- 在应用负载不变的情况下可以通过如下方法进行WAL优化
登录后复制
shell
1) 延长checkpoint时间间隔
FPI产生于checkpoint之后第一次变脏的page，在下次checkpoint到来之前，已经输出过FPI的page是不在需要再次输出FPI。因此checkpoint时间间隔越长，FPI产生的频度会越低。增大checkpoint_timeout和max_wal_size可以延长checkpoint时间间隔。
2) 增加hot_update比例
普通的update经常需要更新2个数据块，并且可能还需要更新索引page，这些又都有可能产生FPI。而hot_update只修改1个数据块，需要写的WAL量也大大减少，通过调整fillfactor可以达到目标
3) 压缩
PG9.5新增了一个wal_compression 参数，设置为on可以对FPI进行压缩，消减WAL的大小，另外还可以在外部通过ssl/ssh的压缩功能减少主备间的通信流量，已经自定义归档脚本对归档的wal进行压缩- 验证
登录后复制
shell
1) 初始化pgbench
-bash-4.2$ pgbench -i
Password:
dropping old tables...
NOTICE: table "pgbench_accounts" does not exist, skipping
NOTICE: table "pgbench_branches" does not exist, skipping
NOTICE: table "pgbench_history" does not exist, skipping
NOTICE: table "pgbench_tellers" does not exist, skipping
creating tables...
generating data (client-side)...
100000 of 100000 tuples (100%) done (elapsed 0.13 s, remaining 0.00 s)
vacuuming...
creating primary keys...
done in 0.42 s (drop tables 0.00 s, create tables 0.02 s, client-side generate 0.21 s, vacuum 0.09 s, primary keys 0.09 s).
2) 验证--延长checkpoint
a. 第一次
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 1GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout = 5min
#生成数据
-bash-4.2$ psql -c "checkpoint;select pg_current_wal_lsn()"; pgbench -n -c 100 -j 100 -t 100; psql -c "select pg_current_wal_lsn()"
Password for user postgres:
CHECKPOINT
pg_current_wal_lsn
--------------------
0/A0001C0
(1 row)
Password:
pgbench (15.8)
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1
query mode: simple
number of clients: 100
number of threads: 100
maximum number of tries: 1
number of transactions per client: 100
number of transactions actually processed: 10000/10000
number of failed transactions: 0 (0.000%)
latency average = 347.779 ms
initial connection time = 273.044 ms
tps = 287.538702 (without initial connection time)
Password for user postgres:
pg_current_wal_lsn
--------------------
0/B3EC678
(1 row)
#解析WAL日志
-bash-4.2$ pg_waldump -t 1 --stats=record -s 0/A0001C0 -e 0/B3EC678
WAL statistics between 0/A0001C0 and 0/B3EC678:
Type N (%) Record size (%) FPI size (%) Combined size (%)
---- - --- ----------- --- -------- --- ------------- ---
Transaction/COMMIT 10003 ( 12.62) 340502 ( 6.44) 0 ( 0.00) 340502 ( 1.65)
Standby/RUNNING_XACTS 3 ( 0.00) 956 ( 0.02) 0 ( 0.00) 956 ( 0.00)
Standby/INVALIDATIONS 3 ( 0.00) 366 ( 0.01) 0 ( 0.00) 366 ( 0.00)
Heap2/PRUNE 3901 ( 4.92) 271240 ( 5.13) 0 ( 0.00) 271240 ( 1.31)
Heap2/VACUUM 2 ( 0.00) 118 ( 0.00) 0 ( 0.00) 118 ( 0.00)
Heap2/VISIBLE 3 ( 0.00) 192 ( 0.00) 98304 ( 0.64) 98496 ( 0.48)
Heap/INSERT 9905 ( 12.50) 786061 ( 14.86) 0 ( 0.00) 786061 ( 3.80)
Heap/UPDATE 3140 ( 3.96) 530419 ( 10.03) 7548 ( 0.05) 537967 ( 2.60)
Heap/HOT_UPDATE 26752 ( 33.76) 1932113 ( 36.53) 25988 ( 0.17) 1958101 ( 9.48)
Heap/LOCK 22033 ( 27.81) 1191797 ( 22.53) 13202280 ( 85.88) 14394077 ( 69.66)
Heap/INPLACE 8 ( 0.01) 1631 ( 0.03) 24716 ( 0.16) 26347 ( 0.13)
Heap/INSERT+INIT 101 ( 0.13) 7979 ( 0.15) 0 ( 0.00) 7979 ( 0.04)
Heap/UPDATE+INIT 115 ( 0.15) 18374 ( 0.35) 0 ( 0.00) 18374 ( 0.09)
Btree/INSERT_LEAF 3261 ( 4.12) 207923 ( 3.93) 2014420 ( 13.10) 2222343 ( 10.76)
