欢迎您访问程序员文章站本站旨在为大家提供分享程序员计算机编程知识!
您现在的位置是: 首页

TiDB 中的TiKV 性能参数调优实战

程序员文章站 2022-07-13 10:49:22
...

概述:

TiKV 最底层使用的是 RocksDB 做为持久化存储,所以 TiKV 的很多性能相关的参数都是与 RocksDB 相关的。TiKV 使用了两个 RocksDB 实例,默认 RocksDB 实例存储 KV 数据,Raft RocksDB 实例(简称 RaftDB)存储 Raft 数据。

TiKV 使用了 RocksDB 的 Column Families 特性。

  • 默认 RocksDB 实例将 KV 数据存储在内部的 defaultwrite 和 lock 3 个 CF 内。

    • default CF 存储的是真正的数据,与其对应的参数位于 [rocksdb.defaultcf] 项中;
    • write CF 存储的是数据的版本信息(MVCC)以及索引相关的数据,相关的参数位于 [rocksdb.writecf] 项中;
    • lock CF 存储的是锁信息,系统使用默认参数。
  • Raft RocksDB 实例存储 Raft log。

    • default CF 主要存储的是 Raft log,与其对应的参数位于 [raftdb.defaultcf] 项中。

每个 CF 都有单独的 block-cache,用于缓存数据块,加速 RocksDB 的读取速度,block-cache 的大小通过参数 block-cache-size 控制,block-cache-size 越大,能够缓存的热点数据越多,对读取操作越有利,同时占用的系统内存也会越多。

每个 CF 有各自的 write-buffer,大小通过 write-buffer-size 控制。

tikv的版本:
# ./tikv-server --version
TiKV 
Release Version:   2.0.6
Git Commit Hash:   57c83dc4ebc93d38d77dc8f7d66db224760766cc
Git Commit Branch: release-2.0
UTC Build Time:    2018-08-03 11:28:38
Rust Version:      1.27.0-nightly (48fa6f963 2018-04-05)

tikv的分组:
#grep '^[[]' last_tikv.toml 
[readpool.storage]
[readpool.coprocessor]
[server]
[server.labels]
[storage]
[pd]
[metric]
[raftstore]
[coprocessor]
[rocksdb]
[rocksdb.defaultcf]
[rocksdb.writecf]
[rocksdb.lockcf]
[rocksdb.raftcf]
[raftdb]
[raftdb.defaultcf]
[security]
[import]
注释:
查看文件 不取含[]的行:
#grep '^[^[]' last_tikv.toml

在tikv2.0版本默认的参数多达300+个参数:

log-level = "info"
log-file = "/data/deploy/log/tikv.log"
[readpool.storage]
high-concurrency = 4
normal-concurrency = 4
low-concurrency = 4
max-tasks-high = 8000
max-tasks-normal = 8000
max-tasks-low = 8000
stack-size = "10MB"

[readpool.coprocessor]
high-concurrency = 38
normal-concurrency = 38
low-concurrency = 38
max-tasks-high = 76000
max-tasks-normal = 76000
max-tasks-low = 76000
stack-size = "10MB"

[server]
addr = "0.0.0.0:20160"
advertise-addr = "10.19.75.102:20160"
notify-capacity = 40960
messages-per-tick = 4096
grpc-compression-type = "none"
grpc-concurrency = 4
grpc-concurrent-stream = 1024
grpc-raft-conn-num = 10
grpc-stream-initial-window-size = "2MB"
grpc-keepalive-time = "10s"
grpc-keepalive-timeout = "3s"
concurrent-send-snap-limit = 32
concurrent-recv-snap-limit = 32
end-point-max-tasks = 2000
end-point-recursion-limit = 1000
end-point-stream-channel-size = 8
end-point-batch-row-limit = 64
end-point-stream-batch-row-limit = 128
end-point-request-max-handle-duration = "1m"
snap-max-write-bytes-per-sec = "100MB"
snap-max-total-size = "0KB"

