fix

Author: digoal

201611/20161121_01.md

@@ -404,37 +404,37 @@ unix_socket_directories = '.'
 tcp_keepalives_idle = 60
 tcp_keepalives_interval = 10
 tcp_keepalives_count = 10
-shared_buffers = 128GB
-maintenance_work_mem = 4GB
+shared_buffers = 128GB                      # 1/4 主机内存
+maintenance_work_mem = 2GB                  # min( 2G, (1/4 主机内存)/autovacuum_max_workers )
 dynamic_shared_memory_type = posix
 vacuum_cost_delay = 0
 bgwriter_delay = 10ms
 bgwriter_lru_maxpages = 1000
 bgwriter_lru_multiplier = 10.0
-bgwriter_flush_after = 0  # IO很好的机器，不需要考虑平滑调度
+bgwriter_flush_after = 0                    # IO很好的机器，不需要考虑平滑调度
 max_worker_processes = 128
-max_parallel_workers_per_gather = 0   #  如果需要使用并行查询，设置为大于1
+max_parallel_workers_per_gather = 0         #  如果需要使用并行查询，设置为大于1 ，不建议超过 主机cores-2
 old_snapshot_threshold = -1
 backend_flush_after = 0  # IO很好的机器，不需要考虑平滑调度, 否则建议128~256kB
 wal_level = replica
 synchronous_commit = off
-full_page_writes = on
-wal_buffers = 1GB
+full_page_writes = on   # 支持原子写超过BLOCK_SIZE的块设备，在对齐后可以关闭。或者支持cow的文件系统可以关闭。
+wal_buffers = 1GB       # min( 2047MB, shared_buffers/32 ) = 512MB
 wal_writer_delay = 10ms
 wal_writer_flush_after = 0  # IO很好的机器，不需要考虑平滑调度, 否则建议128~256kB
-checkpoint_timeout = 30min  # 不建议频繁做检查点，否则XLOG会产生很多的FULL PAGE WRITE。
-max_wal_size = 256GB  # 建议是SHARED BUFFER的2倍
-min_wal_size = 64GB
-checkpoint_completion_target = 0.05  # 硬盘好的情况下，可以让检查点快速结束，恢复时也可以快速达到一致状态。否则建议0.5~0.9
-checkpoint_flush_after = 0  # IO很好的机器，不需要考虑平滑调度, 否则建议128~256kB
+checkpoint_timeout = 30min  # 不建议频繁做检查点，否则XLOG会产生很多的FULL PAGE WRITE(when full_page_writes=on)。
+max_wal_size = 256GB       # 建议是SHARED BUFFER的2倍
+min_wal_size = 64GB        # max_wal_size/4
+checkpoint_completion_target = 0.05          # 硬盘好的情况下，可以让检查点快速结束，恢复时也可以快速达到一致状态。否则建议0.5~0.9
+checkpoint_flush_after = 0                   # IO很好的机器，不需要考虑平滑调度, 否则建议128~256kB
 archive_mode = on
 archive_command = '/bin/date'      #  后期再修改，如  'test ! -f /disk1/digoal/arch/%f && cp %p /disk1/digoal/arch/%f'
 max_wal_senders = 8
 random_page_cost = 1.3  # IO很好的机器，不需要考虑离散和顺序扫描的成本差异
 parallel_tuple_cost = 0
 parallel_setup_cost = 0
 min_parallel_relation_size = 0
-effective_cache_size = 300GB  # 看着办，剩下的都是OS可用的CACHE。
+effective_cache_size = 300GB                          # 看着办，扣掉会话连接RSS，shared buffer, autovacuum worker, 剩下的都是OS可用的CACHE。
 force_parallel_mode = off
 log_destination = 'csvlog'
 logging_collector = on
@@ -445,13 +445,13 @@ log_disconnections = on
 log_error_verbosity = verbose
 log_timezone = 'PRC'
 vacuum_defer_cleanup_age = 0
-hot_standby_feedback = off    # 建议关闭，以免备库长事务导致 主库无法回收垃圾而膨胀。
+hot_standby_feedback = off                             # 建议关闭，以免备库长事务导致 主库无法回收垃圾而膨胀。
 max_standby_archive_delay = 300s
 max_standby_streaming_delay = 300s
 autovacuum = on
 log_autovacuum_min_duration = 0
-autovacuum_max_workers = 16  # CPU核多，并且IO好的情况下，可多点，但是注意16*autovacuum mem，会消耗较多内存，所以内存也要有基础。  
-autovacuum_naptime = 45s     # 建议不要太高频率，否则会因为vacuum产生较多的XLOG。
+autovacuum_max_workers = 16                            # CPU核多，并且IO好的情况下，可多点，但是注意16*autovacuum mem，会消耗较多内存，所以内存也要有基础。  
+autovacuum_naptime = 45s                               # 建议不要太高频率，否则会因为vacuum产生较多的XLOG。
 autovacuum_vacuum_scale_factor = 0.1
 autovacuum_analyze_scale_factor = 0.1
 autovacuum_freeze_max_age = 1600000000

