Tip:
Highlight text to annotate it
X
the broadcast is now starting
all attendee far in listen-only mode hello everyone
my name is the least down and I am pecan as director of product marketing
we will begin in just one moment but first I'd like to conduct
a bit of housecleaning first please raise your hand using the hand icon
located in the code within our control panel to let me know that you can hear
me
hey ready and kids think people can hear me which is great
great thank you very much everyone next during this webinar
you will be on mute should you have any questions during the discussion please
enter them in the questions field within the control panel
at the end of the weapon aren't we will take time to answer as many questions as
possible
Anderson unaddressed will be answered in a follow-up blood Coast entry and
percussionist MySQL performance
plaque and the speakers will probably reach out to you individually if it's a
more in-depth
and kinda question in addition to recording
up this webinars will be made available to everyone within 48 hours
so with that said I'd like to thank you for attending today's webinar
multi-day to send to have MySQL with continue in tents tent
it will be presented today by iconic consultant Peter past
aunt we also have a special gets
start today Robert Hutchins who is continuing
Teil and will be starting the show today so with that said I'll turn the floor
over graphic
ahead Robert thank you very much a Lisa
Lisa can you hear me OK yes i cant a perfectly wonderful well it's my
pleasure to be here today
respectfully on a pleasure to have the opportunity to work with Peterborough us
from
I from o'connor so as lisa says we'll be talking about multi data center bicycle
with continuing
tungsten so I'd like to dive right yen
and iight just i tell you a little bit
about lot we're going to be talking about today so first of all
we'd like to start by discussing *** multi data center MySQL so important and
just give you an example of some the complexities that are involved
in achieving it successfully will then go to talk about continuing tungsten
and how it fills the gaps that exist in
building these types of systems well then switch over to peer
to who's going to do it demonstration up a couple different types
of faulty data center topologies and then conclude with a few practical steps
to
multi data center bicycle I should say in passing
that I I if you're not familiar with continuing
I we r he vendor up
I clustering and replication for
particularly from my sequel databases but we also cover many other different
data bytes types as well
I we're particularly focused on this type of problem of
I enabling people to build large systems using
replication an off-the-shelf bicycle and our main product
yes Canton cuts which will be talking so
with that let's dive in and start with this question about
why is it that we want multi data center isql
so there's a number of reasons why you might wanna spread
heater in my sequel across multiple locations
I've listed the three men once I'd like to just go through them briefly
the first and most obvious reason to spread data out
is for high availability so if you have a problem in one location
what you wanna do is have data copies as up to date as possible
in other locations that are way from the problem
and so you can keep running is another reason to move data in
in today's systems and that is that many of us are building systems
that have users were spread around the globe so
in order to give those users adequate performance we want to move the
applications
and the data is close to my sequel rx close to the user's location as possible
so for example if you're have customers in the United States
you prefer to have data centers located within a relatively
job short distance in terms of the latency
so using data centers for example Amazon Easter Amazon West
on the other hand if you have users in Europe he probably like to have them
operating of the data center European reach
and it is a final topic about multi data center MySQL
which that's witches *** Asher II
as your business grows and as you have use are spread out over different
locations
it turns out that many of those users have different legal requirements for
are keeping their data in some locations but not others
so for example it's very common to have data that may not
leave the host country up that it's generated in
I'll this is quite common with privacy regulations so it's yet another reason
for wanting to have lost my sequel in different locations
arm but perhaps for the data partitioned within those locations
we'll talk less about that issue in this talk and focus mostly on high
availability
performance test
which will turn into the HA by topic I'd like to give you just a quick example
that shows that even though conceptually it's very soon
people to say have copies updated multiple locations
we actually drop down a little bit and consider what you have to do cheap that
its somewhat more complicated than it looks first glance
so let's take a simple example up building systems
in aum's on which is something nice many other people and
in this call have familiarity with now
amazon has a model where their data centers are divided into regions
within those regions they have what are called availability salts