[server.labels]

[storage]
data-dir = "/data/deploy/data"
gc-ratio-threshold = 1.1
max-key-size = 4096
scheduler-notify-capacity = 10240
scheduler-messages-per-tick = 1024
scheduler-concurrency = 2048000
scheduler-worker-pool-size = 8
scheduler-pending-write-threshold = "100MB"

[pd]
endpoints = ["10.19.85.149:2379", "10.19.15.103:2379", "10.19.189.221:2379"]

[metric]
interval = "15s"
address = "10.19.85.149:9091"
job = "tikv"

[raftstore]
sync-log = true
raftdb-path = ""
capacity = "0KB"
raft-base-tick-interval = "1s"
raft-heartbeat-ticks = 2
raft-election-timeout-ticks = 10
raft-min-election-timeout-ticks = 0
raft-max-election-timeout-ticks = 0
raft-max-size-per-msg = "1MB"
raft-max-inflight-msgs = 256
raft-entry-max-size = "8MB"
raft-log-gc-tick-interval = "10s"
raft-log-gc-threshold = 50
raft-log-gc-count-limit = 73728
raft-log-gc-size-limit = "72MB"
split-region-check-tick-interval = "10s"
region-split-check-diff = "6MB"
region-compact-check-interval = "5m"
clean-stale-peer-delay = "10m"
region-compact-check-step = 100
region-compact-min-tombstones = 10000
pd-heartbeat-tick-interval = "1m"
pd-store-heartbeat-tick-interval = "10s"
snap-mgr-gc-tick-interval = "1m"
snap-gc-timeout = "4h"
lock-cf-compact-interval = "10m"
lock-cf-compact-bytes-threshold = "256MB"
notify-capacity = 40960
messages-per-tick = 4096
max-peer-down-duration = "5m"
max-leader-missing-duration = "2h"
abnormal-leader-missing-duration = "10m"
peer-stale-state-check-interval = "5m"
snap-apply-batch-size = "10MB"
consistency-check-interval = "0s"
report-region-flow-interval = "1m"
raft-store-max-leader-lease = "9s"
right-derive-when-split = true
allow-remove-leader = false
merge-max-log-gap = 10
merge-check-tick-interval = "10s"
use-delete-range = false
cleanup-import-sst-interval = "10m"

[coprocessor]
split-region-on-table = true
region-max-size = "144MB"
region-split-size = "96MB"

[rocksdb]
wal-recovery-mode = 2
wal-dir = ""
wal-ttl-seconds = 0
wal-size-limit = "0KB"
max-total-wal-size = "4GB"
max-background-jobs = 6
max-manifest-file-size = "20MB"
create-if-missing = true
max-open-files = 40960
enable-statistics = true
stats-dump-period = "10m"
compaction-readahead-size = "0KB"
info-log-max-size = "1GB"
info-log-roll-time = "0s"
info-log-keep-log-file-num = 10
info-log-dir = ""
rate-bytes-per-sec = "0KB"
bytes-per-sync = "1MB"
wal-bytes-per-sync = "512KB"
max-sub-compactions = 1
writable-file-max-buffer-size = "1MB"
use-direct-io-for-flush-and-compaction = false
enable-pipelined-write = true

[rocksdb.defaultcf]
block-size = "64KB"
block-cache-size = "48331MB"
disable-block-cache = false
cache-index-and-filter-blocks = true
pin-l0-filter-and-index-blocks = true
use-bloom-filter = true
whole-key-filtering = true
bloom-filter-bits-per-key = 10
block-based-bloom-filter = false
read-amp-bytes-per-bit = 0
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1
max-bytes-for-level-base = "512MB"
target-file-size-base = "8MB"
level0-file-num-compaction-trigger = 4
level0-slowdown-writes-trigger = 20
level0-stop-writes-trigger = 36
max-compaction-bytes = "2GB"
compaction-pri = 3
dynamic-level-bytes = false
num-levels = 7
max-bytes-for-level-multiplier = 10
compaction-style = 0
disable-auto-compactions = false
soft-pending-compaction-bytes-limit = "64GB"
hard-pending-compaction-bytes-limit = "256GB"

