本文分享自華爲雲社區《【防過載檢查項】》，作者： 譡裏個檔。
1. GUC參數檢查
目的：針對不同版本建議設定不同的參數值，當前先檢查出來，後續diagnosis會給出建議值
SELECT
split_part((substring(version() from '\((.*)\)')), ' ', 2) AS version,
(EXISTS (SELECT 1 FROM (SELECT count(DISTINCT node_name) AS dn_cnt
FROM pgxc_node WHERE node_type = 'D' AND node_host <> 'localhost' GROUP BY node_host) WHERE dn_cnt <> 1)
) AS sdn_per_node,
node_name, name AS guc_name,
CASE WHEN unit = 'ms' THEN setting::bigint/1000 || ' s'
WHEN unit = 's' THEN setting || ' s'
WHEN unit = 'KB' THEN pg_size_pretty(setting*1024)
WHEN unit = '8KB' THEN pg_size_pretty(setting*1024*8)
ELSE setting
END AS setting, unit,
CASE WHEN name = 'default_distribution_mode' THEN
CASE WHEN setting <> 'roundrobin' THEN 'roundrobin' END
WHEN name = 'autovacuum' THEN
CASE WHEN setting <> 'on' THEN 'on' END
WHEN name = 'autovacuum_max_workers' THEN
CASE WHEN setting::int > 6 THEN '6' END
WHEN name = 'autovacuum_max_workers' THEN
CASE WHEN setting::int > 3 THEN '3' END
WHEN name = 'session_timeout' THEN
CASE WHEN setting::int > 600 OR setting::int = 0 THEN '<= 10min' END
WHEN name = 'statement_timeout' THEN
CASE WHEN setting::int > 24* 60 * 1000 OR setting::int = 0 THEN '<= 24h' END
WHEN name = 'sql_use_spacelimit' THEN
CASE WHEN setting::int > 300*1024*1024 OR setting IN ('0', '-1') THEN '<= 300GB' END
WHEN name = 'temp_file_limit' THEN
CASE WHEN setting::int > 100*1024*1024 OR setting IN ('0', '-1') THEN '<= 100GB' END
WHEN name = 'udf_memory_limit' THEN
CASE WHEN setting::int > 1*1024*1024 THEN '<= 1GB' END
WHEN name = 'query_dop' THEN
CASE WHEN setting::int = 0 THEN ' -4' END
WHEN name = 'max_streams_per_query' THEN
CASE WHEN setting::int = -1 THEN '50' END
WHEN name = 'max_connections' THEN
CASE WHEN node_name = 'dn_6001_6002' AND setting::int < 5000 THEN '>=5000' END
ELSE ''
END AS diagnosis
FROM pgxc_parallel_query('all', 'SELECT
pgxc_node_str() AS node_name, name, setting, unit
FROM pg_settings
WHERE pgxc_node_str() IN (''cn_5001'',''dn_6001_6002'')
AND name in (''max_streams_per_query'',''query_dop''
,''sql_use_spacelimit'',''temp_file_limit'',''default_distribution_mode'',
''autovacuum_mode'',''autovacuum'',''autovacuum_max_workers_hstore'',''autovacuum_max_workers'',
''session_timeout'',''statement_timeout'',''ddl_lock_timeout'',''idle_in_transaction_timeout'',
''max_connections'',''min_pool_size'',''max_pool_size'',
''max_stream_pool'',''max_active_statements'',''max_prepared_transactions'',
''cstore_buffers'',''shared_buffers'', ''max_process_memory'', ''udf_memory_limit'',
''max_process_memory_balanced'', ''bbox_dump_count'', ''enable_bbox_dump'')')
AS (node_name name, name text, setting text, unit text)
ORDER BY node_name, name;
2. 大表檢查
目的：識別大表，建議客戶整改，避免磁盤過載
8.1.3版本使用如下SQL
SELECT
CASE WHEN (skewsize > avgsize * 0.10 AND skewsize > 50 * 1024) THEN 'skew table'
WHEN (reloptions::text LIKE '%orientation=column%' AND reloptions::text LIKE '%compression=no%') THEN 'uncompressed column table'
