PostgreSQL逻辑订阅处理流程解析

1. 表同步阶段处理流程概述

订阅端执行CREATE SUBSCRIPTION后，在后台进行表数据同步。
每个表的数据同步状态记录在pg_subscription_rel.srsubstate中,一共有4种状态码。

• 'i':初始化
• 'd':正在copy数据
• 's':已同步
• 'r':准备好 (普通复制)

从执行CREATE SUBSCRIPTION开始订阅端的相关处理流程概述如下:

1. 设置每个表的为srsubstate中'i'(SUBREL_STATE_INIT)
2. logical replication launcher启动一个logial replication apply worker进程
3. logial replication apply worker进程连接到订阅端开始接受订阅消息，此时表尚未完成初始同步(状态为i或d)，跳过所有insert,update和delete消息的处理。
4. logial replication apply worker进程为每个未同步的表启动logial replication sync worker进程（每个订阅最多同时启动max_sync_workers_per_subscription个sync worker)
5. logial replication sync worker进程连接到订阅端并同步初始数据
   • 5.1 创建临时复制槽，并记录快照位置。
   • 5.2 设置表同步状态为'd'(SUBREL_STATE_DATASYNC）
   • 5.3 copy表数据
   • 5.4 设置表同步状态为SUBREL_STATE_SYNCWAIT(内部状态)，并等待apply worker更新状态为SUBREL_STATE_CATCHUP(内部状态)
6. logial replication apply worker进程更新表同步状态为SUBREL_STATE_CATCHUP(内部状态)，记录最新lsn，并等待sync worker更新状态为SUBREL_STATE_SYNCDONE
7. logial replication sync worker进程完成初始数据同步
   • 7.1 检查apply worker当前处理的订阅消息位置是否已经走到了快照位置前面，如果是从订阅端接受消息并处理直到追上apply worker。
   • 7.2 设置表同步状态为's'(SUBREL_STATE_SYNCDONE)
   • 7.3 进程退出
8. logial replication apply worker进程继续接受订阅消息并处理
   • 8.1 接受到insert,update和delete消息，如果是同步点(进入's'或'r'状态时的lsn位置)之后的消息进行应用。
   • 8.2 接受到commit消息
     • 8.2.1 更新复制源状态，确保apply worker crash时可以找到正确的开始位置
     • 8.2.2 提交事务
     • 8.2.3 更新统计信息
     • 8.2.4 将所有处于's'(SUBREL_STATE_SYNCDONE)同步状态的表更新为'r'(SUBREL_STATE_READY)
   • 8.3 暂时没有新的消息处理
     • 8.3.1 向发布端发送订阅位置反馈
     • 8.3.2 如果不在事务块里，同步表状态。将所有处于's'(SUBREL_STATE_SYNCDONE)同步状态的表更新为'r'(SUBREL_STATE_READY)

2. 表同步后的持续逻辑复制

订阅表进入同步状态(状态码是‘s’或'r')后，发布端的变更都会通过消息通知订阅端；
订阅端apply worker按照订阅消息的接受顺序(即发布端事务提交顺序)对每个表apply变更，并反馈apply位置，用于监视复制延迟。

通过调试，确认发布端发生更新时，发送给订阅端的数据包。

2.1 插入订阅表

insert into tbx3 values(100);

发布端修改订阅表时，在事务提交时，发布端依次发送下面的消息到订阅端

• B(BEGIN)
• R(RELATION)
• I(INSERT)
• C(COMMIT)
  更新复制源pg_replication_origin_status中的remote_lsn和local_lsn，该位点对应于每个订阅表最后一次事务提交的位置。
• k(KEEPALIVE)
• k(KEEPALIVE)
  2个keepalive消息，会更新统计表中的位置
  • 发布端pg_stat_replication:write_lsn,flush_lsn,replay_lsn
  • 发布端pg_get_replication_slots():confirmed_flush_lsn
  • 订阅端更新pg_stat_subscription:latest_end_lsn

2.2 插入非订阅表

insert into tbx10 values(100);