[rocksdb.writecf]
block-size = "64KB"
block-cache-size = "28998MB"
disable-block-cache = false
cache-index-and-filter-blocks = true
pin-l0-filter-and-index-blocks = true
use-bloom-filter = true
whole-key-filtering = false
bloom-filter-bits-per-key = 10
block-based-bloom-filter = false
read-amp-bytes-per-bit = 0
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1
max-bytes-for-level-base = "512MB"
target-file-size-base = "8MB"
level0-file-num-compaction-trigger = 4
level0-slowdown-writes-trigger = 20
level0-stop-writes-trigger = 36
max-compaction-bytes = "2GB"
compaction-pri = 3
dynamic-level-bytes = false
num-levels = 7
max-bytes-for-level-multiplier = 10
compaction-style = 0
disable-auto-compactions = false
soft-pending-compaction-bytes-limit = "64GB"
hard-pending-compaction-bytes-limit = "256GB"

[rocksdb.lockcf]
block-size = "16KB"
block-cache-size = "1GB"
disable-block-cache = false
cache-index-and-filter-blocks = true
pin-l0-filter-and-index-blocks = true
use-bloom-filter = true
whole-key-filtering = true
bloom-filter-bits-per-key = 10
block-based-bloom-filter = false
read-amp-bytes-per-bit = 0
compression-per-level = ["no", "no", "no", "no", "no", "no", "no"]
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1
max-bytes-for-level-base = "128MB"
target-file-size-base = "8MB"
level0-file-num-compaction-trigger = 1
level0-slowdown-writes-trigger = 20
level0-stop-writes-trigger = 36
max-compaction-bytes = "2GB"
compaction-pri = 0
dynamic-level-bytes = false
num-levels = 7
max-bytes-for-level-multiplier = 10
compaction-style = 0
disable-auto-compactions = false
soft-pending-compaction-bytes-limit = "64GB"
hard-pending-compaction-bytes-limit = "256GB"

[rocksdb.raftcf]
block-size = "16KB"
block-cache-size = "128MB"
disable-block-cache = false
cache-index-and-filter-blocks = true
pin-l0-filter-and-index-blocks = true
use-bloom-filter = true
whole-key-filtering = true
bloom-filter-bits-per-key = 10
block-based-bloom-filter = false
read-amp-bytes-per-bit = 0
compression-per-level = ["no", "no", "no", "no", "no", "no", "no"]
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1
max-bytes-for-level-base = "128MB"
target-file-size-base = "8MB"
level0-file-num-compaction-trigger = 1
level0-slowdown-writes-trigger = 20
level0-stop-writes-trigger = 36
max-compaction-bytes = "2GB"
compaction-pri = 0
dynamic-level-bytes = false
num-levels = 7
max-bytes-for-level-multiplier = 10
compaction-style = 0
disable-auto-compactions = false
soft-pending-compaction-bytes-limit = "64GB"
hard-pending-compaction-bytes-limit = "256GB"

[raftdb]
wal-recovery-mode = 2
wal-dir = ""
wal-ttl-seconds = 0
wal-size-limit = "0KB"
max-total-wal-size = "4GB"
max-manifest-file-size = "20MB"
create-if-missing = true
max-open-files = 40960
enable-statistics = true
stats-dump-period = "10m"
compaction-readahead-size = "0KB"
info-log-max-size = "1GB"
info-log-roll-time = "0s"
info-log-keep-log-file-num = 10
info-log-dir = ""
max-sub-compactions = 1
writable-file-max-buffer-size = "1MB"
use-direct-io-for-flush-and-compaction = false
enable-pipelined-write = true
allow-concurrent-memtable-write = false
bytes-per-sync = "1MB"
wal-bytes-per-sync = "512KB"

