MySQL Replication Performance Tuning for Fun and Profit!

Copyright © 2017, Oracle and/or its afliates. All rights reserved.
MySQL Replicatio erfirmaoce Tuoiog fir Fuo aod
rifit!
26ith
Sepitember 2017
Víitir Oliveira (viitir.s.p.iliveira@iracle.cim)
Seoiir erfirmaoce Eogioeer
1

Safe Harbor Statement
The following is intended to outline our general product directon. It is intended for
informaton purposes only, and may not be incorporated into any contract. It is not a
commitment to deliver any material, code, or functonality, and should not be relied upon
in making purchasing decisions. The development, release, and tming of any features or
functonality described for Oracle’s products remains at the sole discreton of Oracle.
Copyright © 2017, Oracle and/or its afliates. All rights reserved. Dublin, 26th September 2017 2

Copyright © 2017, Oracle and/or its afliates. All rights reserved. Dublin, 26th September 2017
Replicaton Performance
Asynchronous Replicaton
Group Replicaton
Impact of the latest improvements
Conclusion
Contents
3
1
2
3
4
5
This session is about Asynchronous and Group Replicaton performance and the optons that
may help support the most demanding workloads.
Some of the latest development eforts to improve the performance of MySQL replicaton
will also be presented.

Replicatio perfirmaoce
●
MySQL Replicaton overview
●
Replicaton performance
●
Replicaton delay
1

1.1 MySQL Replicatio Overview (1/3)
Replicatio
– MySQL Replicaton transfers database changes from one server to its replicas.
Sitarits wheo itraosactios are prepared
– Replicaton preserves the data changes in a binary log fle on the master just before the
transactons are ready to commit;
Traosfers daitabase chaoges io a bioary lig
– The binary log fle contains an engine-agnostc representaton of the data changes;
– It is sent to replicas whenever they request it or can be used for backup, point-in-tme recovery
and integraton purposes.
5

1.1 MySQL Replicaton Overview (2/3)
6
binary
Log
INSERT ...
DB
relay
log
INSERT ...
binary
log
INSERT...
DB
User
Applicaton “INSERT...”
networknetwork
Server
A
Server
B

1.1 MySQL Replicaton Overview (3/3)
Sitaitemeoit- vs Riw-based bioary ligs
– The binlog can either carry the statements issued by the user or the deterministc row-
changes resultng from those statements;
Are asyochrioiusly applied
– The slaves stream the binary log from the master to an intermediate fle on the slaves, the
relay log, which is discarded once it is no longer needed;
– Workloads are applied asynchronously using a best efort approach; it competes for resources
with other workloads on the slaves.
Usiog siogle ir mult-ithreaded schedulers
– The slaves scan the relay log and apply the data changes to the database using one of the
available appliers: single-threaded, DATABASE or LOGICAL_CLOCK.
7

1.2 Replicatio perfirmaoce (1/3)
erfirmaoce
– Performance can be many things, lets focus on concrete metrics like transacton ithriughpuit
and laiteocy.
(efciently commit as many transactons per second/hour as possible, having the smallest
delay between the moment a transactons starts untl all its efects are seen)
Twi meitrics fir replicatio ithriughpuit:
– susitaioed ithriughpuit : the rate the system as a whole can withstand without having some
servers lagging behind others in the long run;
– peak ithriughpuit : the maximum transacton rate that clients can obtain from the client-facing
servers.
8

1.2 Replicaton performance (2/3)
eak vs susitaioed ithriughpuit
– Peak and sustained capacity are similar when the slaves deliver enough throughput to handle
the masters’ workload;
– But some workloads are very demanding from replicaton point-of-view, and may be harder to
process by the slaves.
The eak iti Susitaioed ithriughpuit gap
– Is an exitra margin that can be used for shirit periids
● someone posts a link to a site on slashdot, or is referenced in the news;
● for some tme that site has more visitors than usual, then returns to normal.
– When demand regularly exceeds sustained throughput, its tme to resize the underlying
system!
9

