TEST01数据来自DBA_OBJECTS，重复512次，共4500W行，创建2个索引
create index idx_test01_n1 on test01(object_id,data_object_id);
create index idx_test01_n2 on test01(data_object_id,object_id);测试SQL
select *
from test01 where object_id in(2,3)
start with object_id = 2
connect by nocycle prior object_id = data_object_id;
select count(*)
from test01
start with object_id = 2
connect by nocycle prior object_id = data_object_id;注意：生产环境不会有这样的大量重复数据，object_id和data_object_id也不是层级关系，严格来说我这个测试不严谨(虽然不严谨，但是能反映问题)，主要是数据不好构造，见谅
Oracle19c
SQL1
SQL> select *
2 from test01 where object_id in(2,3)
3 start with object_id = 2
4 connect by nocycle prior object_id = data_object_id;
512 rows selected.
Elapsed: 00:00:01.98
Execution Plan
----------------------------------------------------------
Plan hash value: 3708411942
-------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 261K| 51M| 180K (4)| 00:00:08 |
|* 1 | FILTER | | | | | |
|* 2 | CONNECT BY WITH FILTERING | | | | | |
| 3 | TABLE ACCESS BY INDEX ROWID BATCHED| TEST01 | 505 | 49490 | 510 (0)| 00:00:01 |
|* 4 | INDEX RANGE SCAN | IDX_TEST01_N1 | 505 | | 4 (0)| 00:00:01 |
|* 5 | HASH JOIN | | 260K| 27M| 173K (1)| 00:00:07 |
| 6 | CONNECT BY PUMP | | | | | |
|* 7 | TABLE ACCESS FULL | TEST01 | 4696K| 438M| 172K (1)| 00:00:07 |
-------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("TEST01"."OBJECT_ID"=2 OR "TEST01"."OBJECT_ID"=3)
2 - access("TEST01"."DATA_OBJECT_ID"=PRIOR "TEST01"."OBJECT_ID")
4 - access("OBJECT_ID"=2)
5 - access("connect$_by$_pump$_002"."prior object_id "="DATA_OBJECT_ID")
7 - filter("DATA_OBJECT_ID" IS NOT NULL)
Note
-----
- this is an adaptive plan
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
19032 consistent gets
0 physical reads
0 redo size
7595 bytes sent via SQL*Net to client
820 bytes received via SQL*Net from client
36 SQL*Net roundtrips to/from client
5 sorts (memory)
0 sorts (disk)
512 rows processedSQL2
SQL> select count(*)
2 from test01
3 start with object_id = 2
4 connect by nocycle prior object_id = data_object_id;
Elapsed: 00:00:01.34
Execution Plan
----------------------------------------------------------
Plan hash value: 3365808593
--------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 26 | 8172 (20)| 00:00:01 |
| 1 | SORT AGGREGATE | | 1 | 26 | | |
|* 2 | CONNECT BY WITH FILTERING| | | | | |
|* 3 | INDEX RANGE SCAN | IDX_TEST01_N1 | 505 | 3535 | 4 (0)| 00:00:01 |
| 4 | NESTED LOOPS | | 260K| 5088K| 6585 (1)| 00:00:01 |
| 5 | CONNECT BY PUMP | | | | | |
|* 6 | INDEX RANGE SCAN | IDX_TEST01_N2 | 516 | 3612 | 13 (0)| 00:00:01 |
--------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("DATA_OBJECT_ID"=PRIOR "OBJECT_ID")
3 - access("OBJECT_ID"=2)
6 - access("connect$_by$_pump$_002"."prior object_id "="DATA_OBJECT_ID")
filter("DATA_OBJECT_ID" IS NOT NULL)
Note
-----
- this is an adaptive plan
Statistics
----------------------------------------------------------
1 recursive calls
0 db block gets
88 consistent gets
0 physical reads
0 redo size
362 bytes sent via SQL*Net to client
429 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
5 sorts (memory)
0 sorts (disk)
1 rows processed两条SQL默认走的是CONNECT_BY_FILTERING(/*+ CONNECT_BY_FILTERING /)
测试NO_CONNECT_BY_FILTERING(/+ NO_CONNECT_BY_FILTERING */)看看速度
SQL1
SQL> select /*+ NO_CONNECT_BY_FILTERING */ *
2 from test01 where object_id in(2,3)
3 start with object_id = 2
4 connect by nocycle prior object_id = data_object_id;
512 rows selected.