[raftdb.defaultcf]
block-size = "64KB"
block-cache-size = "2GB"
disable-block-cache = false
cache-index-and-filter-blocks = true
pin-l0-filter-and-index-blocks = true
use-bloom-filter = false
whole-key-filtering = true
bloom-filter-bits-per-key = 10
block-based-bloom-filter = false
read-amp-bytes-per-bit = 0
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1
max-bytes-for-level-base = "512MB"
target-file-size-base = "8MB"
level0-file-num-compaction-trigger = 4
level0-slowdown-writes-trigger = 20
level0-stop-writes-trigger = 36
max-compaction-bytes = "2GB"
compaction-pri = 0
dynamic-level-bytes = false
num-levels = 7
max-bytes-for-level-multiplier = 10
compaction-style = 0
disable-auto-compactions = false
soft-pending-compaction-bytes-limit = "64GB"
hard-pending-compaction-bytes-limit = "256GB"

[security]
ca-path = ""
cert-path = ""
key-path = ""

[import]
import-dir = "/tmp/tikv/import"
num-threads = 8
stream-channel-window = 128

参数比对说明:

TiKV中的默认参数很多,其实下列几组的参数名是一样的:

[rocksdb.defaultcf]
[rocksdb.writecf]
[rocksdb.lockcf]
[rocksdb.raftcf]
[raftdb.defaultcf]

TiKV中的上述组的多数参数配置是一致的,参数有差异的如下:

block-size 
block-cache-size
compression-per-level
max-bytes-for-level-base
level0-file-num-compaction-trigger
compaction-pri

其中[rocksdb]和[raftdb]组中的参数一致,[rocksdb] 多了如下参数:

rate-bytes-per-sec = "0KB"

TiDB官方对TiKV参数的部分解释:(此部分参数仅供参考,可能不适用2.0版本)

# 日志级别,可选值为:trace,debug,info,warn,error,off
log-level = "info"

[server]
# 监听地址
# addr = "127.0.0.1:20160"

# 建议使用默认值
# notify-capacity = 40960
# messages-per-tick = 4096

# gRPC 线程池大小
# grpc-concurrency = 4
# TiKV 每个实例之间的 gRPC 连接数
# grpc-raft-conn-num = 10

# TiDB 过来的大部分读请求都会发送到 TiKV 的 coprocessor 进行处理,该参数用于设置
# coprocessor 线程的个数,如果业务是读请求比较多,增加 coprocessor 的线程数,但应比系统的
# CPU 核数小。例如:TiKV 所在的机器有 32 core,在重读的场景下甚至可以将该参数设置为 30。在没有
# 设置该参数的情况下,TiKV 会自动将该值设置为 CPU 总核数乘以 0.8。
# end-point-concurrency = 8

# 可以给 TiKV 实例打标签,用于副本的调度
# labels = {zone = "cn-east-1", host = "118", disk = "ssd"}

[storage]
# 数据目录
# data-dir = "/tmp/tikv/store"

# 通常情况下使用默认值就可以了。在导数据的情况下建议将该参数设置为 1024000。
# scheduler-concurrency = 102400
# 该参数控制写入线程的个数,当写入操作比较频繁的时候,需要把该参数调大。使用 top -H -p tikv-pid
# 发现名称为 sched-worker-pool 的线程都特别忙,这个时候就需要将 scheduler-worker-pool-size
# 参数调大,增加写线程的个数。
# scheduler-worker-pool-size = 4

[pd]
# pd 的地址
# endpoints = ["127.0.0.1:2379","127.0.0.2:2379","127.0.0.3:2379"]

[metric]
# 将 metrics 推送给 Prometheus pushgateway 的时间间隔
interval = "15s"
# Prometheus pushgateway 的地址
address = ""
job = "tikv"

[raftstore]
# 默认为 true,表示强制将数据刷到磁盘上。如果是非金融安全级别的业务场景,建议设置成 false,
# 以便获得更高的性能。
sync-log = true