发布端产生了和订阅表无关修改，在事务提交时，发布端依次发送下面的消息到订阅端

• B(BEGIN)
• C(COMMIT)
  未产生实际事务，也不更新pg_replication_origin_status
• k(KEEPALIVE)
• k(KEEPALIVE)
  2个'k' keepalive消息，会更新统计表中的位置

3. 异常处理

3.1 sync worker

1. SQL错误(如主键冲突):worker进程异常退出，之后apply worker创建一个新的sync worker重试。错误解除前每5秒重试一次。
2. 表被锁:等待
3. 更新或删除的记录不存在:正常执行，检测不到错误，也么没有日志输出(输出一条DEBUG1级别的日志)。

3.2 apply worker

1. SQL错误(如主键冲突):worker进程异常退出，之后logical replication launcher创建一个新的apply worker重试。错误解除前每5秒重试一次。
2. 表被锁:等待
3. 更新或删除的记录不存在:正常执行，检测不到错误，也么没有日志输出(输出一条DEBUG1级别的日志)。

错误日志示例:

2018-07-28 20:11:56.018 UTC [470] ERROR:  duplicate key value violates unique constraint "tbx3_pkey"
2018-07-28 20:11:56.018 UTC [470] DETAIL:  Key (id)=(2) already exists.
2018-07-28 20:11:56.022 UTC [47] LOG:  worker process: logical replication worker for subscription 74283 (PID 470) exited with exit code 1
2018-07-28 20:12:01.029 UTC [471] LOG:  logical replication apply worker for subscription "sub_shard" has started
2018-07-28 20:12:01.049 UTC [471] ERROR:  duplicate key value violates unique constraint "tbx3_pkey"
2018-07-28 20:12:01.049 UTC [471] DETAIL:  Key (id)=(2) already exists.
2018-07-28 20:12:01.058 UTC [47] LOG:  worker process: logical replication worker for subscription 74283 (PID 471) exited with exit code 1
2018-07-28 20:12:06.070 UTC [472] LOG:  logical replication apply worker for subscription "sub_shard" has started
2018-07-28 20:12:06.089 UTC [472] ERROR:  duplicate key value violates unique constraint "tbx3_pkey"
2018-07-28 20:12:06.089 UTC [472] DETAIL:  Key (id)=(2) already exists.

4. 限制

1. 不复制数据库模式和DDL命令。
2. 不复制序列数据。序列字段(serial / GENERATED ... AS IDENTITY)的值会被复制，但序列的值不会更新
3. 不复制TRUNCATE命令。
4. 不复制大对象
5. 复制只能从基表到基表。也就是说，发布和订阅端的表必须是普通表，而不是视图， 物化视图，分区根表或外部表。订阅继承表的父表，只会复制父表的变更。
6. 只支持触发器的一部分功能
7. 不支持双向复制，会导致WAL循环。
8. 不支持在同一个实例上的两个数据库上创建订阅
9. 不支持在备机上创建订阅
10. 订阅表上没有合适的REPLICA IDENTITY时，发布端执行UPDATE/DELETE会报错

注意事项

1. CREATE SUBSCRIPTION命令执行时，要等待发布端正在执行的事务结束。
2. sync worker初始同步数据时，开启了"REPEATABLE READ"事务，期间产生的垃圾不能被回收。
3. 订阅生效期间，发布端所有事务产生的WAL必须在该事务结束时才能被回收。
4. 订阅端UPDATE/DELETE找不到数据时，没有任何错误输出。

5. 表同步阶段相关代码解析

发布端Backend进程

CREATE PUBLICATION
  CreatePublication()
       CatalogTupleInsert(rel, tup);  // 在pg_publication系统表中插入此发布信息
       PublicationAddTables(puboid, rels, true, NULL);//
         publication_add_relation()
           check_publication_add_relation();// 检查表类型,不支持的表报错。只支持普通表('r')，且不是unloged和临时表
         CatalogTupleInsert(rel, tup);      // 在pg_publication_rel系统表中插入订阅和表的映射