1.2 Replicaton performance (3/3)
Ao example
– In Sysbench RW the slave delivers the same throughput as the master;
But when reads are stripped from those transactons the extra work the slave must perform
becomes visible afer a certain level of throughput.
(Note that when replying GR has transactons network ordered already, which justfes the lower peak throughput).
10
8 16 32 64 128 256
0
5 000
10 000
15 000
20 000
Peak vs Sustained Throughput: Workload with reads (balanced)
Asynchronous Replication PEAK Group Replication PEAK
Asynchronous Replication SUSTAINED Group Replication SUSTAINED
number of client threads
Sysbench RW transactions/second
8 16 32 64 128 256
0
10 000
20 000
30 000
40 000
Peak vs Sustained Throughput: Write-only workload
Asynchronous Replication PEAK Group Replication PEAK
Asynchronous Replication SUSTAINED Group Replication SUSTAINED
Sysbench OLTP RW - Write-only transactions/second

1.3 Replicatio delay (1/3)
A replicaited itraosactio cao be divided io several executio tmes
– The tme to execuite ithe itraosactio untl it is ready to be replicated (storage engine);
– The tme to exitracit aod itraosfer the transacton to a remote master (data changes extracton,
transfer over the network, store remotely);
– The tme to read aod schedule remiite executio (read relay log, wait for dependencies, wait
for an available worker);
– The tme to execuite ithe itraosactio remiitely (storage engine).
11
Front-end
skip-log-bin
binlog
S e rv er
A
INS E RT...
s y nc
sync_binlog=0
re la y log
S e rv e r
B
INS E RT...
D um p
T hre a d
IO
T hrea d
S QL
T hrea d
binlog
s y ncWnW1 ...
Back-end
sync_binlog=0
sync_binlog=1
MTS
sync_binlog=1
client-visible latency replication latency

1.3 Replicaton delay (2/3)
12
Front-end
skip-log-bin
binlog
Server
A
INSERT...
sync
sync_binlog=0
relaylog
Server
B
INSERT...
SQL
Thread
binlog
sync
log-slave-updates=OFF
WnW1 ...
Back-end
sync_binlog=0
sync_binlog=1
MTS
sync_binlog=1
client-visible latency replication latency
networknetwork
gr

1.3 Replicaton delay (3/3)
The replicatio delay shiuld oiit griw iver tme
– That happens when demand remains below the sustained capacity;
– If the demand exceeds the sustained capacity the replicaton delay will increase over tme;
● workload needs to return to manageable levels before exhaustng the bufering resources;
● if peaks are not intertwined with low-demand periods the slave may be unable catch up with
the back-log.
The mioimum replicatio delay is ithe tme iit itakes iti repeait ithe itraosactio io ithe slave
– Transactons don’t start to replicate untl they are almost ready to commit (prepared) on the
master.
13

Asyochrioius Replicatio2
●
The replicaton applier
●
Executng transactons in parallel
●
Replicaton applier optons

2.1 The replicatio applier
– The replicaton applier, or slave applier, takes the database changes from the master and
executes them locally on each slave;
– It reads the relay log and applies the transacton events as they are read, either directly or
using auxiliary worker threads;
– The applier throughput determines the maximum sustained capacity of a replicated system.
15

2.1 The replicaton applier
Applier schedulers
– On MySQL 5.5 and below the applier was single-threaded, which started to show performance
limitatons on write-intensive workloads when processors started having several executon
cores;
● It is stll useful for non-write-intensive workloads;
– MySQL 5.6 introduced the frst mult-threaded applier to parallelize transactons working on
diferent databases;
● achieves good scalability on workloads with many schemas;
16

2.1 The replicaton applier
Geoeral purpise parallel scheduler
– MySQL 5.7 started parallelizing transactons working within the same database with the
introducton of the LOGICAL_CLOCK scheduler;
– The LOGICAL_CLOCK scheduler achieves good scalability using the intra-schema parallelism,
using dependency tracking on the master.
– Group Replicaton and MySQL 8.0 use the LOGICAL_CLOCK scheduler with transacton writeset
dependency tracking.
Optios:
--slave-parallel-itype=LOGICAL_CLOCK
--slave-parallel-wirkers=<o ithreads>
17

2.2 Executog itraosactios io parallel
Trackiog iodepeodeoit itraosactios
– With LOGICAL_CLOCK transactons are allowed to execute in parallel in the slave if the master
has considered them independent, or non-confictng, when they were executed there.
Hiw iit wirks
– The dependency informaton is transferred from the master to the slaves using a pair
(last_commited, sequence_number) for each transacton in the binlog;
– A transacton is only scheduled to run on the slave if the transacton with the
sequence_number equal to this transacton last_commited has already been commited;
– No subsequent transactons are scheduled to run untl that last_commited conditon is
fulflled for that partcular transacton.
18

