I am using 5.7.28-31-57-log - Percona XtraDB Cluster (GPL), Release rel31 in production enviroment
one of the nodes started for no apparent reason agresively flushing bufferpool …
what I see as strange is that only file under pressure xb_doublewrite but no other file have heavy IO
EDIT:
innodb_adaptive_flushing_lwm is at 15
disabling innodb_adaptive_flushing had no effect
incereasing innodb_buffer_pool_size by several GB had no effect
If somebody met similar issue … I would be happy for any pointers where to look further
I suspect InnoDB detected idle state and decided to flush dirty pages. Can you show charts for InnoDB buffer pool dirty pages during this time?
you are right all dirty pages are gone … problem is that only xb_doublewrite is stressed, .idb data files are not stressed at all..
Bytes Filename
30.97M /xtradbcluster/xb_doublewrite
1.04M /innodblogs/ib_logfile2
17.89KB /foo-slow.log
16.00KB /tablespacefilex.ibd
8.57KB /proc/stat
8.00KB /tablespacefiley.ibd
2.51KB /proc/mounts
2.50KB /proc/vmstat
2.33KB /proc/net/netstat
2.08KB /usr/local/percona/pmm-client/textfile-collector/example.prom
^C
normaly I have folowing options:
innodb_io_capacity=1000
innodb_io_capacity_max=2000
innodb_purge_batch_size=4608
innodb_adaptive_flushing_lwm=15
so innodb IO is at innodb_io_capacity_max … when I attempted to change innodb_io_capacity to 100 nothing changed