订阅端Backend进程

CREATE SUBSCRIPTION
   CreateSubscription()
       CatalogTupleInsert(rel, tup);  //在pg_subscription系统表中插入此订阅信息
       replorigin_create(originname); //在pg_replication_origin系统表中插入此订阅对应的复制源
       foreach(lc, tables)            // 设置每个表的pg_subscription_rel.srsubstate
         table_state = copy_data ? SUBREL_STATE_INIT : SUBREL_STATE_READY; // ★★★1 如果拷贝数据，设置每个表的pg_subscription_rel.srsubstate='i'
         SetSubscriptionRelState(subid, relid, table_state,InvalidXLogRecPtr, false); 
       walrcv_create_slot(wrconn, slotname, false,CRS_NOEXPORT_SNAPSHOT, &lsn);
       ApplyLauncherWakeupAtCommit(); //唤醒logical replication launcher进程

订阅端logical replication launcher进程

ApplyLauncherMain()
  sublist = get_subscription_list(); //从pg_subscription获取订阅列表
  foreach(lc, sublist)
    logicalrep_worker_launch(..., InvalidOid); // 对enabled且没有创建worker的订阅创建apply worker。apply worker如果已超过max_logical_replication_workers(默认4)报错
      RegisterDynamicBackgroundWorker(&bgw, &bgw_handle);// 注册后台工作进程，入口函数为"ApplyWorkerMain"

订阅端 logical apply worker进程

ApplyWorkerMain
  replorigin_session_setup(originid); // 从共享内存中查找并设置复制源，如果不存在使用新的，复制源名称为pg_${订阅OID}。
  origin_startpos = replorigin_session_get_progress(false);// 获取复制源的remote_lsn
  walrcv_connect(MySubscription->conninfo, true, MySubscription->name,&err); // 连接到订阅端
  walrcv_startstreaming(wrconn, &options); // 开始流复制
  LogicalRepApplyLoop(origin_startpos);   // Apply进程主循环
    for(;;)
      len = walrcv_receive(wrconn, &buf, &fd);
      if (c == 'w')  // 'w'消息的处理
          UpdateWorkerStats(last_received, send_time, false);更新worker统计信息(last_lsn,last_send_time,last_recv_time)
          apply_dispatch(&s); // 分发逻辑复制命令
            switch (action)
              case 'B': /* BEGIN */
                 apply_handle_begin(s);
              case 'C': /* COMMIT */
                 apply_handle_commit(s);
                   if (IsTransactionState() && !am_tablesync_worker()) // 当发布端的事务更新不涉及订阅表时，仍会发送B和C消息，此时不在事务中，跳过下面操作
                     replorigin_session_origin_lsn = commit_data.end_lsn;  // 更新复制源状态，确保apply worker crash时可以找到正确的开始位置
                     replorigin_session_origin_timestamp = commit_data.committime;
                     CommitTransactionCommand(); // 提交事务
                     pgstat_report_stat(false); // 更新统计信息
                   process_syncing_tables(commit_data.end_lsn); // 对处于同步中的表，协调sync worker和apply worker进程同步状态
                     process_syncing_tables_for_apply(current_lsn);
                       GetSubscriptionNotReadyRelations(MySubscription->oid); // 从pg_subscription_rel中获取订阅中所有非ready状态的表。
                       foreach(lc, table_states) // 处理每个非ready状态的表
                        if (rstate->state == SUBREL_STATE_SYNCDONE)
                        {
                            if (current_lsn >= rstate->lsn)
                            {
                                rstate->state = SUBREL_STATE_READY;                  //处理第一个事务后，从syncdone->ready状态，但这个事务不需要和这个表相关。
                                rstate->lsn = current_lsn;
                                SetSubscriptionRelState(MyLogicalRepWorker->subid,    // 更新pg_subscription_rel
                                                        rstate->relid, rstate->state,
                                                        rstate->lsn, true);
                            }
                          }
                        else
                        {
                            syncworker = logicalrep_worker_find(MyLogicalRepWorker->subid,
                                                                rstate->relid, false);
                            if (syncworker)
                            {
                                /* Found one, update our copy of its state */
                                rstate->state = syncworker->relstate;
                                rstate->lsn = syncworker->relstate_lsn;
                                if (rstate->state == SUBREL_STATE_SYNCWAIT)
                                {
                                    /*
                                     * Sync worker is waiting for apply.  Tell sync worker it
                                     * can catchup now.
                                     */
                                    syncworker->relstate = SUBREL_STATE_CATCHUP;  // ★★★3 SUBREL_STATE_SYNCWAIT -> SUBREL_STATE_CATCHUP
                                    syncworker->relstate_lsn =
                                        Max(syncworker->relstate_lsn, current_lsn);
                                }