Elapsed: 00:01:14.85
Execution Plan
----------------------------------------------------------
Plan hash value: 1234393137
---------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
---------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 261K| 51M| 1166K (86)| 00:00:46 |
|* 1 | FILTER | | | | | |
|* 2 | CONNECT BY NO FILTERING WITH START-WITH| | | | | |
| 3 | TABLE ACCESS FULL | TEST01 | 44M| 4162M| 173K (1)| 00:00:07 |
---------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("TEST01"."OBJECT_ID"=2 OR "TEST01"."OBJECT_ID"=3)
2 - access("TEST01"."DATA_OBJECT_ID"=PRIOR "TEST01"."OBJECT_ID")
filter("OBJECT_ID"=2)
Statistics
----------------------------------------------------------
5449 recursive calls
1596691 db block gets
635906 consistent gets
676955 physical reads ---SQL已经反复多次执行，buffer cache确保能容纳TEST01，关闭了直接路径读
0 redo size
7595 bytes sent via SQL*Net to client
851 bytes received via SQL*Net from client
36 SQL*Net roundtrips to/from client
1 sorts (memory)
1 sorts (disk)
512 rows processedSQL2
SQL> select /*+ NO_CONNECT_BY_FILTERING */ count(*)
2 from test01
3 start with object_id = 2
4 connect by nocycle prior object_id = data_object_id;
Elapsed: 00:00:43.64
Execution Plan
----------------------------------------------------------
Plan hash value: 3439160689
-----------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 26 | 325K (48)| 00:00:13 |
| 1 | SORT AGGREGATE | | 1 | 26 | | |
|* 2 | CONNECT BY WITHOUT FILTERING| | | | | |
|* 3 | INDEX RANGE SCAN | IDX_TEST01_N1 | 505 | 3535 | 4 (0)| 00:00:01 |
| 4 | TABLE ACCESS FULL | TEST01 | 44M| 297M| 172K (1)| 00:00:07 |
-----------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("DATA_OBJECT_ID"=PRIOR "OBJECT_ID")
3 - access("OBJECT_ID"=2)
Statistics
----------------------------------------------------------
727 recursive calls
1660389 db block gets
635910 consistent gets
90036 physical reads ---SQL已经反复多次执行，buffer cache确保能容纳TEST01，关闭了直接路径读
0 redo size
362 bytes sent via SQL*Net to client
460 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
1 sorts (disk)
1 rows processed
SQL> select /*+ NO_CONNECT_BY_FILTERING FULL(TEST01) */ count(*)
2 from test01
3 start with object_id = 2
4 connect by nocycle prior object_id = data_object_id;
Elapsed: 00:00:44.11
Execution Plan
----------------------------------------------------------
Plan hash value: 1977146487
---------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
---------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 26 | 325K (48)| 00:00:13 |
| 1 | SORT AGGREGATE | | 1 | 26 | | |
|* 2 | CONNECT BY NO FILTERING WITH START-WITH| | | | | |
| 3 | TABLE ACCESS FULL | TEST01 | 44M| 297M| 172K (1)| 00:00:07 |
---------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("DATA_OBJECT_ID"=PRIOR "OBJECT_ID")
filter("OBJECT_ID"=2)
Hint Report (identified by operation id / Query Block Name / Object Alias):
Total hints for statement: 1 (N - Unresolved (1))
---------------------------------------------------------------------------
1 - SEL$1
N - FULL(TEST01)
Statistics
----------------------------------------------------------
727 recursive calls
1660389 db block gets
635906 consistent gets
90030 physical reads ---SQL已经反复多次执行，buffer cache确保能容纳TEST01，关闭了直接路径读
0 redo size
362 bytes sent via SQL*Net to client
473 bytes received via SQL*Net from client
2 SQL*Net roundtrips to/from client
1 sorts (memory)
1 sorts (disk)
1 rows processed崖山23.5.1
SQL1
SQL> select *
from test01 where object_id in(2,3)
start with object_id = 2
connect by nocycle prior object_id = data_object_id; 2 3 4
Execution Plan
----------------------------------------------------------------
SQL hash value: 2436060758
Optimizer: ADOPT_C
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| Id | Operation type | Name | Owner | E - Rows | A - Rows | Cost(%CPU) | A - Time | Loops | Memory | Disk | Partition info |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| 0 | SELECT STATEMENT | | | | | | | | | | |
|* 1 | RESULT | | | 768| | 1089546( 0)| | | | | |
|* 2 | CONNECT BY HASH | | | 44537344| | 1086511( 0)| | | | | |
| 3 | TABLE ACCESS FULL | TEST01 | SCOTT | 44537344| | 1086015( 0)| | | | | |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
Operation Information (identified by operation id):
---------------------------------------------------
1 - Predicate : filter("TEST01"."OBJECT_ID" IN BINARY SEARCH : [2, 3])
2 - Predicate : access(PRIOR "TEST01"."OBJECT_ID" = "TEST01"."DATA_OBJECT_ID")
filter("TEST01"."OBJECT_ID" = 2)
Statistics
----------------------------------------------------------------------------------------------------
18 rows fetched.