# Raft RocksDB 目录。默认值是 [storage.data-dir] 的 raft 子目录。
# 如果机器上有多块磁盘,可以将 Raft RocksDB 的数据放在不同的盘上,提高 TiKV 的性能。
# raftdb-dir = "/tmp/tikv/store/raft"

region-max-size = "384MB"
# region 分裂阈值
region-split-size = "256MB"
# 当 region 写入的数据量超过该阈值的时候,TiKV 会检查该 region 是否需要分裂。为了减少检查过程
# 中扫描数据的成本,数据过程中可以将该值设置为32MB,正常运行状态下使用默认值即可。
region-split-check-diff = "32MB"

[rocksdb]
# RocksDB 进行后台任务的最大线程数,后台任务包括 compaction 和 flush。具体 RocksDB 为什么需要进行 compaction,
# 请参考 RocksDB 的相关资料。在写流量比较大的时候(例如导数据),建议开启更多的线程,
# 但应小于 CPU 的核数。例如在导数据的时候,32 核 CPU 的机器,可以设置成 28。
# max-background-jobs = 8

# RocksDB 能够打开的最大文件句柄数。
# max-open-files = 40960

# RocksDB MANIFEST 文件的大小限制.
# 更详细的信息请参考:https://github.com/facebook/rocksdb/wiki/MANIFEST
max-manifest-file-size = "20MB"

# RocksDB write-ahead logs 目录。如果机器上有两块盘,可以将 RocksDB 的数据和 WAL 日志放在
# 不同的盘上,提高 TiKV 的性能。
# wal-dir = "/tmp/tikv/store"

# 下面两个参数用于怎样处理 RocksDB 归档 WAL。
# 更多详细信息请参考:https://github.com/facebook/rocksdb/wiki/How-to-persist-in-memory-RocksDB-database%3F
# wal-ttl-seconds = 0
# wal-size-limit = 0

# RocksDB WAL 日志的最大总大小,通常情况下使用默认值就可以了。
# max-total-wal-size = "4GB"

# 可以通过该参数打开或者关闭 RocksDB 的统计信息。
# enable-statistics = true

# 开启 RocksDB compaction 过程中的预读功能,如果使用的是机械磁盘,建议该值至少为2MB。
# compaction-readahead-size = "2MB"

[rocksdb.defaultcf]
# 数据块大小。RocksDB 是按照 block 为单元对数据进行压缩的,同时 block 也是缓存在 block-cache
# 中的最小单元(类似其他数据库的 page 概念)。
block-size = "64KB"

# RocksDB 每一层数据的压缩方式,可选的值为:no,snappy,zlib,bzip2,lz4,lz4hc,zstd。
# no:no:lz4:lz4:lz4:zstd:zstd 表示 level0 和 level1 不压缩,level2 到 level4 采用 lz4 压缩算法,
# level5 和 level6 采用 zstd 压缩算法,。
# no 表示没有压缩,lz4 是速度和压缩比较为中庸的压缩算法,zlib 的压缩比很高,对存储空间比较友
# 好,但是压缩速度比较慢,压缩的时候需要占用较多的 CPU 资源。不同的机器需要根据 CPU 以及 I/O 资
# 源情况来配置怎样的压缩方式。例如:如果采用的压缩方式为"no:no:lz4:lz4:lz4:zstd:zstd",在大量
# 写入数据的情况下(导数据),发现系统的 I/O 压力很大(使用 iostat 发现 %util 持续 100% 或者使
# 用 top 命令发现 iowait 特别多),而 CPU 的资源还比较充裕,这个时候可以考虑将 level0 和
# level1 开启压缩,用 CPU 资源换取 I/O 资源。如果采用的压缩方式
# 为"no:no:lz4:lz4:lz4:zstd:zstd",在大量写入数据的情况下,发现系统的 I/O 压力不大,但是 CPU
# 资源已经吃光了,top -H 发现有大量的 bg 开头的线程(RocksDB 的 compaction 线程)在运行,这
# 个时候可以考虑用 I/O 资源换取 CPU 资源,将压缩方式改成"no:no:no:lz4:lz4:zstd:zstd"。总之,目
# 的是为了最大限度地利用系统的现有资源,使 TiKV 的性能在现有的资源情况下充分发挥。
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]