Btree/VACUUM 2 ( 0.00) 122 ( 0.00) 0 ( 0.00) 122 ( 0.00)
-------- -------- -------- --------
Total 79232 5289793 [25.60%] 15373256 [74.40%] 20663049 [100%]
b. 第二次(调整参数后)
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
#生产数据
-bash-4.2$ psql -c "checkpoint;select pg_current_wal_lsn()"; pgbench -n -c 100 -j 100 -t 100; psql -c "select pg_current_wal_lsn()"
Password for user postgres:
CHECKPOINT
pg_current_wal_lsn
--------------------
0/C000188
(1 row)
Password:
pgbench (15.8)
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1
query mode: simple
number of clients: 100
number of threads: 100
maximum number of tries: 1
number of transactions per client: 100
number of transactions actually processed: 10000/10000
number of failed transactions: 0 (0.000%)
latency average = 341.077 ms
initial connection time = 165.900 ms
tps = 293.189074 (without initial connection time)
Password for user postgres:
pg_current_wal_lsn
--------------------
0/D3AA588
(1 row)
#解析WAL数据
-bash-4.2$ pg_waldump -t 1 --stats=record -s 0/C000188 -e 0/D3AA588
WAL statistics between 0/C000188 and 0/D3AA588:
Type N (%) Record size (%) FPI size (%) Combined size (%)
---- - --- ----------- --- -------- --- ------------- ---
Transaction/COMMIT 10003 ( 13.52) 340374 ( 7.26) 0 ( 0.00) 340374 ( 1.67)
Standby/RUNNING_XACTS 4 ( 0.01) 1269 ( 0.03) 0 ( 0.00) 1269 ( 0.01)
Standby/INVALIDATIONS 2 ( 0.00) 180 ( 0.00) 0 ( 0.00) 180 ( 0.00)
Heap2/PRUNE 4808 ( 6.50) 332597 ( 7.09) 12696384 ( 80.80) 13028981 ( 63.86)
Heap/INSERT 10000 ( 13.52) 787500 ( 16.79) 494892 ( 3.15) 1282392 ( 6.29)
Heap/UPDATE 267 ( 0.36) 41002 ( 0.87) 3720 ( 0.02) 44722 ( 0.22)
Heap/HOT_UPDATE 29733 ( 40.19) 2148707 ( 45.82) 607136 ( 3.86) 2755843 ( 13.51)
Heap/LOCK 18869 ( 25.51) 1018926 ( 21.73) 0 ( 0.00) 1018926 ( 4.99)
Heap/INPLACE 3 ( 0.00) 601 ( 0.01) 24716 ( 0.16) 25317 ( 0.12)
Heap/UPDATE+INIT 11 ( 0.01) 934 ( 0.02) 0 ( 0.00) 934 ( 0.00)
Btree/INSERT_LEAF 278 ( 0.38) 17088 ( 0.36) 1886160 ( 12.00) 1903248 ( 9.33)
-------- -------- -------- --------
Total 73978 4689178 [22.98%] 15713008 [77.02%] 20402186 [100%] #比20663049，WAL减少了
c. 第三次(不做checkpoint)
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
#生产数据
-bash-4.2$ psql -c "select pg_current_wal_lsn()"; pgbench -n -c 100 -j 100 -t 100; psql -c "select pg_current_wal_lsn()"
Password for user postgres:
pg_current_wal_lsn
--------------------
0/D3CB198
(1 row)
Password:
pgbench (15.8)
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1
query mode: simple
number of clients: 100
number of threads: 100
maximum number of tries: 1
number of transactions per client: 100
number of transactions actually processed: 10000/10000
number of failed transactions: 0 (0.000%)
latency average = 218.119 ms
initial connection time = 145.239 ms
tps = 458.464513 (without initial connection time)
Password for user postgres:
pg_current_wal_lsn
--------------------
0/D8AA9E0
(1 row)
#解析WAL日志
-bash-4.2$ pg_waldump -t 1 --stats=record -s 0/D3CB198 -e 0/D8AA9E0
WAL statistics between 0/D3CB198 and 0/D8AA9E0:
Type N (%) Record size (%) FPI size (%) Combined size (%)