Elapsed: 00:00:33.328SQL2
SQL> select count(*)
from test01
start with object_id = 2
connect by nocycle prior object_id = data_object_id; 2 3 4
Execution Plan
----------------------------------------------------------------
SQL hash value: 3323709368
Optimizer: ADOPT_C
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| Id | Operation type | Name | Owner | E - Rows | A - Rows | Cost(%CPU) | A - Time | Loops | Memory | Disk | Partition info |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
| 0 | SELECT STATEMENT | | | | | | | | | | |
| 1 | AGGREGATE | | | 1| | 5237( 0)| | | | | |
|* 2 | CONNECT BY HASH | | | 44537344| | 4480( 0)| | | | | |
| 3 | INDEX FAST FULL SCAN | IDX_TEST01_N1 | SCOTT | 44537344| | 3984( 0)| | | | | |
+----+--------------------------------+----------------------+------------+----------+----------+-------------+----------+----------+----------+----------+--------------------------------+
Operation Information (identified by operation id):
---------------------------------------------------
2 - Predicate : access(PRIOR "TEST01"."OBJECT_ID" = "TEST01"."DATA_OBJECT_ID")
filter("TEST01"."OBJECT_ID" = 2)
Statistics
----------------------------------------------------------------------------------------------------
17 rows fetched.
Elapsed: 00:00:09.036DM8
SQL1
512 rows got
1 #NSET2: [6153845268876035, 2479468763217, 596]
2 #PRJT2: [6153845268876035, 2479468763217, 596]; exp_num(15), is_atom(FALSE)
3 #HASH RIGHT SEMI JOIN2: [6153845268876035, 2479468763217, 596]; key_num(1), MEM_USED(0KB), DISK_USED(0KB) KEY(DMTEMPVIEW_889194340.colname=TEST01.OBJECT_ID) KEY_NULL_EQU(0)
4 #CONST VALUE LIST: [1, 2, 30]; row_num(2), col_num(1)
5 #HIERARCHICAL QUERY: [6153843615896860, 49589375264358, 596]; key_num(1)
6 #BLKUP2: [6541, 1113433, 596]; IDX_TEST01_N1(TEST01)
7 #SSEK2: [6541, 1113433, 596]; scan_type(ASC), IDX_TEST01_N1(TEST01), is_global(0), scan_range[(exp_cast(2),min),(exp_cast(2),max))
8 #CSCN2: [10110, 44537344, 596]; INDEX33555493(TEST01); btr_scan(1)
Statistics
-----------------------------------------------------------------
0 data pages changed
0 undo pages changed
36603084 logical reads
1094314 physical reads
0 redo size
67379 bytes sent to client
382 bytes received from client
4 roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
0 rows processed
32079 io wait time(ms) ---io等待了32秒，SQL已经反复多次执行
107915 exec time(ms)
used time: 00:01:34.376. Execute id is 708.SQL2
SQL> select count(*)
from test01
start with object_id = 2
connect by nocycle prior object_id = data_object_id;2 3 4
LINEID COUNT(*)
---------- --------------------
1 4456960
1 #NSET2: [3177098600204064, 1, 60]
2 #PRJT2: [3177098600204064, 1, 60]; exp_num(1), is_atom(FALSE)
3 #AAGR2: [3177098600204064, 1, 60]; grp_num(0), sfun_num(1), distinct_flag[0]; slave_empty(0)
4 #HIERARCHICAL QUERY: [3177098600204064, 49589375264358, 60]; key_num(1)
5 #SSEK2: [158, 1113433, 60]; scan_type(ASC), IDX_TEST01_N1(TEST01), is_global(0), scan_range[(exp_cast(2),min),(exp_cast(2),max))
6 #SSCN: [5219, 44537344, 60]; IDX_TEST01_N1(TEST01); btr_scan(1); is_global(0)
Statistics
-----------------------------------------------------------------
0 data pages changed
0 undo pages changed
22350276 logical reads
214093 physical reads
0 redo size
138 bytes sent to client
173 bytes received from client
1 roundtrips to/from client
0 sorts (memory)
1 sorts (disk)
0 rows processed
5769 io wait time(ms)
49413 exec time(ms)
used time: 00:00:49.413. Execute id is 711.GaussDB
SQL1
oracle=# explain analyze select *
oracle-# from test01 where object_id in(2,3)
oracle-# start with object_id = 2
oracle-# connect by nocycle prior object_id = data_object_id;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
CTE Scan on tmp_reuslt (cost=1597.86..1606.52 rows=4 width=788) (actual time=26.335..38347.584 rows=512 loops=1)
Filter: (object_id = ANY ('{2,3}'::numeric[]))
Rows Removed by Filter: 4456448
CTE tmp_reuslt