2.2.1 Lock-based dependency tracking (1/5)
Lick ir griup cimmiit based itrackiog
– Initally transactons were considered independent in LOGICAL_CLOCK if they were prepared
within the same group commit (WL#6314);
– MySQL 5.7.2 improved that to allow two transactons to execute in parallel if they hold all their
locks at the same tme (WL#7165);
● That increased the parallelizaton window to include transacton beyond the group commit
boundaries, and removed the serializaton between commit groups for a more stable parallel
executon.
19

20
step 1 step n step n+1 step n+2 step n+3 step n+4
T1T1
Session ASession A
T2T2
Session BSession B
T3T3
Session BSession B
T4T4
Session ASession A
T5T5
Session CSession C
T6T6
Session CSession C
T7T7
Session CSession C
T8T8
Session BSession B
tme
databaseaddressspace

Delayiog griup cimmiit io ithe masiter
– It may be benefcial to delay the commits so that more transactons are prepared in parallel,
partcularly on fast storage systems, using:
●
binlog-group-commit-sync-delay = <n microseconds>
– Don’t wait when many threads are already waitng to commit
●
binlog-group-commit-sync-no-delay-count = <k user threads>
Hiwever
– Due to the nature of this detecton scheme, transactons from the same client are never executed
in parallel.
21

22
step 1 step n step n+1 step n+2 step n+3
T1T1
Session ASession A
T2T2
Session BSession B
T3T3
Session BSession B
T4T4
Session ASession A
T5T5
Session CSession C
T6T6
Session CSession C
T7T7
Session CSession C
T8T8
Session BSession B
tme
T7T7
Session CSession C

Drawbacks if lick-based depeodeocy itrackiog
– The cimmiit delays iocrease transacton latency on the master;
– It may also decreased ithriughpuit if ithe masiter when the number of clients is not sufcient
to properly explore the available capacity;
– And parallelism is reduced when going through intermediate masters:
●
The number of applier-threads on the slave is usually smaller than user-threads, and
transactons from the same thread are never parallelized;
●
Parallelism lost this way is not recovered.
23

2.2.2 Writeset-based dependency tracking (1/4)
Track chaoged riws
– With writeset-based dependency tracking the transactons are considered independent, and
can run in parallel on the slave, if they touch non-overlapping data;
– Overlapping data is detected by hashing the unique keys of each row touched by the
transacton and saving that in memory untl the transacton-history-size limit is reached;
– That same informaton is used by Group Replicaton to detect which writes can be accepted in
mult-master mode.
Wriite seit requiremeoits
– Table must have Primary Keys / Unique Keys;
– Write set extracton must be turned on (default afer MySQL 8.0.3).
24

25
step n step n+1 step n+2 step n+3 step n+4
T1T1
Session ASession A
T2T2
Session BSession B
T3T3
Session BSession B
T4T4
Session ASession A
T5T5
Session CSession C
T6T6
Session CSession C
T7T7
Session CSession C
T8T8
Session BSession B
tme
step 1
Write set history size
binlog-transaction-dependency-tracking=
WRITESET

Limiitatios (falls back iti cimmiit irder)
– Foreign-key constraints unsupported;
– DDL do not have a write set.
Triubles wiith slave cimmiit hisitiry
– Applying transactons in parallel may result in a diferent commit history than that on the
originatng server;
– slave_preserve_commit_order may be useful in those cases that commit order is required;
– WRITESET_SESSION can be used if session order is sufcient.
26

2.2.2 WRITESET_SESSION dependency tracking
27
step 1 step n step n+1 step n+2 step n+3 step n+4
T1T1
Session ASession A
T2T2
Session BSession B
T3T3
Session BSession B
T4T4
Session ASession A
T5T5
Session CSession C
T6T6
Session CSession C
T7T7
Session CSession C
T8T8
Session BSession B
tmeWrite set history size
binlog-transaction-dependency-tracking=
WRITESET_SESSION

2.2.3 Dependency tracking impact
28
1 2 4 8 16 32 64 128 256
0
5000
10000
15000
20000
25000
30000
35000
40000
45000
50000
Applier Thriughpuit: Sysbeoch Updaite Iodex
COMMIT_ORDER WRITESET WRITESET_SESSION
Number if Clieoits io ithe Masiter
Updaites/seciodAppliesioitheReplica
Summary
WRITESET tracking allows a
single threaded workload to be
applied in parallel.
It also delivers the best and
almost constant throughput at
any concurrency level.
WRITESET_SESSION depends on
the number of clients on the
master, but it stll delivers
higher throughput at lower
number of clients than that of
COMMIT_ORDER.

2.3 Replicatio applier iptios
Ficusiog io LOGICAL_CLOCK, leit’s italk abiuit perfirmaoce implicatios if sime iptios:
– Number of worker threads
– Preserve commit order on slave
– Durable vs non-durable setngs
– Log slave updates (atached)
– Binary log format (atached)
29

2.3.1 Number of worker threads
The applier perfirmaoce is biuod iti ithe oumber if wirkers
– slave-parallel-workers=<number of workers>
arallel applier beoefits frim itwi maio facitirs:
– arallel executio
multple transactons can execute in parallel on the CPU, which scales between 5-10X in a
modern machine;
– IO laiteocy hidiog
storage-bound workloads beneft from more workers to hide the storage latency associated
with the writng and syncing to disk.
The efectveoess depeods partcularly io:
– the workload characteristcs and
– the dependency tracking on the master (next session).
30

2.3.2 Number of workers and preserve commit-order
• With the parallel applier the order of the transactons on the slaves is not guaranteed to be ordered;
• To force that, the optons –-slave-preserve-commit-order=ON must be used.
31
0 1 2 4 8 12 16 64
0
5000
10000
15000
20000
25000
Slave Applier Throughput: Sysbench RW
SPCO=OFF SPCO=ON
number of applier threads
transactions/second applied on slave
0 1 2 4 8 12 16 64
0
10000
20000
30000
40000
50000
Slave Applier Throughput: Sysbench Update Index
SPCO=OFF SPCO=ON
number of applier threads
updates/second applied on slave

2.3.3 Number of worker threads and durability
32
16 24 32 64 16 24 32 64
Sysbench OLTP RW Sysbench Update Index
0
5000
10000
15000
20000
25000
30000
35000
40000
45000
Slave Applier Throughput Varying Applier Workers: Sysbench RW
NON-DURABLE DURABLE
number of applier workers
transactions/second applied on slave
Observatios
Non-durable has mostly
constant beneft when
enough threads are
available (16+)
Durable setngs can
beneft from more
threads as they hide
sync latency.

2.3.4 Durable vs non-durable setngs
Ti haodles crashes wiithiuit daita liss:
– the database and binary log must be made durable using sync-binlog=1 and innodb-fush-log-
at-trx-commit=1.
– that comes at the cost of having to wait for a sync on disk to complete.
33
8 16 32 64 128 256
0
5 000
10 000
15 000
Durable vs Non-durable Setting: Sysbench RW Sustained Throughput
Asynch. Replication NON-DURABLE Group Replication NON-DURABLE
Asynch. Replication DURABLE Group Replication DURABLE
transactions/second
8 16 32 64 128 256
0
10 000
20 000
30 000
40 000
50 000
Durable vs Non-durable Setting: Sysbench Update Index Sustained Throughput
Asynch. Replication NON-DURABLE Group Replication NON-DURABLE
Asynch. Replication DURABLE Group Replication DURABLE
updates/second

Griup Replicatio3
●
Transacton fow
●
Performance enhancing
•
Group communicaton
•
Certfcaton
•
Multple group writers
•
Remote applier
●
Flow-control

Group Replicaton
“Single primary/Mult-master updaite everywhere replicaton plugin for MySQL with
built-in auitimatc disitribuited recivery, cioficit deitectio and griup membership.”
35

3.1 Traosactio fiw
Execuite licaly
– Group Replicaton starts replicatng a transacton when it is ready to commit, just before being
writen to the binary log;
Seod io-irder iti all members
– At that point, transactons are broadcast to the network using a group communicaton
protocol (Paxos, similar to the Mencius variant);
Certfy iodepeodeoitly
– All members receive the transactons in order and execute a deterministc certfcaton
algorithms to check if the received transacton can be applied safely;
Apply asyochrioiusly
– On remote members successfully certfed transactons are writen to the relay log and
asynchronously applied by the members, just as happens for other replicaton methods.
– On local member, the prepared transactons is commited to the storage engine.
36

3.1 Transacton fow: writer-side (primary node)
37
Transaction
Hook
Begin
Execute transaction
(until prepare)
Needs to
throttle?
Rollback
Commit
Delay until next
flow-control period
     Yes
Collect writeset
information
Send message for
ordering by GCS
Wait result from
certification thread
Certification
positive?
No
Certification
certification result
Group
Communication

3.1 Transacton fow: receiver-side (all nodes)
38
         Certifier
         Loop
Start Quit?
End
        Yes
Get next transaction
in queue
Certify transaction
Is
transaction
local?
Applier
Certification
positive?
User
Threads
Send certification result
to user thread
Add transaction events
to relay log
  Yes
  Yes
signal
release
GCS
Receive
Thread
Ordered
Transaction
Queue

3.2 erfirmaoce eohaociog
Neitwirk baodwidith aod laiteocy
– Group Replicaton sends messages to all members of the group and wait for the majority to
agree in a specifc order of transactons;
– That consumes network bandwidth and adds at least one network RTT to the transacton
latency.
Certfcatio ithriughpuit
– Transactons are certfed in the agreed message delivery order and writen to the relay log by
a single thread;
– That can become a botleneck if the certfcaton rate is very high or if relay log writes are slow.
Remiite itraosactio applier ithriughpuit
– Remote transactons can be applied by the single or the mult-threaded applier, where it
benefts from WRITESET-based dependency tracking (where it frst appeared).
39

3.2.1 Network bandwidth and latency (1/2)
Optios iti reduce baodwidith usage
– group_replicaton_compression_threshold=<size: 100>
– binlog_row_image=MINIMAL
Reduce laiteocy wiith busy waitog befire sleepiog
– group_replicaton_poll_spin_loops=<num_spins>
40

3.2.1 Network bandwidth and latency (2/2)
Ciocurreoit ciooectios hide oeitwirk laiteocy iocrease
The group communicaton protocol used by Group Replicaton is very efectve on low-latency networks.
But supports high-latency networks, although a higher number of concurrent client threads is required to
achieve the same performance.
41
8 16 32 64 128 256 512 1024
0
2 500
5 000
7 500
10 000
12 500
15 000
Effects of network round-trip latency: Sysbench RW
3 nodes in 10Gbps LAN 2 nodes in 10Gbps LAN, 1 node @ 7ms, 200Mbps
number of client threadsthroughput (TPS)
8 16 32 64 128 256
0
10 000
20 000
30 000
40 000
50 000
Asynchronous vs Group Replication Sustained Throughput
Asynchronous Replication (WS) Group Replication
Sysbench Update Index updates/second

3.2.2 Certfer throughput (1/2)
42
Use high-perfirmaoce sitirage fir ithe relay lig
– Improve the disk bandwidth as each event is sent by the certfer thread to the relay log while
also being read by the applier;
Use fasit itempirary direcitiry
– Transacton writesets are extracted using the IO_CACHE, and that may need to spill-over to
disk on large transactons.
– If using large transactons setng a fast tmpdir may improve throughput:
●
tmpdir=/dev/shm/...

3.2.2 Certfer throughput (2/2)
Reduce certfcatio cimplexiity io mult-masiter
– Simplifying GTID handling increases the throughput of the certfer, as the certfcaton
informaton requires less efort.
●
group_replicaton_gtd_assignment_block_size=<size: 10M>
– This may need to be changed for very high transacton rate.
43

3.2.3 Multple group writers
Iocrease read/wriite capaciity usiog mult-masiter
– When transactons writen to a group have a large read component, the throughput may be
larger if the transacton are sent to multple members.
– Some restrictons on mult-master must be considered (check docs).
44
8 16 32 64 128 256
0
5 000
10 000
15 000
20 000
Multi-master Sustained Throughput: Sysbench OLTP RW
1-writer non-durable 2-writers non-durable
1-writer durable 2-writers durable
transactions/second
8 16 32 64 128 256
0
5 000
10 000
15 000
20 000
Multi-master Sustained Throughput: Sysbench OLTP RW Write-only
1-writer non-durable 2-writers non-durable
1-writer durable 2-writers durable
transactions/second

3.3 Fliw-cioitril (1/4)
Optmize ithe sysitem as a while
– Sometmes it is benefcial to delay some parts of a distributed system it to improve the
throughput of the system as a whole;
– In Group Replicaton, it is used to
● keep writers operatng below the sustained capacity of the system;
● reduce bufering stress on the replicaton pipeline;
● protect the correct executon of the system.
45

3.3 Flow-control (2/4)
Desigoed as a safeity measure
– Flow-control should be DISABLED in a properly sized deployment, where system administrators
are fully aware of
● The implicatons of high-availability and
● of the real capacity of the underlying system to deal with the workload.
– If fow-control is disabled:
● members will not be delayed in any way, writers will achieve the maximum network
throughput and if the slaves are operatng normally, they should keep up with the writers;
– Even if fow-control is ENABLED, throtling will never be actve while the system is operatng
below its sustained capacity.
46

Ti beter suppirit uobalaoced sysitems aod uofrieodly wirkliads:
– Reduce the number of transactons aborts in mult-master:
● Rows from transactons in the relay log cannot be used by new transactons.
● Otherwise, the transacton will be forced to fail at certfcaton tme.
– Keep members closer for faster failover:
● Failing over to an up-to-date member is faster as the back-log to apply is smaller.
– Keep members closer for reduced replicaton lag:
● Applicatons using one writer and multple readers will see more up-to-date informaton when
reading from members other than the writer.
47

Make sure oew members cao always jiio wriite-ioiteosive griups
– Nodes entering the group need to catch up previous work while also storing current work to
apply later.
– This is more demanding then just applying remote transactons;
– In case excess capacity is not available, the cluster will need to be put at lower throughput for
new members to catch up.
48

3.3.1 Flow-control fow
49
        Flow-control
        Loop
Start Quit?
End
        Yes
Find members with
excessive queueing
Send stats message
to group
Are all
members
ok?
Throttling
Active?
quota = min. capacity /
number of writers
(minus 10%, up to 5%
of thresholds)
Release throttling
gradually
(50% increase per step)
  Yes

Stats
Messages
Receiver
no
wait one second
& release trans
Member
Execution
Stats
Writer decision
The writers throttle themselves
when they see a large queue on
other members
Delayed members don’t have to
spend extra time dealing with that.
No micro-management
The flow-control is coarse grained,
no micro-management.
The synchrony between nodes is
checked and corrective actions are
performed at regular intervals.

3.3.2 Flow-control optons (MySQL 5.7)
Basic ciofguratio iptios
– group-replicaton-fow-control-mode = QUOTA | DISABLED
– group-replicaton-fow-control-certfer-threshold = 0..n
– group-replicaton-fow-control-applier-threshold = 0..n
Certfer/applier ithreshilds
– The thresholds are the point at which the fow-control system will delay the writes at the
master;
– The default is set to 25000 and should be kept larger then one second of sustained commit
rate;
– But some members will be up to 25000 transactons delayed, if, and only if, they are unable to
keep up with the writer members.
50

3.3.3 Flow-control optons (MySQL 8.0.3)
MySQL 8.0.3 ioitriduces additioal iptios iti foe-ituoe ithe heuristcs:
● group_replicaton_fow_control_min_quota = X commits/s
● group_replicaton_fow_control_min_recovery_quota = X commits/s
● group_replicaton_fow_control_max_commit_quota = X commits/s
● group_replicaton_fow_control_member_quota_percent = Y %
● group_replicaton_fow_control_period = Z seconds
● group_replicaton_fow_control_hold_percent = Y %
● group_replicaton_fow_control_release_percent = Y %
Aod wheo griup-replicatio-fiw-cioitril-mide = DISABLED
● The member will no longer be considered delayed by the others.
51

Impact of the latest improvements4
●
Asynchronous replicaton
●
Group Replicaton throughput
●
Group Replicaton recovery

4.1 Imprived ithriughpuit ait liw ciocurreocy
– Asynchronous Replicaton benefts from Group Replicaton’s writeset dependency tracking to
eliminate replicaton lag at low concurrency;
– The beneft is in parallel executon, as delays on low concurrency were mostly caused by
reduced parallelism in the binary log.
53
8 16 32 64
0
10 000
20 000
30 000
40 000
50 000
Asynchronous Replication Sustained Throughput
Sysbench Update Index, durable settings
MySQL 5.7 MySQL 8.0.3
updates/second
8 16 32 64
0
10 000
20 000
30 000
40 000
50 000
Asynchronous Replication Sustained Throughput
Sysbench Update Index, non-durable settings
number of client threadsupdates/second