---- - --- ----------- --- -------- --- ------------- ---
Transaction/COMMIT 10005 ( 13.46) 340714 ( 7.23) 0 ( 0.00) 340714 ( 6.90)
Standby/RUNNING_XACTS 2 ( 0.00) 655 ( 0.01) 0 ( 0.00) 655 ( 0.01)
Standby/INVALIDATIONS 2 ( 0.00) 228 ( 0.00) 0 ( 0.00) 228 ( 0.00)
Heap2/PRUNE 4639 ( 6.24) 319782 ( 6.78) 0 ( 0.00) 319782 ( 6.47)
Heap2/VACUUM 8 ( 0.01) 420 ( 0.01) 0 ( 0.00) 420 ( 0.01)
Heap2/VISIBLE 5 ( 0.01) 295 ( 0.01) 0 ( 0.00) 295 ( 0.01)
Heap/INSERT 10000 ( 13.46) 790000 ( 16.76) 0 ( 0.00) 790000 ( 15.99)
Heap/UPDATE 266 ( 0.36) 44061 ( 0.93) 21252 ( 9.44) 65313 ( 1.32)
Heap/HOT_UPDATE 29735 ( 40.01) 2152030 ( 45.65) 21000 ( 9.33) 2173030 ( 44.00)
Heap/LOCK 19341 ( 26.03) 1044414 ( 22.16) 0 ( 0.00) 1044414 ( 21.15)
Heap/INPLACE 7 ( 0.01) 1664 ( 0.04) 0 ( 0.00) 1664 ( 0.03)
Heap/UPDATE+INIT 16 ( 0.02) 1530 ( 0.03) 0 ( 0.00) 1530 ( 0.03)
Btree/INSERT_LEAF 282 ( 0.38) 17982 ( 0.38) 182900 ( 81.23) 200882 ( 4.07)
Btree/VACUUM 2 ( 0.00) 136 ( 0.00) 0 ( 0.00) 136 ( 0.00)
-------- -------- -------- --------
Total 74310 4713911 [95.44%] 225152 [4.56%] 4939063 [100%]
3) 验证--增加hot_update比例
a. 第一次(fillfactor为100)
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
#生产数据
-bash-4.2$ psql -c "checkpoint;select pg_current_wal_lsn()"; pgbench -n -c 100 -j 100 -t 100; psql -c "select pg_current_wal_lsn()"
Password for user postgres:
CHECKPOINT
pg_current_wal_lsn
--------------------
0/D8AFBB8
(1 row)
Password:
pgbench (15.8)
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1
query mode: simple
number of clients: 100
number of threads: 100
maximum number of tries: 1
number of transactions per client: 100
number of transactions actually processed: 10000/10000
number of failed transactions: 0 (0.000%)
latency average = 400.769 ms
initial connection time = 188.981 ms
tps = 249.520098 (without initial connection time)
Password for user postgres:
pg_current_wal_lsn
--------------------
0/EDBD258
(1 row)
#解析WAL数据
-bash-4.2$ pg_waldump -t 1 --stats=record -s 0/D8AFBB8 -e 0/EDBD258
WAL statistics between 0/D8AFBB8 and 0/EDBD258:
Type N (%) Record size (%) FPI size (%) Combined size (%)
---- - --- ----------- --- -------- --- ------------- ---
Transaction/COMMIT 10005 ( 13.47) 340714 ( 7.21) 0 ( 0.00) 340714 ( 1.56)
Storage/TRUNCATE 1 ( 0.00) 46 ( 0.00) 0 ( 0.00) 46 ( 0.00)
Standby/LOCK 1 ( 0.00) 42 ( 0.00) 0 ( 0.00) 42 ( 0.00)
Standby/RUNNING_XACTS 3 ( 0.00) 1263 ( 0.03) 0 ( 0.00) 1263 ( 0.01)
Standby/INVALIDATIONS 2 ( 0.00) 180 ( 0.00) 0 ( 0.00) 180 ( 0.00)
Heap2/PRUNE 4408 ( 5.94) 319323 ( 6.75) 12892876 ( 75.27) 13212199 ( 60.45)
Heap2/VACUUM 8 ( 0.01) 430 ( 0.01) 0 ( 0.00) 430 ( 0.00)
Heap2/VISIBLE 113 ( 0.15) 6682 ( 0.14) 98304 ( 0.57) 104986 ( 0.48)
Heap/INSERT 10000 ( 13.47) 787500 ( 16.66) 1534892 ( 8.96) 2322392 ( 10.63)
Heap/UPDATE 345 ( 0.46) 57061 ( 1.21) 0 ( 0.00) 57061 ( 0.26)
Heap/HOT_UPDATE 29644 ( 39.92) 2142423 ( 45.31) 591736 ( 3.45) 2734159 ( 12.51)
Heap/LOCK 19322 ( 26.02) 1043388 ( 22.07) 0 ( 0.00) 1043388 ( 4.77)
Heap/INPLACE 6 ( 0.01) 1210 ( 0.03) 24716 ( 0.14) 25926 ( 0.12)
Heap/UPDATE+INIT 28 ( 0.04) 4440 ( 0.09) 0 ( 0.00) 4440 ( 0.02)
Btree/INSERT_LEAF 373 ( 0.50) 23135 ( 0.49) 1986480 ( 11.60) 2009615 ( 9.19)