some availabilty zone is actually a separate site it's an actual false clear
skill
data center that's within a short distance say
10 or 20 kilometers up the other data centers
in that region so let's consider what happens
if we have a a failure so
one common type affair Amazon that has happened a number of times over the
years
is to have a problem with elastic block store this is your nap
network attached storage when you have a DPS problem
it means that the applications work on that site maybe
either down or at the very least severely compromised
so in that case applications running in a single AZ
will be now Amazon of course recommends that
that you actually spread your applications across
multiple a seasoned single region and that's a
that's a best practice that's followed by many people but in fact
it's not quite as simple as that because just going to
spreading across the AC's doesn't fully protect you from
EPS failures have been a number of cases already over the past two to three years
where the EBS failure actually lead to poisoning
I love arm other Amazon infrastructure services
running in other hi AZ's within the same reach
so at that point if you have the misfortune to hit one of those failures
your applications now could be knocked down across the region
so in that case it's important to have the applications
running in the next region but the way you do that and
and the way that you he set these up is very important
because if the applications are in anyway coupled
back to the real to the applications that are running the first failed region
you could now have degraded performance in those regions as well
so it's very important that as you thinking about this
to not only think about just the distribution a few data and the
management of that data in different locations
but we also want to thank very hard about exactly how we do that
distribution
to eliminate problems like over tight coupling between regions
so this leads to some requirements if you're going to
the old systems that depend on multi data center
my sequel so one really important
a requirement is to be able to handle local high availability
within geographic regions arm for performance reasons
and and you'll see an example of this the tamil in a little bit
it's very important for the arm
for your data to be as close to your applications as possible
that means that if you have a failure of a single bicycle database
where you have to take it offline for for me its you'd like to be able to
replace it
with a copy that's running within that region at the very least
and even better within the same availabilty sup
so that is to you need the ability to have high availability that
the coverage you for failover as well as planned me it's an upgrade
be on that you need the ability to do advanced replication that to link up
the region's and this includes the ability to support flexible topologies
this multiple ways regions can be linked the replication mechanism needs to be
very tolerant of faults as well as family see
so one of the things of course there up
that you see is you never cross winds is that is tremendous variation
even Amazon Manchester CID ok this tremendous variation
network performance you can go from say 100 milliseconds
pink tie between two sites I'll that actually could go considerably higher
if Amazon is having issues which is kinda rare
but more likely because somebody else shared infrastructure is doing something
it's consuming up and
so and then finally you need to have the ability to do transactions
filtering so that you can actually control the types of things that move
between sites
there's a final constraint on this when you're building this and that is that
as you as you start to set up the small TTC systems
it's very common to actually be doing this
with the system it's already operating successfully
within a single hom Amazon
feel guilty so on or at most a single creature
so what you want to be able to do it introduce these changes with minimal
changes to applications
and without needing to migrate your data to new data our database type
so this is sorta overview of why you want to
to go to mall TTC little bit inside the summer the complexities
we see when the system set these systems up
and then what are some other requirements to be able to deploy
successfully
saw this problem so with this
I'd like to jump in and just explain a little bit about how continuing tungsten
fills these gaps we see in terms a
I'm ability to hi to handle these problems
so the tents to
act years fits in with my sequin a very simple way
so we view my sequel is a building block
as you are by by no I ship building my sequel based systems
I'm I it's a very very reliable server
excellent for oil TP has very good scaling
is very well known so the idea is that
on you're going to build systems using bicycle you doing that already
but I'm still really comes and picks up and handles the capabilities
where my sequel saw and these include the following: so
obvious things like an infield single database being able to do automated
feel being able to do transparent maintenance lady wraps
running so you've got 24 by 7 taps being able to take a database
out of service completely transparent without reconfiguring
restarting applications appeal to load
sequel load balance equal transactions crossed it depicts
see got extra copies you want scale create disaster recovery site so this is
beginning to edge into the multi
region capabilities are have the ability to have data
I'm dynamically I
or are replicated will tell you about sites
and then be on that to go to form of the master replication we have sites
operating as peers
and then finally arm and this is a capability what discuss
any further in this on in this webinar be able to replicate the transactions
in real time database types and this is becoming increasingly important because
not only your system's geographically separated
but now it is most large businesses
use more than just bicycle so he may have geographically distributed
instances of my sequel the of feed into a data warehouse
single location so these are the capabilities
tungsten supplies what I'd like to do is 90 yen on the capabilities that are
particularly
relevant for the problem that we're discussing this webinar today
so the site that we're showing right now
gives you the basic tungsten costra architecture as I mentioned
the court this is off the shelf bicycle see use
any bicycle version starting with 50 that pleases you
aren't so for example if you're using pequot night could be per car to 5 hyper
cons
art six whatever which ever built you
you prefer you start with that and then I'll
tense than ads the following to those
on to those databases you have runny first of all we had
ASIC replication sows when we started out of
thus some years ago week built
enterprise quality replication which global transaction a East special
features designed to enable failover
our transaction filtering I'm Palo replication things like that that's then
had a teacher the hosts next to that we have
managers that lasted deep Europe your management so the managers former
group in this week if your cluster you go in there in
and talk to those two do you management and then finally
and this assert the secret sauce in these in these clusters we have
something called the tax
connector which isn't layer in between it's a proxy
understands the bicycle network protocol and
your applications connect to their and it makes this cluster of the shop
databases
look like a single application to users
a single database to user applications
so that's the basic architecture I'm
let me see you in just a little bit on the connectivity because as I've
mentioned this is a
a pretty special ice feature that the gives a lot of the power to this
so touched and connectors mentioned is a as a prop
it's very fast uses Java our sponsors threat reached on
connection coming in so it works very well S&P hardware
are can run anywhere on the network I common way to do is put on Sep message
at servers
but you can also put it behind elastic load balancer
up between each AG proxy hardware load balancers
up by default when transactions come in and people
and applications make connections we brought them to the master
wherever yes but we do have options to enable be great splitting
at one at that features at this is that this
on is that when you when we do recognize a read it is then load balanced across
all available slaves
if no slaves are available Cassat master so this is a pretty powerful feature
um to enable you to get full utilization out of your host
and it also is key to enabling HA
because having indirection layer allows us then
to do transparent reconfigurations
system that the applications now so let me give you an example
let's talk about graceful failover so this is an operation that you wanna do
if you need to me that's I'm you wanna take for example let's just take a
a typical example in in Amazon is you might wanna take a server offline so
that you can recruit pitch in
on a larger instance I sent him aside to get better performance
so this is triggered by Kamiya costs which and basically what it does
years arm the the connectors Hall
the car transaction little crack transactions and
halt of further activity and allow the majors
to reconfigure replication to promote
a new master in this case %ah host db2
on and then they let the connections continue
this is designed in such a way that it's always your data loss
I'll so the we we ensure that the transactions
stopped I'll weekend
said the transaction through to ensure that the slate that were promoting
is fully up to date aren't any other specialists features a
killing thanks to permit Lake committee transactions from come again
up things like that is completely transparent to applications
so that you can do this without having to reconfigure applications to restart
are have broken connections and the of and the reconfiguration which is
under the covers of course feel a complex is complete transpire
use issues single command so this is a very powerful feature
one that's commonly used to do your arm
maintenance within a single region or on a single
built it up a one thing I should point out is that these
these notes can of course be spread across multiple built built it sounds
I'm if you operating used on-premises data centers
take a course span data centers that are short distance par
heart fail over on the other hand are requires we do
an automated failover so this is triggered by automatic rules
the run in the manager's so when we detected a master has failed
which we do by just com saying that it's non-responsive
we go ahead find the most advanced late and
are promoted it's very fast for process crash because thats unambiguous
we know the process is gone so immediately picked slick
wrote that promotion on typically occurs within seconds even for very large
instances
and then if you have a host crash or something dropped
of the network that can be a to a minute
arm you do require admin for recovery don't automatically recover that we do
expect that somebody
will go look at the system and decided needs to be recovered
that can be done with single Camii
so is another case that we cover which is a heart failure rate
where you not only have arm a single note fail but you actually have multiple
neck feels
up here notes feel and in this case it actually leads to
on los so like a lot of a high availability software times
does implement or between managers on
the managers maintain the cluster we expect to have a majority in order to
come
continue so when you get a case where for example you get
are failures with multiple managers so that the one surviving manager
for example here at node up can only see itself it doesn't have a majority
the entire cost a cluster stops and once again this feature
works across the Amazon availability sons but we do not meet in Coram
but we do not recommend trying to be this across regions are you tend to lose
or because a a variation where network connectivity
so that covers the basic a high-availability features
within a single region and that's that's what other
sort the Lex on what your gonna build these systems are supported systems
right now like to do is talk a little bit more about that the multi-site
features
and for this I'm going to go ahead and
like give you a little bit of an explanation how replication works
because
that the point to point replication constrain portent here
so she may be familiar one of the I'm
key things that continue offer assistance to replicator it on
coda Google dot com are it most transactions
between point A to point B typically by sequel I'll
databases but it also can extract or call arm
and you can basically set up any
topology you what it knows how to replicate between clusters
the replicators have a kinda interesting property that they can run multiple
replication services within a so a single process
I we have global transaction nineties transaction filter install cancer
thanks expect in a in enterprise replicate early colgate
here's just a simple example that illustrate some the flexibility
we have two clusters one Amazon East one named
East one in West the are we wanna
I poll transactions from this are from these clusters
and are fan them into a single database which perhaps is used reporting
so we have a East service on that replicator pulling from
East slave we have a West service
pulling from the West master of course could be configured differently input
from different
up replicators and there's a number if useful features like this
but I think this illustrates that the basic flexibility
you have to work with so here's how we'll use this for multi-site
arms computing so here's an example
a for multi master what we do here is set up the east and west
on I
clusters on with their own application stacks talking directly to them
and we set up extra replication services that basic he
just link between the arm the east and west locations
so these operate outsider clusters and
are completely independent up so
if in just for example if you commit a transaction in West it goes into the
West master
its replicated within the cluster to the sleigh meanwhile
West service over on the Eastside
pick set up and then its supplied both
rap it's there so this is normal operation
let's just take a couple examples of failures let's sled
one of the what are the failure one other replicators is down for
maintenance or failed
on the east side that's fine replication continues to that replicator gets back
in which case it mom it it
it that continues replicating to the instance that it's pointed out
there's no reconfiguration system this 30 handles
similarly if we have a local failure on West
the master fails and we then promote our slave
I area its I again
no on there's no reconfiguration necessary in a cross-site
replication it handles it automatically
and in fact arm you can go to all the way to a full site feel so what's at the
West region goes down completely
in that case what happens is replica patient just stops between sites
so the West replicators that are are over on the East
site ever trying to pull thanks from the West region
just don't synchronizing owed and begin to
I simply poll I on
and go ahead and and not just basically
remain in that mode until the web site comes back similarly
East replicators will either stop if they're in within the region
they may fail in just have to be restarted but there's no reconfiguration
necessary
once you get West back up and running again he just restart everything
your system begins to run again
so let me talk to you %uh a little bit about another
topology you can use for multi-site are multi data center
bicycle and as will discuss
shortly bicycle doesn't work for every single application
in some cases applications chignon
multi masters at work for every single application in some cases you still have
to use basic
master/slave in this is actually built into the cluster
to what we call composite a decir its and in this case
the clusters on each site themselves have a master-slave
replication area relationship to each other and connectors
move the on I with the data
from one site to the other arm so there's a very interesting feature here
which yes that in this particular configuration West
is the is the primary and we call that master that has the primary master
arm the East is the sleigh so it has to serve a master has what we call a relay
and the conductivity in the in the Transit Connect ur knows where the
primary master is
so if it sees a connection coming into right what automatically forwarded
out to the west so this is needs
you don't have to worry about reconfiguring applications
plus you also have the ability to continue to do reads on whichever sites
closest
lets show an example the failover sleep now to get no configuration
necessary it's handled within the cluster on
the West master fails we promote slave we automatically
ensure that the replication Lake continues to the
arm to the east relay moreover
the arm the the at the application connectivity
is then seamlessly redirected to the new master the west side
similarly if we have a full site failure where
the west side goes completely dark arm
you can then go ahead and using a simple command do a manual failover
to the slave site I'm and then basically all your applications
will our point over there this does have some consequences for
application performance will discuss those shortly
at at this point I think that we're
ready to go ahead and do a demonstration these capabilities
arm so very quickly
Peter are are you ready to go
left I am headgear hear me yes i cud
just a slight shown which arm do you wanna just
I describe the two parties up front a couple pictures up them
if you'd like or we can just sold over to your
I'll to demo okay so first female
examined the the topology on the left hand side
and reviewer make inhibitor make nasty figures in US used one which would
ultimately
donde comprar deeper center
and after that we were reconfigured the whole cluster to the second case
radziwill have an application
in the in the EU using this Craftster
and biggest simulated the switchover
from PU to US trash Coast
and you will see how bad it works let's see
okay soul lackluster
let me go ahead and make you presenter Peter and let's get going
just fine in the list here here we go
I'll let you know when I see your slides
kid you should see my skin
I did of you go
location so re have
a bunch of for much chance to your
Ralph for $10 each region
sry forms across Thurso PW
stands for thanks from beginner and
US their stance for the united states and
use for the east soul USC want to speed up the process for nodes you don't want
to
be across the notes you SW want to stay across the notes
and industry benchmark in Old image region and able
put on the application which in this case usespans
so that's examiner the current situation a bit
so like with any other
the continent punks *** cluster
the dinner reach 100 can use the she should clear their utility
to to check the cluster status
what it's probably for morning class thing
and by interesting game in her
it wasn't in previous babe in arms
is
the open-top more market wrap Cpl and if you remember
ran Robert talked about to
that is the deputy editor for each possible connection and it doesn't
matter
bust meant that if Sandra one of and shell
you have descript to have a nice the
but I just rebuilt a of
what this happen and perhaps freakish /url
disputed it actually has the the same output in order notes
so you can examine the you can
you can examine the you can be examined the spirit
regardless booknotes you are running this
and wraps check the benchmark same Roxy for some small script
first two years as big a person it reached just performed his bench
in this case from the East Coast benchmarking note to the east coast rust
and let's start that
that's right away just to have some idea about
how can that cross the perform
an we can ship it's the wrong granted a certificate to transactions a second
and I have a script called SB
limited doctor
the same church is different in a way that Miller running
distributed transaction date and
this means dad
each second we will generate can you transactions
so this doesn't mean that people do tend to transactions every second it means
that he will generate stand you
and if I start this is
it will start to doing those transactions
and if I start additionally the jump
then in every turn
this transaction we r this transaction mister
and they are replicated to have the other teacher
for in order to to do this we have
three ski mountain as detest
USC as because through a stop as you can this be
you so each region is replicated to each region
so you will see you see how how it
works and right now what
I want to examine is that in the East Coast
I want to do it feel better right now
I must it is the the first known in the East Coast
let's reach to the second one made the switch from
you wanna see to
and it starts to do the switching but I'm more curious
about the benchmarking node so you can see that
that around 68 seconds