                                /* If we told worker to catch up, wait for it. */
                                if (rstate->state == SUBREL_STATE_SYNCWAIT)
                                {
                                    /* Signal the sync worker, as it may be waiting for us. */
                                    if (syncworker->proc)
                                        logicalrep_worker_wakeup_ptr(syncworker);

                                    wait_for_relation_state_change(rstate->relid,
                                                                   SUBREL_STATE_SYNCDONE); // 等待sync worker将表的同步状态设置为SUBREL_STATE_SYNCDONE
                                }
                            }
                            else
                            {
                                /*
                                 * If there is no sync worker for this table yet, count
                                 * running sync workers for this subscription, while we have
                                 * the lock.
                                 */
                                        logicalrep_worker_launch(MyLogicalRepWorker->dbid,   // 如果这个表没有对应的sync worker，且sync worker数未超过max_sync_workers_per_subscription，启动一个。
                                                                 MySubscription->oid,
                                                                 MySubscription->name,
                                                                 MyLogicalRepWorker->userid,
                                                                 rstate->relid);
                            }
            else if (c == 'k') // 'k'消息的处理
              send_feedback(last_received, reply_requested, false); // 向订阅端发生反馈
              UpdateWorkerStats(last_received, timestamp, true);    // 更新worker统计信息(last_lsn,last_send_time,last_recv_time,reply_lsn,send_time) 
          case I': /* INSERT */
             apply_handle_insert(s);
               relid = logicalrep_read_insert(s, &newtup);
               if (!should_apply_changes_for_rel(rel))return; 
                    if (am_tablesync_worker())
                        return MyLogicalRepWorker->relid == rel->localreloid; // 对sync worker，只apply其负责同步的表
                    else
                        return (rel->state == SUBREL_STATE_READY ||           // 对apply worker, 同步状态为SUBREL_STATE_SYNCDONE时，只同步syncdone位置之后的wal
                                (rel->state == SUBREL_STATE_SYNCDONE &&
                                 rel->statelsn <= remote_final_lsn));
               ExecSimpleRelationInsert(estate, remoteslot);  // 插入记录
                 ExecBRInsertTriggers(estate, resultRelInfo, slot); // 处理BEFORE ROW INSERT Triggers
                 simple_heap_insert(rel, tuple);
                 ExecARInsertTriggers(estate, resultRelInfo, tuple,recheckIndexes, NULL); // 处理AFTER ROW INSERT Triggers
               AfterTriggerEndQuery(estate);  // 处理 queued AFTER triggers

          ...
    send_feedback(last_received, false, false);//没有新的消息要处理，向发布端发送位置反馈
    process_syncing_tables(last_received);//如果不在事务块里，同步表状态