WHEN (x.pclocatortype = 'R' AND avgsize > 10 * 1024) THEN 'large replicattion table'
WHEN (pg_stat_get_dead_tuples(c.oid) >100000 AND pg_stat_get_dead_tuples(c.oid)/(pg_stat_get_dead_tuples(c.oid)+pg_stat_get_live_tuples(c.oid)) > 0.4) THEN 'dirty table'
ELSE 'normal large table'
END AS diagnostic,
t1.schemaname, -- 表的schema
t1.tablename, -- 表名
a.rolname AS tableowner,
x.pgroup AS nodegroup,
CASE x.pclocatortype WHEN 'H' THEN 'Hash'
WHEN 'N' THEN 'Round Robin'
WHEN 'R' THEN 'Replicate'
END AS locatortype,
CASE WHEN c.parttype = 'p' THEN true ELSE false END AS ispartitioned,
CASE WHEN reloptions::text LIKE '%orientation=column%' THEN 'column'
WHEN reloptions::text LIKE '%orientation=row%' THEN 'row'
END AS orientation,
t1.dnnum, -- 表的node group的DN數
t1.totalsize AS "totalsize(MB)", -- 表的size ，單位MB
t1.avgsize AS "avgsize(MB)", -- 平均每個DN上數據量，單位MB
t1.skewsize AS "skewsize(MB)", -- 不同DN上數據size的最大差值，單位MB
t1.skewdn, -- 數據量最大的DN
t1.maxratio, -- 數據量最大DN的size/平均size
t1.minratio, -- 數據量最小DN的size/平均size
t1.skewratio -- 不同DN上數據size的最大差值/平均size
FROM ( -- 預處理，識別傾斜表
SELECT
schemaname,
tablename,
skewdn,
dnnum,
totalsize,
avgsize,
skewsize,
(maxsize/avgsize)::numeric(20,2) AS maxratio,
(minsize/avgsize)::numeric(20,2) AS minratio,
(skewsize/avgsize)::numeric(20,2) AS skewratio
FROM (
SELECT
schemaname,tablename,skewdn,count(1) AS dnnum,sum(dnsize) AS totalsize,
avg(dnsize) AS avgsize,max(dnsize) AS maxsize,min(dnsize) AS minsize, (max(dnsize) - min(dnsize)) AS skewsize
FROM ( --對每個表的數據按照DN數據量大小排序，以及獲取傾斜的dn
SELECT
schemaname,
tablename,
nodename,
(dnsize/1024/1024)::bigint AS dnsize, -- 單位換算爲MB
first_value(nodename) over (PARTITION BY schemaname, tablename ORDER BY dnsize DESC, nodename) AS skewdn -- --數據量最大的DN
FROM ( -- 獲取大於10GB的表
SELECT
schemaname, tablename,(rd).nodename, ((rd).dnsize + 1) AS dnsize
FROM (
SELECT
schemaname, tablename, gs_table_distribution(schemaname, tablename) AS rd
FROM gs_table_distribution()
WHERE schemaname NOT IN ('pg_catalog', 'dbms_om', 'cstore')
AND relkind = 'r'
GROUP BY schemaname, tablename
HAVING sum(dnsize) > 50.0 * 1024 * 1024 * 1024 -- 總大小大於100GB
)
)
)
GROUP BY schemaname,tablename, skewdn
)
) t1
INNER JOIN pg_class c ON c.relname = t1.tablename
LEFT JOIN pg_namespace n ON n.nspname = t1.schemaname
LEFT JOIN pg_authid a ON a.oid = c.relowner
LEFT JOIN pgxc_class x ON x.pcrelid = c.oid
WHERE c.reloptions::text NOT LIKE '%internal_mask%'
ORDER BY totalsize DESC, diagnostic, skewsize DESC
;
8.2.1和8.2.0版本使用如下
-- 大表診斷
SELECT
CASE WHEN (skewsize > avgsize * 0.10 AND skewsize > 50 * 1024) THEN 'skew table'
WHEN (reloptions::text LIKE '%orientation=column%' AND reloptions::text LIKE '%compression=no%') THEN 'uncompressed column table'
WHEN (x.pclocatortype = 'R' AND avgsize > 10 * 1024) THEN 'large replicattion table'
WHEN (pg_stat_get_dead_tuples(c.oid) >100000 AND pg_stat_get_dead_tuples(c.oid)/(pg_stat_get_dead_tuples(c.oid)+pg_stat_get_live_tuples(c.oid)) > 0.4) THEN 'dirty table'