# RocksDB memtable 的大小。
write-buffer-size = "128MB"

# 最多允许几个 memtable 存在。写入到 RocksDB 的数据首先会记录到 WAL 日志里面,然后会插入到
# memtable 里面,当 memtable 的大小到达了 write-buffer-size 限定的大小的时候,当前的
# memtable 会变成只读的,然后生成一个新的 memtable 接收新的写入。只读的 memtable 会被
# RocksDB 的 flush 线程(max-background-flushes 参数能够控制 flush 线程的最大个数)
# flush 到磁盘,成为 level0 的一个 sst 文件。当 flush 线程忙不过来,导致等待 flush 到磁盘的
# memtable 的数量到达 max-write-buffer-number 限定的个数的时候,RocksDB 会将新的写入
# stall 住,stall 是 RocksDB 的一种流控机制。在导数据的时候可以将 max-write-buffer-number
# 的值设置的更大一点,例如 10。
max-write-buffer-number = 5

# 当 level0 的 sst 文件个数到达 level0-slowdown-writes-trigger 指定的限度的时候,
# RocksDB 会尝试减慢写入的速度。因为 level0 的 sst 太多会导致 RocksDB 的读放大上升。
# level0-slowdown-writes-trigger 和 level0-stop-writes-trigger 是 RocksDB 进行流控的
# 另一个表现。当 level0 的 sst 的文件个数到达 4(默认值),level0 的 sst 文件会和 level1 中
# 有 overlap 的 sst 文件进行 compaction,缓解读放大的问题。
level0-slowdown-writes-trigger = 20

# 当 level0 的 sst 文件个数到达 level0-stop-writes-trigger 指定的限度的时候,RocksDB 会
# stall 住新的写入。
level0-stop-writes-trigger = 36

# 当 level1 的数据量大小达到 max-bytes-for-level-base 限定的值的时候,会触发 level1 的
# sst 和 level2 种有 overlap 的 sst 进行 compaction。
# 黄金定律:max-bytes-for-level-base 的设置的第一参考原则就是保证和 level0 的数据量大致相
# 等,这样能够减少不必要的 compaction。例如压缩方式为"no:no:lz4:lz4:lz4:lz4:lz4",那么
# max-bytes-for-level-base 的值应该是 write-buffer-size 的大小乘以 4,因为 level0 和
# level1 都没有压缩,而且 level0 触发 compaction 的条件是 sst 的个数到达 4(默认值)。在
# level0 和 level1 都采取了压缩的情况下,就需要分析下 RocksDB 的日志,看一个 memtable 的压
# 缩成一个 sst 文件的大小大概是多少,例如 32MB,那么 max-bytes-for-level-base 的建议值就应
# 该是 32MB * 4 = 128MB。
max-bytes-for-level-base = "512MB"

# sst 文件的大小。level0 的 sst 文件的大小受 write-buffer-size 和 level0 采用的压缩算法的
# 影响,target-file-size-base 参数用于控制 level1-level6 单个 sst 文件的大小。
target-file-size-base = "32MB"

# 在不配置该参数的情况下,TiKV 会将该值设置为系统总内存量的 40%。如果需要在单个物理机上部署多个
# TiKV 节点,需要显式配置该参数,否则 TiKV 容易出现 OOM 的问题。
# block-cache-size = "1GB"

[rocksdb.writecf]
# 保持和 rocksdb.defaultcf.compression-per-level 一致。
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]