-> StartWith Operator (cost=4.52..1597.86 rows=385 width=195) (actual time=26.315..34366.881 rows=4456960 loops=1)
Start With pseudo atts: array_key_4
-> Recursive Union (cost=4.52..1597.86 rows=385 width=195) (actual time=0.133..25430.129 rows=4719104 loops=1)
-> Bitmap Heap Scan on test01 (cost=4.52..144.22 rows=35 width=195) (actual time=0.132..1.459 rows=512 loops=1)
Recheck Cond: (object_id = 2)
Heap Blocks: exact=512
-> Bitmap Index Scan on idx_test01_n1 (cost=0.00..4.51 rows=35 width=0) (actual time=0.078..0.078 rows=512 loops=1)
Index Cond: (object_id = 2)
-> Nested Loop (cost=4.52..144.59 rows=35 width=195) (actual time=16530.470..24550.782 rows=4718592 loops=4456960)
-> WorkTable Scan on tmp_reuslt (cost=0.00..0.02 rows=1 width=32) (actual time=239.643..486.658 rows=4456960 loops=4456960)
-> Bitmap Heap Scan on test01 (cost=4.52..144.22 rows=35 width=195) (actual time=12952.710..20084.627 rows=4718592 loops=4456960)
Recheck Cond: (data_object_id = tmp_reuslt.object_id)
Heap Blocks: exact=4718592
-> Bitmap Index Scan on idx_test01_n2 (cost=0.00..4.51 rows=35 width=0) (actual time=11181.021..11181.021 rows=4718592 loops=4456960)
Index Cond: (data_object_id = tmp_reuslt.object_id)
Total runtime: 38366.669 ms
(20 rows)
Time: 38374.623 msSQL2
oracle=# explain analyze select count(*)
oracle-# from test01
oracle-# start with object_id = 2
oracle-# connect by nocycle prior object_id = data_object_id;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------------------------------
Aggregate (cost=1606.52..1606.53 rows=1 width=8) (actual time=37241.527..37241.527 rows=1 loops=1)
CTE tmp_reuslt
-> StartWith Operator (cost=4.52..1597.86 rows=385 width=195) (actual time=24.808..33771.341 rows=4456960 loops=1)
Start With pseudo atts: array_key_4
-> Recursive Union (cost=4.52..1597.86 rows=385 width=195) (actual time=0.142..25120.971 rows=4719104 loops=1)
-> Bitmap Heap Scan on test01 (cost=4.52..144.22 rows=35 width=195) (actual time=0.141..1.207 rows=512 loops=1)
Recheck Cond: (object_id = 2)
Heap Blocks: exact=512
-> Bitmap Index Scan on idx_test01_n1 (cost=0.00..4.51 rows=35 width=0) (actual time=0.087..0.087 rows=512 loops=1)
Index Cond: (object_id = 2)
-> Nested Loop (cost=4.52..144.59 rows=35 width=195) (actual time=16282.481..24271.672 rows=4718592 loops=4456960)
-> WorkTable Scan on tmp_reuslt (cost=0.00..0.02 rows=1 width=32) (actual time=237.553..496.508 rows=4456960 loops=4456960)
-> Bitmap Heap Scan on test01 (cost=4.52..144.22 rows=35 width=195) (actual time=12734.711..19878.902 rows=4718592 loops=4456960)
Recheck Cond: (data_object_id = tmp_reuslt.object_id)
Heap Blocks: exact=4718592
-> Bitmap Index Scan on idx_test01_n2 (cost=0.00..4.51 rows=35 width=0) (actual time=10978.502..10978.502 rows=4718592 loops=4456960)
Index Cond: (data_object_id = tmp_reuslt.object_id)
-> CTE Scan on tmp_reuslt (cost=0.00..7.70 rows=385 width=0) (actual time=24.814..36609.173 rows=4456960 loops=1)
Total runtime: 37265.295 ms
(19 rows)
Time: 37266.820 ms测试结果汇总如下图所示
现在来研究一下Oracle为什么比其他数据库快那么多，先把SQL改写为WITH递归(以SQL1为例)
SQL> with cte(owner, object_name, subobject_name, object_id, data_object_id,
2 object_type, created, last_ddl_time, timestamp, status,
3 temporary, generated, secondary, namespace, edition_name,
4 lv, path) as
5 (select owner, object_name, subobject_name, object_id, data_object_id,
6 object_type, created, last_ddl_time, timestamp, status,
7 temporary, generated, secondary, namespace, edition_name,
8 1, '|' || object_id || '|'
9 from test01
10 where object_id = 2
11 union all
12 select /*+ use_nl(t,c) */
13 t.owner, t.object_name, t.subobject_name, t.object_id, t.data_object_id,
14 t.object_type, t.created, t.last_ddl_time, t.timestamp, t.status,
15 t.temporary, t.generated, t.secondary, t.namespace, t.edition_name,
16 c.lv + 1,
17 c.path || t.object_id || '|'
18 from test01 t join cte c
19 on c.object_id = t.data_object_id
20 where c.path not like '%|' || t.object_id || '|%'
21 )
22 select *
23 from cte
24 where object_id in (2,3);
512 rows selected.