WHEN (reloptions::text LIKE '%orientation=column%') THEN
CASE WHEN (SELECT total_cu_count > 0 AND (zero_size_cu_count + small_cu_count)/total_cu_count > 0.5 FROM get_col_cu_info(t1.schemaname, t1.tablename)) THEN 'small cu table'
ELSE 'normal large table'
END
ELSE 'normal large table'
END AS diagnostic,
t1.schemaname, -- 表的schema
t1.tablename, -- 表名
a.rolname AS tableowner,
x.pgroup AS nodegroup,
CASE x.pclocatortype WHEN 'H' THEN 'Hash'
WHEN 'N' THEN 'Round Robin'
WHEN 'R' THEN 'Replicate'
END AS locatortype,
CASE WHEN c.parttype = 'p' THEN true ELSE false END AS ispartitioned,
CASE WHEN reloptions::text LIKE '%orientation=column%' THEN 'column'
WHEN reloptions::text LIKE '%orientation=row%' THEN 'row'
END AS orientation,
t1.dnnum, -- 表的node group的DN數
t1.totalsize AS "totalsize(MB)", -- 表的size ，單位MB
t1.avgsize AS "avgsize(MB)", -- 平均每個DN上數據量，單位MB
t1.skewsize AS "skewsize(MB)", -- 不同DN上數據size的最大差值，單位MB
t1.skewdn, -- 數據量最大的DN
t1.maxratio, -- 數據量最大DN的size/平均size
t1.minratio, -- 數據量最小DN的size/平均size
t1.skewratio -- 不同DN上數據size的最大差值/平均size
FROM ( -- 預處理，識別傾斜表
SELECT
schemaname,
tablename,
skewdn,
dnnum,
totalsize,
avgsize,
skewsize,
(maxsize/avgsize)::numeric(20,2) AS maxratio,
(minsize/avgsize)::numeric(20,2) AS minratio,
(skewsize/avgsize)::numeric(20,2) AS skewratio
FROM (
SELECT
schemaname,
tablename,
skewdn,
count(1) AS dnnum,
sum(dnsize) AS totalsize,
avg(dnsize) AS avgsize,
max(dnsize) AS maxsize,
min(dnsize) AS minsize,
(max(dnsize) - min(dnsize)) AS skewsize
FROM ( --對每個表的數據按照DN數據量大小排序，以及獲取傾斜的dn
SELECT
schemaname,
tablename,
nodename,
(dnsize/1024/1024)::bigint AS dnsize, -- 單位換算爲MB
first_value(nodename) over (PARTITION BY schemaname, tablename ORDER BY dnsize DESC, nodename) AS skewdn -- --數據量最大的DN
FROM ( -- 獲取大於10GB的表
SELECT
schemaname, tablename,(rd).nodename, ((rd).dnsize + 1) AS dnsize
FROM (
SELECT
schemaname, tablename, gs_table_distribution(schemaname, tablename) AS rd
FROM gs_table_distribution()
WHERE schemaname NOT IN ('pg_catalog', 'dbms_om', 'cstore')
AND relkind = 'r'
GROUP BY schemaname, tablename
HAVING sum(dnsize) > 50* 1024 * 1024 * 1024.0 -- 總大小大於100GB
)
)
)
GROUP BY schemaname,tablename, skewdn
)
) t1
INNER JOIN pg_class c ON c.relname = t1.tablename
LEFT JOIN pg_namespace n ON n.nspname = t1.schemaname
LEFT JOIN pg_authid a ON a.oid = c.relowner
LEFT JOIN pgxc_class x ON x.pcrelid = c.oid
WHERE c.reloptions::text NOT LIKE '%internal_mask%'
ORDER BY totalsize DESC, diagnostic, skewsize DESC
;
8.3.0版本使用
-- 大表診斷
SELECT
CASE WHEN (skewsize > avgsize * 0.10 AND skewsize > 50 * 1024) THEN 'skew table'
WHEN (reloptions::text LIKE '%orientation=column%' AND reloptions::text LIKE '%compression=no%') THEN 'uncompressed column table'
WHEN (x.pclocatortype = 'R' AND avgsize > 10 * 1024) THEN 'large replicattion table'
WHEN (pg_stat_get_dead_tuples(c.oid) >100000 AND pg_stat_get_dead_tuples(c.oid)/(pg_stat_get_dead_tuples(c.oid)+pg_stat_get_live_tuples(c.oid)) > 0.4) THEN 'dirty table'