订阅端 logical sync worker进程

ApplyWorkerMain() //apply worker和sync worker使用相同的入口函数
  LogicalRepSyncTableStart(&origin_startpos);
    GetSubscriptionRelState()(MyLogicalRepWorker->subid,MyLogicalRepWorker->relid,&relstate_lsn, true);// 从pg_subscription_rel中获取订阅的复制lsn
    walrcv_connect(MySubscription->conninfo, true, slotname, &err);
    switch (MyLogicalRepWorker->relstate)
    {
        case SUBREL_STATE_INIT:
        case SUBREL_STATE_DATASYNC:
            {
                MyLogicalRepWorker->relstate = SUBREL_STATE_DATASYNC;
                MyLogicalRepWorker->relstate_lsn = InvalidXLogRecPtr;
                SetSubscriptionRelState(MyLogicalRepWorker->subid,
                                        MyLogicalRepWorker->relid,
                                        MyLogicalRepWorker->relstate,
                                        MyLogicalRepWorker->relstate_lsn,
                                        true);
                res = walrcv_exec(wrconn,                              // 开始事务
                                  "BEGIN READ ONLY ISOLATION LEVEL "
                                  "REPEATABLE READ", 0, NULL);
                walrcv_create_slot(wrconn, slotname, true,             // 使用快照创建临时复制槽，并记录快照位置。
                                   CRS_USE_SNAPSHOT, origin_startpos);
                copy_table(rel);                                       // copy表数据
                walrcv_exec(wrconn, "COMMIT", 0, NULL);
                MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCWAIT;  // ★★★2 更新表同步状态为SUBREL_STATE_SYNCWAIT
                MyLogicalRepWorker->relstate_lsn = *origin_startpos;
                wait_for_worker_state_change(SUBREL_STATE_CATCHUP);    // 等待apply worker将状态变更为SUBREL_STATE_CATCHUP
                if (*origin_startpos >= MyLogicalRepWorker->relstate_lsn) // 如果sync worker落后于apply worker，sync worker跳过此步继续apply WAL；
                {
                    /*
                     * Update the new state in catalog.  No need to bother
                     * with the shmem state as we are exiting for good.
                     */
                    SetSubscriptionRelState(MyLogicalRepWorker->subid,    // ★★★4 把同步状态从SUBREL_STATE_CATCHUP更新到SUBREL_STATE_SYNCDONE并退出
                                            MyLogicalRepWorker->relid,
                                            SUBREL_STATE_SYNCDONE,
                                            *origin_startpos,
                                            true);
                    finish_sync_worker();
                }
                break;
            }
        case SUBREL_STATE_SYNCDONE:
        case SUBREL_STATE_READY:
        case SUBREL_STATE_UNKNOWN:
            finish_sync_worker();
            break;    
    }
  options.startpoint = origin_startpos;
  walrcv_startstreaming(wrconn, &options);// 开始流复制，以同步快照位置作为流的开始位置
  LogicalRepApplyLoop(origin_startpos);   // Apply进程主循环
    for(;;)
      len = walrcv_receive(wrconn, &buf, &fd);
      UpdateWorkerStats(last_received, send_time, false); 更新worker统计信息(last_lsn,last_send_time,last_recv_time)
      apply_dispatch(&s); // 分发逻辑复制命令
        switch (action)
          case 'B': /* BEGIN */
             apply_handle_begin(s);
          case 'C': /* COMMIT */
             apply_handle_commit(s);
               process_syncing_tables(commit_data.end_lsn); // 对处于同步中的表，协调sync worker和apply worker进程同步状态
                 process_syncing_tables_for_sync(current_lsn);
                    if (MyLogicalRepWorker->relstate == SUBREL_STATE_CATCHUP &&
                        current_lsn >= MyLogicalRepWorker->relstate_lsn)
                    {
                        TimeLineID    tli;

                        MyLogicalRepWorker->relstate = SUBREL_STATE_SYNCDONE; // ★★★4 把同步状态从SUBREL_STATE_CATCHUP更新到SUBREL_STATE_SYNCDONE
                        MyLogicalRepWorker->relstate_lsn = current_lsn;

                        SpinLockRelease(&MyLogicalRepWorker->relmutex);

                        SetSubscriptionRelState(MyLogicalRepWorker->subid,
                                                MyLogicalRepWorker->relid,
                                                MyLogicalRepWorker->relstate,
                                                MyLogicalRepWorker->relstate_lsn,
                                                true);

                        walrcv_endstreaming(wrconn, &tli);
                        finish_sync_worker();
                    }
          case I': /* INSERT */
             apply_handle_insert(s);

6.1 参考

• https://yq.aliyun.com/articles/71128

• http://www.postgres.cn/docs/10/logical-replication.html