Elapsed: 00:00:28.09
Execution Plan
----------------------------------------------------------
Plan hash value: 3294323809
-----------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 3839G| 7758T| 18E (0)|999:59:59 |
|* 1 | VIEW | | 3839G| 7758T| 18E (0)|999:59:59 |
| 2 | UNION ALL (RECURSIVE WITH) BREADTH FIRST| | | | | |
| 3 | TABLE ACCESS BY INDEX ROWID BATCHED | TEST01 | 505 | 49490 | 510 (0)| 00:00:01 |
|* 4 | INDEX RANGE SCAN | IDX_TEST01_N1 | 505 | | 4 (0)| 00:00:01 |
| 5 | NESTED LOOPS | | 3839G| 7423T| 18E (0)|999:59:59 |
| 6 | NESTED LOOPS | | 3869G| 7423T| 18E (0)|999:59:59 |
| 7 | RECURSIVE WITH PUMP | | | | | |
|* 8 | INDEX RANGE SCAN | IDX_TEST01_N2 | 26 | | 13 (0)| 00:00:01 |
| 9 | TABLE ACCESS BY INDEX ROWID | TEST01 | 26 | 2548 | 39 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter("OBJECT_ID"=2 OR "OBJECT_ID"=3)
4 - access("OBJECT_ID"=2)
8 - access("C"."OBJECT_ID"="T"."DATA_OBJECT_ID")
filter("T"."DATA_OBJECT_ID" IS NOT NULL AND "C"."PATH" NOT LIKE
'%|'||TO_CHAR("T"."OBJECT_ID")||'|%')
Hint Report (identified by operation id / Query Block Name / Object Alias):
Total hints for statement: 1 (U - Unused (1))
---------------------------------------------------------------------------
5 - SEL$F1D6E378 / C@SEL$1
U - use_nl(t,c)
Statistics
----------------------------------------------------------
209 recursive calls
29564091 db block gets
4541494 consistent gets
15368 physical reads
0 redo size
8227 bytes sent via SQL*Net to client
1711 bytes received via SQL*Net from client
36 SQL*Net roundtrips to/from client
2 sorts (memory)
1 sorts (disk)
512 rows processed注意观察执行计划，UNION ALL (RECURSIVE WITH) BREADTH FIRST说明WITH递归用的是广度优先算法，把SQL改写为WITH SEARCH高级递归写法，指定深度优先，执行计划显示为A-TIME样式
with cte (
owner, object_name, subobject_name, object_id, data_object_id,
object_type, created, last_ddl_time, timestamp, status,
temporary, generated, secondary, namespace, edition_name, lvl
) as (
select
owner, object_name, subobject_name, object_id, data_object_id,
object_type, created, last_ddl_time, timestamp, status,
temporary, generated, secondary, namespace, edition_name,
1
from test01
where object_id = 2
union all
select /*+ use_nl(t,c) */
t.owner, t.object_name, t.subobject_name, t.object_id, t.data_object_id,
t.object_type, t.created, t.last_ddl_time, t.timestamp, t.status,
t.temporary, t.generated, t.secondary, t.namespace, t.edition_name,
c.lvl + 1
from test01 t
join cte c on c.object_id = t.data_object_id
)
search depth first by object_id set order1
cycle object_id set is_cycle to 'Y' default 'N'
select * from cte
where object_id in (2,3) and is_cycle = 'N';
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
-----------------------------------------------------------------------------------------------------------------------------------------------
SQL_ID cna60m0r6uhkf, child number 0
-------------------------------------
with cte ( owner, object_name, subobject_name, object_id,
data_object_id, object_type, created, last_ddl_time, timestamp,
status, temporary, generated, secondary, namespace, edition_name,
lvl ) as ( select owner, object_name, subobject_name,
object_id, data_object_id, object_type, created, last_ddl_time,
timestamp, status, temporary, generated, secondary, namespace,
edition_name, 1 from test01 where object_id = 2
union all select /*+ use_nl(t,c) */ t.owner, t.object_name,
t.subobject_name, t.object_id, t.data_object_id, t.object_type,
t.created, t.last_ddl_time, t.timestamp, t.status, t.temporary,
t.generated, t.secondary, t.namespace, t.edition_name, c.lvl +