WHEN (reloptions::text LIKE '%orientation=column%') THEN
CASE WHEN (SELECT total_cu_count > 0 AND (zero_cu_count + small_cu_count)/total_cu_count > 0.5 FROM pgxc_get_small_cu_info(c.oid)) THEN 'small cu table'
ELSE 'normal large table'
END
ELSE 'normal large table'
END AS diagnostic,
t1.schemaname, -- 表的schema
t1.tablename, -- 表名
a.rolname AS tableowner,
x.pgroup AS nodegroup,
CASE x.pclocatortype WHEN 'H' THEN 'Hash'
WHEN 'N' THEN 'Round Robin'
WHEN 'R' THEN 'Replicate'
END AS locatortype,
CASE WHEN c.parttype = 'p' THEN true ELSE false END AS ispartitioned,
CASE WHEN reloptions::text LIKE '%orientation=column%' THEN 'column'
WHEN reloptions::text LIKE '%orientation=row%' THEN 'row'
END AS orientation,
t1.dnnum, -- 表的node group的DN數
t1.totalsize AS "totalsize(MB)", -- 表的size ，單位MB
t1.avgsize AS "avgsize(MB)", -- 平均每個DN上數據量，單位MB
t1.skewsize AS "skewsize(MB)", -- 不同DN上數據size的最大差值，單位MB
t1.skewdn, -- 數據量最大的DN
t1.maxratio, -- 數據量最大DN的size/平均size
t1.minratio, -- 數據量最小DN的size/平均size
t1.skewratio -- 不同DN上數據size的最大差值/平均size
FROM ( -- 預處理，識別傾斜表
SELECT
schemaname,
tablename,
skewdn,
dnnum,
totalsize,
avgsize,
skewsize,
(maxsize/avgsize)::numeric(20,2) AS maxratio,
(minsize/avgsize)::numeric(20,2) AS minratio,
(skewsize/avgsize)::numeric(20,2) AS skewratio
FROM (
SELECT
schemaname,
tablename,
skewdn,
count(1) AS dnnum,
sum(dnsize) AS totalsize,
avg(dnsize) AS avgsize,
max(dnsize) AS maxsize,
min(dnsize) AS minsize,
(max(dnsize) - min(dnsize)) AS skewsize
FROM ( --對每個表的數據按照DN數據量大小排序，以及獲取傾斜的dn
SELECT
schemaname,
tablename,
nodename,
(dnsize/1024/1024)::bigint AS dnsize, -- 單位換算爲MB
first_value(nodename) over (PARTITION BY schemaname, tablename ORDER BY dnsize DESC, nodename) AS skewdn -- --數據量最大的DN
FROM ( -- 獲取大於10GB的表
SELECT
schemaname, tablename,(rd).nodename, ((rd).dnsize + 1) AS dnsize
FROM (
SELECT
schemaname, tablename, gs_table_distribution(schemaname, tablename) AS rd
FROM gs_table_distribution()
WHERE schemaname NOT IN ('pg_catalog', 'dbms_om', 'cstore')
AND relkind = 'r'
GROUP BY schemaname, tablename
HAVING sum(dnsize) > 50.0 * 1024 * 1024 * 1024 -- 總大小大於100GB
)
)
)
GROUP BY schemaname,tablename, skewdn
)
) t1
INNER JOIN pg_class c ON c.relname = t1.tablename
LEFT JOIN pg_namespace n ON n.nspname = t1.schemaname
LEFT JOIN pg_authid a ON a.oid = c.relowner
LEFT JOIN pgxc_class x ON x.pcrelid = c.oid
WHERE c.reloptions::text NOT LIKE '%internal_mask%'
ORDER BY totalsize DESC, diagnostic, skewsize DESC
;
針對不同的診斷結果使用如下診斷措施
| 類別 |
建議手段 |
| skew table |
根據業務選擇關聯常用，並且數據不傾斜的列作爲修改分佈列，如果找不到合適的分佈列，可以把表修改爲RoundRobin分佈 ALTER TABLE ctmes_tgmesmbi.mbi_hlottsthisDISTRIBUTE BY ROUNDROBIN;複製
|
| uncompressed column table |
列存表的壓縮效果非常好，一般推薦使用壓縮，至少使用low級別壓縮 ALTER TABLE customer_address SET(compression=low); VACUUM FULL customer_address;複製
|
| dirty table |
說明表檢測碎片率比較高，需要通過VACUUM整理表 VACUUM FULL customer_address;複製
|
| small cu table |
說明表小CU比較多，需要通過VACUUM整理表 VACUUM FULL customer_address;複製
|
| large replicattion table |
複製表在每個DN上都有一份全量數據，當表磁盤空間佔用過大時，需要把表修改爲HASH表。一般複製表都是維表，存在主鍵，直接把分佈列修改爲主鍵 ALTER TABLE ctmes_tgmesmbi.mbi_hlottsthisDISTRIBUTE BY HASH(id);複製 |
| normal large table |
3. 冗餘索引診斷
目的：識別冗餘索引，建議客戶刪除。可以降低磁盤空間，並降低大規模數據導入的時候的xlog規模
-- optimizable policy爲duplicate的檢查項
-- 表明兩個索引字段和字段順序完全一致
-- 建議直接刪除optimizable index指定的索引；
-- optimizable policy爲redundancy檢查項表明
-- optimizable index指定的索引的索引列剛好是base index的索引列的前面字段
-- 建議直接刪除optimizable index指定的索引；
-- optimizable policy爲optimizable檢查項
-- 表明optimizable index和base index這兩個索引的索引列完全重複，但是索引列的順序不一致
-- 這種場景需要人工介入分析是否可以優化
WITH info AS(
SELECT
quote_ident(n.nspname) || '.' || quote_ident(c.relname) AS tablename,
pgroup AS nodegroup,
x.indrelid AS indrelid,
x.indexrelid AS indexrelid,
indisunique,
indisprimary,
indnatts,
indkey,
indexprs
FROM pg_index x
INNER JOIN pg_class c ON c.oid = x.indrelid
INNER JOIN pg_class i ON i.oid = x.indexrelid
LEFT JOIN pg_namespace n ON n.oid = c.relnamespace
INNER JOIN pgxc_class xc ON xc.pcrelid = c.oid
WHERE c.relkind = 'r' AND c.oid >= 16384 AND (c.reloptions IS NULL OR c.reloptions::text NOT LIKE '%internal_mask%')
AND i.relkind = 'i' AND i.oid >= 16384
AND x.indpred IS NULL
),
base AS(
SELECT
tablename,
nodegroup,
i.indrelid,
i.indexrelid baseidx,
i.indisunique AS base_unique,
i.indisprimary AS base_primary,
x.indexrelid AS optidx,
x.indisunique AS opt_unique,
x.indisprimary AS opt_primary,
CASE WHEN opt_primary > base_primary OR opt_unique > base_unique THEN true ELSE false END AS swap,
CASE WHEN i.indkey = x.indkey AND coalesce(pg_get_expr(i.indexprs, i.indrelid), 'NULL') = coalesce(pg_get_expr(x.indexprs, x.indrelid), 'NULL') THEN 'duplicate'::text
WHEN x.indexprs IS NOT NULL OR i.indexprs IS NOT NULL THEN NULL::text
WHEN strpos(i.indkey::text, x.indkey::text||' ') = 1 OR strpos(x.indkey::text, i.indkey::text||' ') = 1 THEN 'redundancy'::text
WHEN i.indkey @> x.indkey AND x.indkey @> i.indkey THEN 'optimizable'::text
ELSE NULL
END AS optpolicy
FROM info i
INNER JOIN pg_index x ON (i.indrelid = x.indrelid AND i.indexrelid > x.indexrelid)
WHERE x.indpred IS NULL AND optpolicy IS NOT NULL
),
tmp AS(
SELECT
tablename,
indrelid,
nodegroup,
CASE WHEN swap THEN optidx ELSE baseidx END AS base_idx,
CASE WHEN swap THEN opt_primary ELSE base_primary END AS base_primary,
CASE WHEN swap THEN opt_unique ELSE base_unique END AS base_unique,
CASE WHEN swap THEN baseidx ELSE optidx END AS opt_idx,
CASE WHEN swap THEN base_primary ELSE opt_primary END AS opt_primary,
CASE WHEN swap THEN base_unique ELSE opt_unique END AS opt_unique,
optpolicy
FROM base
)
SELECT
tablename,
nodegroup,
base_idx::regclass::text AS base_index,
base_primary,
base_unique,
substring(pg_get_indexdef(base_idx) from 'USING .+\)') AS base_idxdef,
opt_idx::regclass::text AS opt_index,
opt_primary,
opt_unique,
substring(pg_get_indexdef(opt_idx) from 'USING .+\)') AS opt_idxdef,
optpolicy,
pg_get_tabledef(indrelid)
FROM tmp
ORDER BY 1, 2, 3
;