昨天看到微信群中，有人提问：pg对于标记为删除的页，是否会扫描到？
今天做了一下测试，发现如果删除的是表的最后连续的几个页(根据ctid来确定数据插入先后，只讨论有insert的情况)中的数据，最后几个页经过vacuum后，会被释放回操作系统，自然不会被扫描到，但如果不是最后的页会不会扫描到，下面测试一下。
1. 测试无索引的情况：
swrd=# d t1
Table "swrd.t1"
Column | Type | Collation | Nullable | Default
--------+-------------------+-----------+----------+---------
c1 | integer | | |
c2 | character varying | | |
c3 | character varying | | |
c4 | character varying | | |
swrd=# insert into t1 select i,md5(i::text),md5(i::text),md5(i::text) from generate_series(1,10000) t(i);
INSERT 0 10000
swrd=# select max(ctid) from t1;
max
----------
(40,80)
(1 row)
在t1中插入1w条记录，t1表是一共占用了41个页，查看记录，一个块中存放248条记录。
##全表扫描，可以看到buffers数是41个。
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 ;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------
Seq Scan on swrd.t1 (cost=0.00..141.00 rows=10000 width=103) (actual time=0.020..1.696 rows=10000 loops=1)
Output: c1, c2, c3, c4
Buffers: shared hit=41
Planning Time: 0.142 ms
Execution Time: 2.650 ms
(5 rows)
##删除几个页，看是否有变化，不做vacuum
##删除第13页到16页
swrd=# alter table t1 set(autovacuum_enabled=off);
ALTER TABLE
swrd=# delete from t1 where c1>=2977 and c1<=4216;
DELETE 1240
##仍然读取的buffer个数是41个，因为页还没做清理，所以很正常，下面做一下vacuum
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 ;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Seq Scan on swrd.t1 (cost=0.00..128.60 rows=8760 width=103) (actual time=0.066..2.961 rows=8760 loops=1)
Output: c1, c2, c3, c4
Buffers: shared read=41
Planning Time: 0.694 ms
Execution Time: 3.836 ms
(5 rows)
##做一下vacuum
swrd=# vacuum VERBOSE t1;
INFO: vacuuming "swrd.t1"
INFO: "t1": removed 0 row versions in 5 pages
INFO: "t1": found 0 removable, 8760 nonremovable row versions in 41 out of 41 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 46060313
There were 0 unused item pointers.
Skipped 0 pages due to buffer pins, 0 frozen pages.
0 pages are entirely empty.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
INFO: vacuuming "pg_toast.pg_toast_37066"
INFO: index "pg_toast_37066_index" now contains 0 row versions in 1 pages
DETAIL: 0 index row versions were removed.
0 index pages have been deleted, 0 are currently reusable.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
INFO: "pg_toast_37066": found 0 removable, 0 nonremovable row versions in 0 out of 0 pages
DETAIL: 0 dead row versions cannot be removed yet, oldest xmin: 46060313
There were 0 unused item pointers.
Skipped 0 pages due to buffer pins, 0 frozen pages.
0 pages are entirely empty.
CPU: user: 0.00 s, system: 0.00 s, elapsed: 0.00 s.
VACUUM
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 ;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Seq Scan on swrd.t1 (cost=0.00..128.60 rows=8760 width=103) (actual time=0.011..1.162 rows=8760 loops=1)
Output: c1, c2, c3, c4
Buffers: shared hit=41
Planning Time: 0.092 ms
Execution Time: 1.972 ms
(5 rows)
扫描的page仍然是41个。在做vacuum full后，扫描的页数才减少了5个。
swrd=# vacuum FULL t1;
VACUUM
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 ;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------
Seq Scan on swrd.t1 (cost=0.00..123.60 rows=8760 width=103) (actual time=0.057..2.040 rows=8760 loops=1)
Output: c1, c2, c3, c4
Buffers: shared read=36
Planning Time: 0.089 ms
Execution Time: 2.879 ms
(5 rows)
2. 测试使用索引的情况：
swrd=# create index on t1(c1);
CREATE INDEX
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 where c1>=2977 and c1<=4216 ;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------
Index Scan using t1_c1_idx on swrd.t1 (cost=0.29..42.08 rows=1240 width=103) (actual time=0.098..0.415 rows=1240 loops=1)
Output: c1, c2, c3, c4
Index Cond: ((t1.c1 >= 2977) AND (t1.c1 <= 4216))
Buffers: shared hit=7
Planning Time: 0.143 ms
Execution Time: 0.558 ms
(6 rows)
c1值在2977-4216共占用了5个页，其他两个页是来自索引的。
使用插件pageinspect看一下索引的结构：
swrd=# select * from bt_metap('t1_c1_idx');
magic | version | root | level | fastroot | fastlevel | oldest_xact | last_cleanup_num_tuples