1 from test01 t join cte c on c.object_id = t.data_object_id )
search depth first by object_id set order1 cycle object_id set is_cycle
to 'Y' default 'N' select * from cte where object_id in (2,3) and
is_cycle = 'N'
Plan hash value: 769798160
----------------------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers | OMem | 1Mem | Used-Mem |
----------------------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 512 |00:00:01.81 | 9792 | | | |
|* 1 | VIEW | | 1 | 18E| 512 |00:00:01.81 | 9792 | | | |
| 2 | UNION ALL (RECURSIVE WITH) DEPTH FIRST| | 1 | | 4719K|00:00:01.37 | 9792 | 1612K| 624K| 1432K (0)|
| 3 | TABLE ACCESS BY INDEX ROWID BATCHED | TEST01 | 1 | 505 | 512 |00:00:00.01 | 516 | | | |
|* 4 | INDEX RANGE SCAN | IDX_TEST01_N1 | 1 | 505 | 512 |00:00:00.01 | 4 | | | |
| 5 | NESTED LOOPS | | 2 | 18E| 9216 |00:00:00.03 | 9276 | | | |
| 6 | NESTED LOOPS | | 2 | 18E| 9216 |00:00:00.01 | 60 | | | |
| 7 | RECURSIVE WITH PUMP | | 2 | | 18 |00:00:00.01 | 0 | | | |
|* 8 | INDEX RANGE SCAN | IDX_TEST01_N2 | 18 | 516 | 9216 |00:00:00.01 | 60 | | | |
| 9 | TABLE ACCESS BY INDEX ROWID | TEST01 | 9216 | 516 | 9216 |00:00:00.03 | 9216 | | | |
----------------------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter((INTERNAL_FUNCTION("OBJECT_ID") AND "IS_CYCLE"='N'))
4 - access("OBJECT_ID"=2)
8 - access("C"."OBJECT_ID"="T"."DATA_OBJECT_ID")
filter("T"."DATA_OBJECT_ID" IS NOT NULL)
43 rows selected.指定广度优先，执行计划显示为A-TIME样式
with cte (
owner, object_name, subobject_name, object_id, data_object_id,
object_type, created, last_ddl_time, timestamp, status,
temporary, generated, secondary, namespace, edition_name, lvl
) as (
select
owner, object_name, subobject_name, object_id, data_object_id,
object_type, created, last_ddl_time, timestamp, status,
temporary, generated, secondary, namespace, edition_name,
1
from test01
where object_id = 2
union all
select /*+ use_nl(t,c) */
t.owner, t.object_name, t.subobject_name, t.object_id, t.data_object_id,
t.object_type, t.created, t.last_ddl_time, t.timestamp, t.status,
t.temporary, t.generated, t.secondary, t.namespace, t.edition_name,
c.lvl + 1
from test01 t
join cte c on c.object_id = t.data_object_id
)
search breadth first by object_id set order1
cycle object_id set is_cycle to 'Y' default 'N'
select * from cte
where object_id in (2,3) and is_cycle = 'N';
SQL> select * from table(dbms_xplan.display_cursor(null,null,'ALLSTATS LAST'));
PLAN_TABLE_OUTPUT
-------------------------------------------------------------------------------------------------------------------------------------------------------------------
SQL_ID 9hbu8p6s226v4, child number 0
-------------------------------------
with cte ( owner, object_name, subobject_name, object_id,
data_object_id, object_type, created, last_ddl_time, timestamp,
status, temporary, generated, secondary, namespace, edition_name,
lvl ) as ( select owner, object_name, subobject_name,
object_id, data_object_id, object_type, created, last_ddl_time,
timestamp, status, temporary, generated, secondary, namespace,
edition_name, 1 from test01 where object_id = 2
union all select /*+ use_nl(t,c) */ t.owner, t.object_name,
t.subobject_name, t.object_id, t.data_object_id, t.object_type,
t.created, t.last_ddl_time, t.timestamp, t.status, t.temporary,
t.generated, t.secondary, t.namespace, t.edition_name, c.lvl +
1 from test01 t join cte c on c.object_id = t.data_object_id )
search breadth first by object_id set order1 cycle object_id set
is_cycle to 'Y' default 'N' select * from cte where object_id in (2,3)
