prepared事务的恢复过程： 当执行prepare时候，PG会把该事务的lock信息当做prepare日志记录的一部分记录在日志文件（xlog）里。当数 据库重新启动，会读这个日志文件（xlog）这条日志记录，把锁“还原”到pg_lock表里。 1. StartupXlog函数发现XLOG_XACT_PREPARE日志记录进行redo，调用函数recreateTwoPhaseFile将 该日志记录中的信息放到 用于判断分布式事务是否提交，作用和PG 的commit log类似，基于simple LRU实现 • 分布式死锁检测 • 本地事务的管理：创建、提交、状态迁移 等 • 加锁和MVCC • 本地事务的死锁检测 • xlog、commit log(CLOG) • 对PREPARE、COMMIT/ABORT PREPARED语句的处理 Greenplum在PG的基础上实现 Greenplum复用PG的实现 33

进行诊断和维护 在某些极端情況下，数据库重启失敗，可能原因是Persistent或Xlog有数据不一致，此时，请在Support说明下进行修复，例如抑制错误、修复 xlog日志等，需要说明的是，未得到support同意，不要用pg_resetxlog去修復xlog，否则可能导致数据不一致；  故障机器隔离 当发现有机器本身有故障，例如经常性自动重启

naming changes in Greenplum 7: Old Name New Name master coordinator pg_log log pg_xlog pg_wal pg_clog pg_xact xlog (includes SQL functions, tools and options) wal pg_xlogdump pg_waldump pg_resetxlog less than 32kB as 32kB. If you specify this value without units, it is taken as WAL blocks, that is XLOG_BLCKSZ bytes, typically 32kB. You can set this parameter only at server start. The contents of the give reasonable results in most cases. Value Range Default Set Classifications -1 - (INT_MAX / XLOG_BLCKSZ) (integer) -1 local, system, restart wal_compression Enables compression of full page writes

姓名：张文杰 邮箱：zhuodao.zwj@alibaba-inc.com 公司：阿里云 2017 年象行中国（杭州 站）第一期 Greenplum数据备份恢复： 1. 数据量较大 2. 不能完全使用Xlog日志备份 3. 需要保证数据完整性和一致性 2017 年象行中国（杭州 站）第一期 Greenplum提供了： 1. 非并行备份和恢复： --pg_dump和pg_dumpall（pg_restore）

postgres 4096 Dec 8 11:22 base 11:22 PG_VERSION drwx------ 3 postgres postgres 4096 Dec 8 11:22 pg_xlog -rw------- 1 postgres postgres 88 Dec 8 11:22 postgresql.auto.conf -rw------- 1 postgres postgres

incorrectly reset to 0. This could cause an infinite loop of retries and failures and lead to the pg_xlog file growing in size with the checkpointer process hanging. The problem was resolved by ensuring "22023","recovery command file ""recovery.conf"" request for standby mode not specified",,,,,,,0,,"xlog.c",5465 Greenplum Database now supports server recovery in a non-continuous mode using standby_mode=on race condition between a checkpoint and xlog COMMIT PREPARE recording. When Greenplum encountered an orphaned prepared transaction that was committed after the xlog was recorded, it returned the error message:

incorrectly reset to 0. This could cause an infinite loop of retries and failures and lead to the pg_xlog file growing in size with the checkpointer process hanging. The problem was resolved by ensuring "22023","recovery command file ""recovery.conf"" request for standby mode not specified",,,,,,,0,,"xlog.c",5465 Greenplum Database now supports server recovery in a non-continuous mode using standby_mode=on race condition between a checkpoint and xlog COMMIT PREPARE recording. When Greenplum encountered an orphaned prepared transaction that was committed after the xlog was recorded, it returned the error message:

incorrectly reset to 0. This could cause an infinite loop of retries and failures and lead to the pg_xlog file growing in size with the checkpointer process hanging. The problem was resolved by ensuring "22023","recovery command file ""recovery.conf"" request for standby mode not specified",,,,,,,0,,"xlog.c",5465 Greenplum Database now supports server recovery in a non-continuous mode using standby_mode=on race condition between a checkpoint and xlog COMMIT PREPARE recording. When Greenplum encountered an orphaned prepared transaction that was committed after the xlog was recorded, it returned the error message:

incorrectly reset to 0. This could cause an infinite loop of retries and failures and lead to the pg_xlog file growing in size with the checkpointer process hanging. The problem was resolved by ensuring "22023","recovery command file ""recovery.conf"" request for standby mode not specified",,,,,,,0,,"xlog.c",5465 Greenplum Database now supports server recovery in a non-continuous mode using standby_mode=on race condition between a checkpoint and xlog COMMIT PREPARE recording. When Greenplum encountered an orphaned prepared transaction that was committed after the xlog was recorded, it returned the error message:

incorrectly reset to 0. This could cause an infinite loop of retries and failures and lead to the pg_xlog file growing in size with the checkpointer process hanging. The problem was resolved by ensuring "22023","recovery command file ""recovery.conf"" request for standby mode not specified",,,,,,,0,,"xlog.c",5465 Greenplum Database now supports server recovery in a non-continuous mode using standby_mode=on race condition between a checkpoint and xlog COMMIT PREPARE recording. When Greenplum encountered an orphaned prepared transaction that was committed after the xlog was recorded, it returned the error message: