pg_dump 使 PostgreSQL 服务器崩溃

Question

在包含大量 blob 的数据库上运行 pg_dump，当执行此查询时 PostgreSQL 崩溃：

pg_dump: reading large objects
pg_dump: error: query failed: SSL SYSCALL error: EOF detected
pg_dump: error: query was: SELECT l.oid, (SELECT rolname FROM pg_catalog.pg_roles WHERE oid = l.lomowner) AS rolname, (SELECT pg_catalog.array_agg(acl ORDER BY row_n) FROM (SELECT acl, row_n FROM pg_catalog.unnest(coalesce(l.lomacl,pg_catalog.acldefault('L',l.lomowner))) WITH ORDINALITY AS perm(acl,row_n) WHERE NOT EXISTS ( SELECT 1 FROM pg_catalog.unnest(coalesce(pip.initprivs,pg_catalog.acldefault('L',l.lomowner))) AS init(init_acl) WHERE acl = init_acl)) as foo) AS lomacl, (SELECT pg_catalog.array_agg(acl ORDER BY row_n) FROM (SELECT acl, row_n FROM pg_catalog.unnest(coalesce(pip.initprivs,pg_catalog.acldefault('L',l.lomowner))) WITH ORDINALITY AS initp(acl,row_n) WHERE NOT EXISTS ( SELECT 1 FROM pg_catalog.unnest(coalesce(l.lomacl,pg_catalog.acldefault('L',l.lomowner))) AS permp(orig_acl) WHERE acl = orig_acl)) as foo) AS rlomacl, NULL AS initlomacl, NULL AS initrlomacl FROM pg_largeobject_metadata l LEFT JOIN pg_init_privs pip ON (l.oid = pip.objoid AND pip.classoid = 'pg_largeobject'::regclass AND pip.objsubid = 0)

我已经对查询进行了实验，lomacl 和 rlomacl 的两个 array_agg() 列似乎是罪魁祸首。

这是 AWS Aurora PostgreSQL 11：

SELECT version();
                                             version
-------------------------------------------------------------------------------------------------
PostgreSQL 11.9 on x86_64-pc-linux-gnu, compiled by x86_64-pc-linux-gnu-gcc (GCC) 7.4.0, 64-bit

日志：

2021-08-19 19:47:46 UTC::@:[46753]:LOG: server process (PID 21837) was terminated by signal 9: Killed
2021-08-19 19:47:46 UTC::@:[46753]:DETAIL: Failed process was running: SELECT l.oid, (SELECT rolname FROM pg_catalog.pg_roles WHERE oid = l.lomowner) AS rolname, (SELECT pg_catalog.array_agg(acl ORDER BY row_n) FROM (SELECT acl, row_n FROM pg_catalog.unnest(coalesce(l.lomacl,pg_catalog.acldefault('L',l.lomowner))) WITH ORDINALITY AS perm(acl,row_n) WHERE NOT EXISTS ( SELECT 1 FROM pg_catalog.unnest(coalesce(pip.initprivs,pg_catalog.acldefault('L',l.lomowner))) AS init(init_acl) WHERE acl = init_acl)) as foo) AS lomacl, (SELECT pg_catalog.array_agg(acl ORDER BY row_n) FROM (SELECT acl, row_n FROM pg_catalog.unnest(coalesce(pip.initprivs,pg_catalog.acldefault('L',l.lomowner))) WITH ORDINALITY AS initp(acl,row_n) WHERE NOT EXISTS ( SELECT 1 FROM pg_catalog.unnest(coalesce(l.lomacl,pg_catalog.acldefault('L',l.lomowner))) AS permp(orig_acl) WHERE acl = orig_acl)) as foo) AS rlomacl, NULL AS initlomacl, NULL AS initrlomacl FROM pg_largeobject_metadata l LEFT JOIN pg_init_privs pip ON (l.oid = pip.objoid AND pip.classoid = 'pg_largeobject'::regclass AND pip.objsubid = 0)
2021-08-19 19:47:46 UTC::@:[46753]:LOG: terminating any other active server processes
2021-08-19 19:47:46 UTC::@:[46753]:FATAL: Can't handle storage runtime process crash
2021-08-19 19:47:46 UTC::@:[46753]:LOG: database system is shut down

任何故障排除步骤/建议？

Answer 1

pg_dump 使用 pg_largeobject_metadata 和 pg_init_privs 之间的连接来获取要转储的大型对象 OID 列表。

现在要么数据库服务器上的内存太低，要么你有很多大对象，而你的work_mem 设置得非常高，以至于数据库服务器机器内存不足。由于您没有在数据库服务器操作系统上禁用内存过量使用，OOM 杀手会杀死您的进程。

要么增加可用 RAM，要么使用更保守的 work_mem 设置。我必须补充一点，在 v13 之前，PostgreSQL 很容易错误地创建大于 work_mem 的哈希值。也许您可以在转储期间将enable_hashjoin 设置为off。