and is_cycle = 'N'
Plan hash value: 3294323809
------------------------------------------------------------------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Starts | E-Rows | A-Rows | A-Time | Buffers | Reads | Writes | OMem | 1Mem | Used-Mem |
------------------------------------------------------------------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | | 512 |00:00:34.83 | 36M| 202K| 134K| | | |
|* 1 | VIEW | | 1 | 18E| 512 |00:00:34.83 | 36M| 202K| 134K| | | |
| 2 | UNION ALL (RECURSIVE WITH) BREADTH FIRST| | 1 | | 4719K|00:00:34.39 | 36M| 202K| 134K| 592M| 7993K| 97M (1)|
| 3 | TABLE ACCESS BY INDEX ROWID BATCHED | TEST01 | 1 | 505 | 512 |00:00:00.01 | 516 | 0 | 0 | | | |
|* 4 | INDEX RANGE SCAN | IDX_TEST01_N1 | 1 | 505 | 512 |00:00:00.01 | 4 | 0 | 0 | | | |
| 5 | NESTED LOOPS | | 2 | 18E| 4718K|00:00:10.06 | 4730K| 67352 | 0 | | | |
| 6 | NESTED LOOPS | | 2 | 18E| 4718K|00:00:03.15 | 11422 | 67352 | 0 | | | |
| 7 | RECURSIVE WITH PUMP | | 2 | | 4456K|00:00:00.54 | 0 | 67352 | 0 | | | |
|* 8 | INDEX RANGE SCAN | IDX_TEST01_N2 | 4456K| 516 | 4718K|00:00:01.88 | 11422 | 0 | 0 | | | |
| 9 | TABLE ACCESS BY INDEX ROWID | TEST01 | 4718K| 516 | 4718K|00:00:05.91 | 4718K| 0 | 0 | | | |
------------------------------------------------------------------------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - filter((INTERNAL_FUNCTION("OBJECT_ID") AND "IS_CYCLE"='N'))
4 - access("OBJECT_ID"=2)
8 - access("C"."OBJECT_ID"="T"."DATA_OBJECT_ID")
filter("T"."DATA_OBJECT_ID" IS NOT NULL)
43 rows selected.Oracle深度优先循环了9216次，Oracle广度优先循环了4456K次，ORACLE START WITH CONNECT BY默认用的是深度优先算法，普通的WITH递归用的是广度优先算法，现在破案了，Oracle之所以跑得快是因为默认采用深度优先算法，把需要处理的数据量给缩小了
测试的几款国产数据库目前还没实现WITH SEARCH高级递归语法，PG已经实现了，在PG18上看看性能如何
PG深度优先
postgres=# explain analyze with recursive cte as (
postgres(# select
postgres(# owner, object_name, subobject_name, object_id, data_object_id,
postgres(# object_type, created, last_ddl_time, timestamp, status,
postgres(# temporary, generated, secondary, namespace, edition_name,
postgres(# 1 as lvl
postgres(# from test01
postgres(# where object_id = 2
postgres(# union all
postgres(# select
postgres(# t.owner, t.object_name, t.subobject_name, t.object_id, t.data_object_id,
postgres(# t.object_type, t.created, t.last_ddl_time, t.timestamp, t.status,
postgres(# t.temporary, t.generated, t.secondary, t.namespace, t.edition_name,
postgres(# c.lvl + 1
postgres(# from test01 t
postgres(# join cte c on c.object_id = t.data_object_id
postgres(# )
postgres-# search depth first by object_id set ordercol
postgres-# cycle object_id set is_cycle using path
postgres-# select * from cte
postgres-# where object_id in (2, 3)
postgres-# and not is_cycle;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------
CTE Scan on cte (cost=31423106.70..31734434.41 rows=69184 width=849) (actual time=0.089..15461.487 rows=512.00 loops=1)
Filter: ((NOT is_cycle) AND (object_id = ANY ('{2,3}'::numeric[])))
Rows Removed by Filter: 4718592
Storage: Disk Maximum Storage: 1249896kB
Buffers: shared hit=22565381, temp read=156224 written=312461
CTE cte
-> Recursive Union (cost=0.56..31423106.70 rows=13836787 width=256) (actual time=0.085..12831.550 rows=4719104.00 loops=1)
Storage: Disk Maximum Storage: 1249921kB
Buffers: shared hit=22565381, temp read=156224 written=156224
-> Index Scan using idx_test01_n1 on test01 (cost=0.56..2085.60 rows=517 width=256) (actual time=0.083..0.611 rows=512.00 loops=1)
Index Cond: (object_id = '2'::numeric)
Index Searches: 1
Buffers: shared hit=517
