the overall in/out degrees count together.*/GddStat     *topstat;/** Pointer to the `global` struct with the same vert id, the global* in/out degrees of vert.*/GddStat       *global;List        *edgesIn;       /* List<Edge>, directed edges to vert */List      *edgesOut;      /* List<Edge>, directed edges from vert *//** The data set and used only by the caller, GDD does not touch or access* it.*/void       *data;
};

3.5 GddMap：本质GddPair数组，保存的是K-V映射关系

/** A <int, pointer> pair.*/
struct GddPair
{int            key;void        *ptr;
};/** A simple int->ptr map, implemented with array.*/
struct GddMap
{int            length;         /* length of the <k,v> array */int            capacity;       /* capacity of the <k,v> array */GddPair      *pairs;         /* array of <k,v> pairs */
};

3 GDD进程介绍和工作总流程

该进程是greenplum内置的辅助进程，postmaster正式接受客户端请求前在后台启动

 {"global deadlock detector process", "global deadlock detector process",BGWORKER_SHMEM_ACCESS | BGWORKER_BACKEND_DATABASE_CONNECTION,BgWorkerStart_RecoveryFinished,0, /* restart immediately if gdd exits with non-zero code */"postgres", "GlobalDeadLockDetectorMain", 0, {0}, 0,GlobalDeadLockDetectorStartRule},        运行规则：bool GlobalDeadLockDetectorStartRule(Datum main_arg)
{if (Gp_role == GP_ROLE_DISPATCH &&gp_enable_global_deadlock_detector)return true;return false;
}

入口函数：GlobalDeadLockDetectorMain

/** GlobalDeadLockDetectorMain*/
void
GlobalDeadLockDetectorMain(Datum main_arg)
{am_global_deadlock_detector = true;                           // 更新标识pqsignal(SIGHUP, sigHupHandler);                               // 注册信号处理函数/* We're now ready to receive signals */BackgroundWorkerUnblockSignals();                         // 打开信号/* Connect to our database */BackgroundWorkerInitializeConnection(DB_FOR_COMMON_ACCESS, NULL, 0);        // 连接数据库/* disable orca here */extern bool optimizer;optimizer = false;GlobalDeadLockDetectorLoop();                            // 死锁检测，轮询，下节重点讲解/* One iteration done, go away */proc_exit(0);
}