--------+---------+------+-------+----------+-----------+-------------+-------------------------
340322 | 3 | 3 | 1 | 3 | 1 | 0 | -1
(1 row)
看到索引的level是1，也就是该索引是2级结构，包含了meta page,root page,leaf page。另外读取的buffer是root page和leaf page。
下面删掉其中的两个页的数据量：
swrd=# delete from t1 where c1 >=3225 and c1<=3720;
DELETE 496
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 where c1>=2977 and c1<=4216
swrd-# ;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------
Index Scan using t1_c1_idx on swrd.t1 (cost=0.29..42.08 rows=1240 width=103) (actual time=0.086..0.428 rows=744 loops=1)
Output: c1, c2, c3, c4
Index Cond: ((t1.c1 >= 2977) AND (t1.c1 <= 4216))
Buffers: shared hit=8 dirtied=1
Planning Time: 0.141 ms
Execution Time: 0.528 ms
(6 rows)
swrd=# explain (ANALYZE,VERBOSE,COSTS,BUFFERS,TIMING) select * from t1 where c1>=2977 and c1<=4216
;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------
Index Scan using t1_c1_idx on swrd.t1 (cost=0.29..42.08 rows=1240 width=103) (actual time=0.064..0.253 rows=744 loops=1)
Output: c1, c2, c3, c4
Index Cond: ((t1.c1 >= 2977) AND (t1.c1 <= 4216))
Buffers: shared hit=5
Planning Time: 0.125 ms
Execution Time: 0.348 ms
(6 rows)
通过使用bt_page_items查看索引中的数据，一个索引页中包含了1473条记录，并可以看到删除的页均在一个leaf page上面。
由于我们之前避免vacuum的干扰，将它关闭了，手动执行一下，可以看到对应的索引记录也被删除了。此处删除的数据是(13,1)到(14,248)的记录。
swrd=# select * from bt_page_items('t1_c1_idx',4);
itemoffset | ctid | itemlen | nulls | vars | data
------------+----------+---------+-------+------+-------------------------
1 | (17,201) | 16 | f | f | 41 11 00 00 00 00 00 00
2 | (11,217) | 16 | f | f | 81 0b 00 00 00 00 00 00
3 | (11,218) | 16 | f | f | 82 0b 00 00 00 00 00 00
4 | (11,219) | 16 | f | f | 83 0b 00 00 00 00 00 00
5 | (11,220) | 16 | f | f | 84 0b 00 00 00 00 00 00
6 | (11,221) | 16 | f | f | 85 0b 00 00 00 00 00 00
7 | (11,222) | 16 | f | f | 86 0b 00 00 00 00 00 00
8 | (11,223) | 16 | f | f | 87 0b 00 00 00 00 00 00
9 | (11,224) | 16 | f | f | 88 0b 00 00 00 00 00 00
10 | (11,225) | 16 | f | f | 89 0b 00 00 00 00 00 00
11 | (11,226) | 16 | f | f | 8a 0b 00 00 00 00 00 00
12 | (11,227) | 16 | f | f | 8b 0b 00 00 00 00 00 00
13 | (11,228) | 16 | f | f | 8c 0b 00 00 00 00 00 00
14 | (11,229) | 16 | f | f | 8d 0b 00 00 00 00 00 00
15 | (11,230) | 16 | f | f | 8e 0b 00 00 00 00 00 00
16 | (11,231) | 16 | f | f | 8f 0b 00 00 00 00 00 00
17 | (11,232) | 16 | f | f | 90 0b 00 00 00 00 00 00
18 | (11,233) | 16 | f | f | 91 0b 00 00 00 00 00 00
19 | (11,234) | 16 | f | f | 92 0b 00 00 00 00 00 00
20 | (11,235) | 16 | f | f | 93 0b 00 00 00 00 00 00
.....
277 | (12,244) | 16 | f | f | 94 0c 00 00 00 00 00 00
278 | (12,245) | 16 | f | f | 95 0c 00 00 00 00 00 00
279 | (12,246) | 16 | f | f | 96 0c 00 00 00 00 00 00
280 | (12,247) | 16 | f | f | 97 0c 00 00 00 00 00 00
281 | (12,248) | 16 | f | f | 98 0c 00 00 00 00 00 00
282 | (15,1) | 16 | f | f | 89 0e 00 00 00 00 00 00
283 | (15,2) | 16 | f | f | 8a 0e 00 00 00 00 00 00
284 | (15,3) | 16 | f | f | 8b 0e 00 00 00 00 00 00
285 | (15,4) | 16 | f | f | 8c 0e 00 00 00 00 00 00
....
769 | (16,240) | 16 | f | f | 70 10 00 00 00 00 00 00
770 | (16,241) | 16 | f | f | 71 10 00 00 00 00 00 00
771 | (16,242) | 16 | f | f | 72 10 00 00 00 00 00 00
772 | (16,243) | 16 | f | f | 73 10 00 00 00 00 00 00
773 | (16,244) | 16 | f | f | 74 10 00 00 00 00 00 00
774 | (16,245) | 16 | f | f | 75 10 00 00 00 00 00 00
775 | (16,246) | 16 | f | f | 76 10 00 00 00 00 00 00
776 | (16,247) | 16 | f | f | 77 10 00 00 00 00 00 00
777 | (16,248) | 16 | f | f | 78 10 00 00 00 00 00 00
778 | (17,1) | 16 | f | f | 79 10 00 00 00 00 00 00
....
看到上面的执行计划，第一次执行时是有8个buffer，其中一个是脏页，这里后面再深入了解一下，再补充上，第二次时，读取的buffer数就变成了5个，我们删除的两个页已经不再了。也就是使用索引扫描时没有扫描到删除的页，这个也很好理解，因为索引是根据ctid来扫描的，删除的页不存在要查询的数据了，自然就不会在扫了。
总结：
- 全表扫描时，读取数据时可能会扫描到删除的页面。（按理说在读数据时，只需要读取buftag对应的页就可以了，但测试发现膨胀产生的脏页仍会被读到。很疑惑。）
- 索引扫描时，不会扫描到删除的页面，但对于有hot操作的记录，仍需要使用原记录的line pointer。