-> Nested Loop (cost=0.56..3128265.32 rows=1383627 width=256) (actual time=3039.122..5690.770 rows=2359296.00 loops=2)
Buffers: shared hit=22564864, temp read=156224 written=1
-> WorkTable Scan on cte c (cost=0.00..103.40 rows=2585 width=100) (actual time=0.043..483.035 rows=2228480.00 loops=2)
Filter: (NOT is_cycle)
Rows Removed by Filter: 131072
Buffers: temp read=156224 written=1
-> Index Scan using idx_test01_n2 on test01 t (cost=0.56..1194.07 rows=535 width=187) (actual time=0.001..0.002 rows=1.06 loops=4456960)
Index Cond: (data_object_id = c.object_id)
Index Searches: 4456960
Buffers: shared hit=22564864
Planning Time: 0.187 ms
Execution Time: 15698.086 ms
(25 rows)
Time: 15698.654 ms (00:15.699)PG广度优先
postgres=# explain analyze with recursive cte as (
postgres(# select
postgres(# owner, object_name, subobject_name, object_id, data_object_id,
postgres(# object_type, created, last_ddl_time, timestamp, status,
postgres(# temporary, generated, secondary, namespace, edition_name,
postgres(# 1 as lvl
postgres(# from test01
postgres(# where object_id = 2
postgres(# union all
postgres(# select
postgres(# t.owner, t.object_name, t.subobject_name, t.object_id, t.data_object_id,
postgres(# t.object_type, t.created, t.last_ddl_time, t.timestamp, t.status,
postgres(# t.temporary, t.generated, t.secondary, t.namespace, t.edition_name,
postgres(# c.lvl + 1
postgres(# from test01 t
postgres(# join cte c on c.object_id = t.data_object_id
postgres(# )
postgres-# search breadth first by object_id set ordercol
postgres-# cycle object_id set is_cycle using path
postgres-# select * from cte
postgres-# where object_id in (2, 3)
postgres-# and not is_cycle;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------------------
CTE Scan on cte (cost=31423106.70..31734434.41 rows=69184 width=849) (actual time=0.037..15149.188 rows=512.00 loops=1)
Filter: ((NOT is_cycle) AND (object_id = ANY ('{2,3}'::numeric[])))
Rows Removed by Filter: 4718592
Storage: Disk Maximum Storage: 1014872kB
Buffers: shared hit=22565381, temp read=126848 written=253707
CTE cte
-> Recursive Union (cost=0.56..31423106.70 rows=13836787 width=256) (actual time=0.033..12543.080 rows=4719104.00 loops=1)
Storage: Disk Maximum Storage: 1014905kB
Buffers: shared hit=22565381, temp read=126848 written=126848
-> Index Scan using idx_test01_n1 on test01 (cost=0.56..2085.60 rows=517 width=256) (actual time=0.032..0.597 rows=512.00 loops=1)
Index Cond: (object_id = '2'::numeric)
Index Searches: 1
Buffers: shared hit=517
-> Nested Loop (cost=0.56..3128265.32 rows=1383627 width=256) (actual time=3022.640..5546.855 rows=2359296.00 loops=2)
Buffers: shared hit=22564864, temp read=126848 written=1
-> WorkTable Scan on cte c (cost=0.00..103.40 rows=2585 width=100) (actual time=0.042..452.027 rows=2228480.00 loops=2)
Filter: (NOT is_cycle)
Rows Removed by Filter: 131072
Buffers: temp read=126848 written=1
-> Index Scan using idx_test01_n2 on test01 t (cost=0.56..1194.07 rows=535 width=187) (actual time=0.001..0.002 rows=1.06 loops=4456960)
Index Cond: (data_object_id = c.object_id)
Index Searches: 4456960
Buffers: shared hit=22564864
Planning Time: 0.212 ms
Execution Time: 15348.194 ms
(25 rows)
Time: 15348.865 ms (00:15.349)PG深度优先循环了4456960次，PG广度优先也循环了4456960次，PG的深度优先没有把数据量减少
崖山CONNECT BY默认也是用的深度优先，但是性能和Oracle比相差太大，得继续优化
最后，针对树形查询，给国产数据库提点建议：
1.实现WITH SEARCH DEPTH/BREADTH FIRST语法，且真正实现深度优先，广度优先算法
2.实现Oracle的/*+ CONNECT_BY_FILTERING / 和 / + NO_CONNECT_BY_FILTERING */功能
3.START WITH CONNECT BY能控制选择走深度优先或广度优先
4.树形查询能够开启并行查询
5.树形查询能够利用批量模式
6.性能追上甚至超越Oracle