4.2 Beter scalabiliity io Griup Replicatio
– Group Replicaton’s achieves higher throughput due to an optmized scalability dependency
tracking to eliminate replicaton lag at low concurrency;
– Group Replicaton is now much close to the maximum sustained throughput of Asynchronous
replicaton, even for small transactons.
54
8 16 32 64 128 256
0
10 000
20 000
30 000
40 000
50 000
Group Replication Sustained Throughput
Sysbench Update Index, non-durable settings
updates/second
8 16 32 64 128 256
0
10 000
20 000
30 000
40 000
50 000
Group Replication Sustained Throughput
Sysbench Update Index, durable settings
updates/second

4.3 Reduced recivery tme
– Group Replicaton can now recover faster using WRITESET dependency tracking and benefts
from the improved applier synchronizaton;
– In this tests the user workload on the writer-node is constant (around 1/3 and 2/3 of the
server capacity for that workload) and we measure how fast a member can enter the group.
55
Sysbench RW at 33% capacity
(workload: 4K TPS on 64 threads)
Sysbench RW at 66% capacity
(workload: 8K TPS on 64 threads)
0
2
4
6
8
10
Group Replication Recovery Time: Sysbench RW
(durable settings, 64 threads, fixed workload)
time to catch-up per time of workload missing (ratio)
Sysbench Update Index at 33% capacity
(workload: 9K updates/s on 64 threads)
Sysbench Update Index at 66% capacity
(workload: 18K updates/s on 64 threads)
0
2
4
6
8
10
Group Replication Recovery Time: Sysbench Update Index
(durable settings, 64 threads, fixed workload)
time to catch-up per unit of workload missing (ratio)

4.4 artal JSON
• MySQL 8.0.3 allows the changes in JSON felds to be sent without updatng the entre feld, signifcantly
reducing the data transferred and stored in the binlog;
• This is partcularly efectve if binlog-images=MINIMAL, since it also prevents the before image to be fully
sent.
(the numbers on the right are from a
specially designed benchmark, where
the tables have 10 JSON felds and in
each transacton only around 10% of
the data changes in one of them)
56
binlog-images=FULL binlog-images=MINIMAL binlog-images=FULL binlog-images=MINIMAL
default PARTIAL_JSON
0
10000
20000
30000
40000
50000
60000
Binary Log Space per Transaction: PARTIAL_JSON
SPACE/TRANS
bytes per transaction

4.4 artal JSON
• Using binlog-row-value-optons=PARTIAL_JSON has implicatons on throughput
– The master works more to extract the changes, so it reduces the throughput;
– However, the throughput of the slave increases, amplifed by MINIMAL images.
57
FULL MINIMAL FULL MINIMAL
0
2000
4000
6000
8000
10000
Throughput on the Slave: PARTIAL_JSON
4 8 16 32 64
Number of threadstransactions per second
binlog-row-image=FULL binlog-row-image=MINIMAL binlog-row-image=FULL binlog-row-image=MINIMAL
0
500
1000
1500
2000
2500
Throughput on the Master: PARTIAL_JSON
4 8 16 32 64
Number of threads
transactions per second

Conclusion5

5 Cioclusiio
– To optmize replicaton performance, focus on the system as a whole
●
sustained throughput, not just client response to the front-end server.
– Replicaton lag should not grow over tme, if it does:
●
make sure there are periods of low demand for the replicas to catch up;
●
make sure you don’t need the data in the replicas to be up-to-date.
59

5 Conclusion
Griup Replicatio
– To optmize Group Replicaton focus on the three areas envolved: group communicaton,
certfer and applier throughputs.
– The performance is compettve with Asynchronous Replicaton
●
Scalable to a signifcant number of members and client threads;
●
High throughput and low-latency even compared to asynchronous replicaton;
●
Optmized for local networks but can withstand higher-latency networks.
60

5 Conclusion
MySQL 8.0 has greait imprivemeoits io replicatio perfirmaoce
– Asynchronous replicaton has now much beter sustained throughput
●
optmized pipeline and writeset dependency tracking;
– Group Replicaton has
●
higher throughput, even for small transactons;
●
members joining a group can recover faster.
61

Thank you.
Any questons?
• Documentaton
– htp://dev.mysql.com/doc/refman/5.7/en/group-replicaton.html
• Performance blog posts related to Group Replicaton:
– htp://mysqlhighavailability.com/category/performance/
62

MySQL Replication Performance Tuning for Fun and Profit!

MySQL Replication Performance Tuning for Fun and Profit!

In this document

More Related Content

What's hot

Viewers also liked

Similar to MySQL Replication Performance Tuning for Fun and Profit!

Recently uploaded

MySQL Replication Performance Tuning for Fun and Profit!