# 保持和 rocksdb.defaultcf.write-buffer-size 一致。
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1

# 保持和 rocksdb.defaultcf.max-bytes-for-level-base 一致。
max-bytes-for-level-base = "512MB"
target-file-size-base = "32MB"

# 在不配置该参数的情况下,TiKV 会将该值设置为系统总内存量的 15%。如果需要在单个物理机上部署多个
# TiKV 节点,需要显式配置该参数。版本信息(MVCC)相关的数据以及索引相关的数据都记录在 write 这
# 个 cf 里面,如果业务的场景下单表索引较多,可以将该参数设置的更大一点。
# block-cache-size = "256MB"

[raftdb]
# RaftDB 能够打开的最大文件句柄数。
# max-open-files = 40960

# 可以通过该参数打开或者关闭 RaftDB 的统计信息。
# enable-statistics = true

# 开启 RaftDB compaction 过程中的预读功能,如果使用的是机械磁盘,建议该值至少为2MB。
# compaction-readahead-size = "2MB"

[raftdb.defaultcf]
# 保持和 rocksdb.defaultcf.compression-per-level 一致。
compression-per-level = ["no", "no", "lz4", "lz4", "lz4", "zstd", "zstd"]

# 保持和 rocksdb.defaultcf.write-buffer-size 一致。
write-buffer-size = "128MB"
max-write-buffer-number = 5
min-write-buffer-number-to-merge = 1

# 保持和 rocksdb.defaultcf.max-bytes-for-level-base 一致。
max-bytes-for-level-base = "512MB"
target-file-size-base = "32MB"

# 通常配置在 256MB 到 2GB 之间,通常情况下使用默认值就可以了,但如果系统资源比较充足可以适当调大点。
block-cache-size = "256MB"

TiDB 系统调优的官方建议:

TiKV 内存使用情况

除了以上列出的 block-cache 以及 write-buffer 会占用系统内存外:

需预留一些内存作为系统的 page cache
TiKV 在处理大的查询的时候(例如 select * from ...)会读取数据然后在内存中生成对应的
数据结构返回给 TiDB,这个过程中 TiKV 会占用一部分内存。

TiKV 机器配置推荐

生产环境中,不建议将 TiKV 部署在 CPU 核数小于 8 或内存低于 32GB 的机器上
如果对写入吞吐要求比较高,建议使用吞吐能力比较好的磁盘
如果对读写的延迟要求非常高,建议使用 IOPS 比较高的 SSD 盘

实际使用过程中需要重点关注的参数:

推荐的TiKV 参数配置   
sync-log = false
grpc-concurrency = 8
grpc-raft-conn-num = 24 
[defaultcf]
block-cache-size = "12GB"
[writecf]
block-cache-size = "5GB"
[raftdb.defaultcf]
block-cache-size = "2GB"

实际的配置:
[server]
grpc-concurrency = 4
grpc-raft-conn-num = 10

[raftstore]
sync-log = true
[rocksdb.defaultcf]
block-cache-size = "48331MB"
[rocksdb.writecf]
block-cache-size = "28998MB"
[raftdb.defaultcf]
block-cache-size = "2GB"

针对48线程的CPU和188G内存的主机配置修改:

[server]
grpc-concurrency = 4   -->8
grpc-raft-conn-num = 10 -->24
[storage]
scheduler-concurrency = 2048000 
scheduler-worker-pool-size = 8  -->16
[coprocessor]
region-max-size = "144MB"  -->384MB
region-split-size = "96MB"  -->256MB
[raftstore]
region-split-check-diff = "6MB" -->32MB
sync-log = true  

[rocksdb]
max-background-jobs = 6   -->32
max-open-files = 40960    -->65535
[rocksdb.defaultcf]
block-cache-size = "48331MB"  -->80G
[rocksdb.writecf]
block-cache-size = "28998MB"   -->30G
[raftdb.defaultcf]
block-cache-size = "2GB"

此外参数需要调整为32MB。
target-file-size-base = "8MB" -->32MB

参考:

https://github.com/facebook/rocksdb/wiki/RocksDB-Tuning-Guide

https://github.com/facebook/mysql-5.6/wiki/my.cnf-tuning