Btree/VACUUM 2 ( 0.00) 146 ( 0.00) 0 ( 0.00) 146 ( 0.00)
-------- -------- -------- --------
Total 74261 4727983 [21.63%] 17129004 [78.37%] 21856987 [100%]
b. 第二次(fillfactor调整为)
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
#调整表的fillfactor
alter table pgbench_accounts set (fillfactor=60);
alter table pgbench_branches set (fillfactor=60);
alter table pgbench_history set (fillfactor=60);
alter table pgbench_tellers set (fillfactor=60);
vacuum full pgbench_accounts;
vacuum full pgbench_branches;
vacuum full pgbench_history ;
vacuum full pgbench_tellers ;
#生成数据
-bash-4.2$ psql -c "checkpoint;select pg_current_wal_lsn()"; pgbench -n -c 100 -j 100 -t 100; psql -c "select pg_current_wal_lsn()"
Password for user postgres:
CHECKPOINT
pg_current_wal_lsn
--------------------
0/FE98040
(1 row)
Password:
pgbench (15.8)
transaction type: <builtin: TPC-B (sort of)>
scaling factor: 1
query mode: simple
number of clients: 100
number of threads: 100
maximum number of tries: 1
number of transactions per client: 100
number of transactions actually processed: 10000/10000
number of failed transactions: 0 (0.000%)
latency average = 413.325 ms
initial connection time = 234.007 ms
tps = 241.940584 (without initial connection time)
Password for user postgres:
pg_current_wal_lsn
--------------------
0/11001FE0
(1 row)
#解析WAL日志
-bash-4.2$ pg_waldump -t 1 --stats=record -s 0/FE98040 -e 0/11001FE0
WAL statistics between 0/FE98040 and 0/11001FE0:
Type N (%) Record size (%) FPI size (%) Combined size (%)
---- - --- ----------- --- -------- --- ------------- ---
Transaction/COMMIT 10001 ( 13.69) 340309 ( 7.38) 0 ( 0.00) 340309 ( 1.88)
Standby/RUNNING_XACTS 3 ( 0.00) 924 ( 0.02) 0 ( 0.00) 924 ( 0.01)
Heap2/PRUNE 4064 ( 5.56) 286153 ( 6.21) 30892 ( 0.23) 317045 ( 1.76)
Heap/INSERT 9900 ( 13.56) 782075 ( 16.97) 2052 ( 0.02) 784127 ( 4.34)
Heap/UPDATE 54 ( 0.07) 6566 ( 0.14) 19576 ( 0.15) 26142 ( 0.14)
Heap/HOT_UPDATE 29934 ( 40.99) 2158766 ( 46.84) 13306060 ( 98.96) 15464826 ( 85.65)
Heap/LOCK 18883 ( 25.86) 1019687 ( 22.12) 32624 ( 0.24) 1052311 ( 5.83)
Heap/INPLACE 3 ( 0.00) 294 ( 0.01) 46204 ( 0.34) 46498 ( 0.26)
Heap/INSERT+INIT 100 ( 0.14) 7900 ( 0.17) 0 ( 0.00) 7900 ( 0.04)
Heap/UPDATE+INIT 19 ( 0.03) 1598 ( 0.03) 0 ( 0.00) 1598 ( 0.01)
Btree/INSERT_LEAF 73 ( 0.10) 4639 ( 0.10) 9160 ( 0.07) 13799 ( 0.08)
-------- -------- -------- --------
Total 73034 4608911 [25.53%] 13446568 [74.47%] 18055479 [100%] #比21856987减少了
4) 验证--压缩
a. 第一次
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
#生成数据
同上，省略。。。
#解析WAL日志
同上，省略。。。
b. 第二次(调整压缩参数)
#配置参数
max_connections = 3000
shared_buffers = 2048MB
max_wal_size = 2GB
min_wal_size = 100MB
wal_keep_size=5000
checkpoint_timeout=30min
wal_compression=on
#生成数据
同上，省略。。。
#解析WAL日志
同上，省略。。。- 优化的副作用
登录后复制
shell
1) 延长checkpoint时间间隔的副作用
导致crash恢复时间边长，crash恢复时需要回放的WAL日志量一般小于max_wal_size的一半，WAL的回放速度(wal_compression=on时)一般是50MB/s~150MB/s之间，可以根据可容忍的最大crash恢复时间，估算出允许的max_wal_size的最大值
2) 调整fillfactor的副作用
过小的设置会浪费存储空间，另外对于频繁更新的表，即使把fillfactor设成100%，每个page里还是要一部分空间被dead tuple占据，不会比设置成一个合适的稍小的fillfactor更节省空间。
3) 设置wal_compression=on的副作用
需要额外占用CPU的资源进行叶索，但影响不大