201710/20171011_01.md

@@ -131,15 +131,16 @@ fio的测试规格。
 ```
 # 规格相关(8C, 64G为例)
 shared_buffers = 16GB                  # 1/4 主机内存 = 16G
-work_mem = 64MB                        # 1/4 主机内存 / 256 = 16G/256 = 64MB                         
-wal_buffers = 512MB                    # min(2047MB, shared_buffers/32) = 512MB
+work_mem = 64MB                        # 1/4 主机内存 / 256 = 16G/256 = 64MB (假设256个并发同时使用work_mem)                        
+wal_buffers = 512MB                    # min( 2047MB, shared_buffers/32 ) = 512MB
 maintenance_work_mem = 2GB             # min( 2G, (1/4 主机内存)/autovacuum_max_workers ) = min(2G, 16/5) = 2GB
 max_wal_size = 32GB                    # shared_buffers*2 = 16*2 = 32GB
 min_wal_size = 8GB                     # max_wal_size/4 = 32/4 = 8GB
 effective_cache_size = 40GB            # 5/8 主机内存 = 40GB
 max_parallel_workers_per_gather = 6    # 主机cores-2 = 8-2 = 6
 
 # 优化器相关
+# https://github.com/digoal/blog/blob/master/201311/20131126_03.md
 seq_page_cost = 1.0
 random_page_cost = 1.2
 cpu_tuple_cost = 0.01
@@ -168,7 +169,7 @@ max_parallel_workers = 128
 wal_level = replica
 fsync = on
 synchronous_commit = off
-full_page_writes = on
+full_page_writes = on     # 支持原子写超过BLOCK_SIZE的块设备，在对齐后可以关闭。或者支持cow的文件系统可以关闭。
 wal_writer_delay = 10ms
 wal_writer_flush_after = 1MB
 checkpoint_timeout = 35min
@@ -204,9 +205,9 @@ autovacuum_multixact_freeze_max_age = 1200000000
 autovacuum_vacuum_cost_delay = 0
 statement_timeout = 0
 vacuum_freeze_min_age = 50000000
-vacuum_freeze_table_age = 500000000
+vacuum_freeze_table_age = 800000000
 vacuum_multixact_freeze_min_age = 50000000
-vacuum_multixact_freeze_table_age = 500000000
+vacuum_multixact_freeze_table_age = 800000000
 datestyle = 'iso, ymd'
 timezone = 'PRC'
 lc_messages = 'en_US.UTF8'

201710/20171018_01.md

@@ -544,6 +544,8 @@ tcp_keepalives_interval = 60
 tcp_keepalives_count = 10  
 shared_buffers = 256MB          # 共享内存，建议设置为系统内存的1/4  .  
 maintenance_work_mem = 64MB     # 系统内存超过32G时，建议设置为1GB。超过64GB时，建议设置为2GB。超过128GB时，建议设置为4GB。  
+work_mem = 64MB                        # 1/4 主机内存 / 256 (假设256个并发同时使用work_mem)
+wal_buffers = 512MB                    # min( 2047MB, shared_buffers/32 ) 
 dynamic_shared_memory_type = posix  
 vacuum_cost_delay = 0  
 bgwriter_delay = 10ms  
@@ -552,16 +554,16 @@ bgwriter_lru_multiplier = 5.0
 effective_io_concurrency = 0  
 max_worker_processes = 128                 
 max_parallel_workers_per_gather = 32        # 建议设置为主机CPU核数的一半。  
-max_parallel_workers = 32                   # 建议设置为主机CPU核数的一半。  
+max_parallel_workers = 32                   # 看业务AP和TP的比例，以及AP TP时间交错分配。实际情况调整。例如 主机CPU cores-2
 wal_level = replica  
 fsync = on  
 synchronous_commit = off  
-full_page_writes = on  
+full_page_writes = on                  # 支持原子写超过BLOCK_SIZE的块设备，在对齐后可以关闭。或者支持cow的文件系统可以关闭。
 wal_writer_delay = 10ms  
 wal_writer_flush_after = 1MB  
 checkpoint_timeout = 35min  
-max_wal_size = 8GB                          # 建议设置为max(8GB, shared_buffers*2)  
-min_wal_size = 80MB  
+max_wal_size = 32GB                    # shared_buffers*2 
+min_wal_size = 8GB                     # max_wal_size/4 
 archive_mode = on  
 archive_command = '/bin/date'  
 max_wal_senders = 10  
@@ -574,7 +576,7 @@ parallel_tuple_cost = 0.1
 parallel_setup_cost = 1000.0  
 min_parallel_table_scan_size = 8MB  
 min_parallel_index_scan_size = 512kB  
-effective_cache_size = 10GB                 # 建议设置为主机内存的3/4。     
+effective_cache_size = 10GB                 # 建议设置为主机内存的5/8。     
 log_destination = 'csvlog'  
 logging_collector = on  
 log_directory = 'log'  
@@ -602,9 +604,9 @@ statement_timeout = 0                                # 单位ms, s, min, h, d.
 lock_timeout = 0                                     # 单位ms, s, min, h, d.  表示锁等待的超时时间，0表示不限制。  
 idle_in_transaction_session_timeout = 2h             # 单位ms, s, min, h, d.  表示空闲事务的超时时间，0表示不限制。  
 vacuum_freeze_min_age = 50000000  
-vacuum_freeze_table_age = 500000000  
-vacuum_multixact_freeze_min_age = 5000000  
-vacuum_multixact_freeze_table_age = 500000000  
+vacuum_freeze_table_age = 800000000  
+vacuum_multixact_freeze_min_age = 50000000  
+vacuum_multixact_freeze_table_age = 800000000  
 datestyle = 'iso, ymd'  
 timezone = 'PRC'  
 lc_messages = 'en_US.UTF8'