transactions were stopped and pierced piercing 0 transactions
and thinker look at the Q lacks the QNX is growing
will be a candidate you can transactions that district on and lost the failover
completed
so here this year 0 transactions
then 10 transactions but with a few of 100
20 bashers benchmarks on the Q
and it works it down to cuba ranked
0 and the failover started their own
68 seconds so a graceful feel over look like
sirkin ish seconds sold not
much and everything goes on
for like previous three
the a bit more matched your keys if
if also
let's see if the use to it from October now and really wish to use the master
lacks do marker
sure
to crash test from thank you should always do it like this
because dishes practically manual kernel panic on
on Linux and action express then turned but you didn't see it because
because of the can occur in a panic and if I checked the benchmarking know the
damn
transection stopped at 150 six seconds and just start to generate a queue
in this case it to a peek a little bit longer and that kind of makes sense
because
bid don't know okay to it looks like to cook a little bit tune
a I have to
Jim in this case you all and fatal what happened and it works from the queue
so suspense didn't have a a big enough time or the film professor on
searched the 42nd so it's a under a minute
it didn't have a big enough time
and when you are building currently available applications like this
the point of this travel what you know you should take home
is that a.m. should expect
that for some transient amount of time
maybe there'd be a bit no data available and in that case
you have to do cewek some and in this case cue
what's possible biggest banks then connecting to server behaving
application
and chairs benches assembly behaving application
because here for because it's cute but
your the time will tell santa wasn't enough
and if she's better if you check the state now
the first noticed the hospital damn and
the second node in the East Coast shot which means that the administrator
student coveted an thats
the thats be even let's display from
more stuff defied
destroy this than I expect
the benchmark to
stopped completely and if a check your
she stop not to
not available and this is not promoted
to master and actually so honorably affirmative
for which should be shown and i'm looking deductions bench
see here you can see cuing and
so goes sir East Coast known mold
the shut itself down because it most courtroom
and and that's the big a beer
here expected your so let me
let me stop the benchmarks and
that's used I
to examine what happens if you want to do
if you don't want to mark the master because in case of multinational you saw
that
the Denon *** Tom Brady that there have been all
conflict between data centers 100 guaranteed that in a way that we use
different schemas indifferent did
did the center's it doesn't matter you know
how you guarantee it can be based on
what the court seabright toward Bart users
you don't to prather did
a synchronous replication and market muster conflict sir
doesn't the back too much in better
accident Libyan are so for jazz after describing
your that topic so lacks
reconfigured the class for and what i want. cool meant
show you years that you can configure
the whole thing through a single I and I find
richer is very similar to might not see enough
so I defined a big but sources
and I define all the possible cause actions
for the applicators for to mark the site apology
on the benchmarking node it's practically the same
X
back bad in hearing in the East Coast
I have an additional connector for benchmarking and if I find your
bpm you start your it will only install the connector
because your current know the cluster topology and its Renault
that this post this on their contractors sold this
a you I night I'll approach is ready
automation friend so long and I
keypad didn't you line I tried riches
back to conduct an ad that compost it which has a competitive
the source so we perhaps the US the master
and you have rash the
rest for *** as the failure oversight so
I've ago had and
the Configure talkback by this
and
you know I'm with pink the the new configuration file in place
and that Shane now I'm playing your menu will pop up the minimum annual
sheftel I am and I also want to do
this long on the bench
King good and now that
the new configuration is is unclear if all I have to do
you should Ron from the stitching that victory
bpm useful
trimmings for the new configuration
and I can
on it on every node even in Prato
it doesn't matter so every node is independent
so you them if you are a
if you want to back to Shafter superior on a proper
manifest for this you don't have to catapult
Beach ordered your you do this
or anything wants all the nodes for
art on the cross for trivia complete so
the corner
installed the connector is complaining that I don't have extra copies from
its kind of that because extra backup is awesome but
but you don't really need bet on the applications
rare so its okay and that's
examine democracy
contribution so if I take a look at
you did the source their mascot is online
and detest lives but the fact change
to to the best data source
a chance US West Coast yet this Talkster the rascals contractors
it would be in smarter to do this from arrest course
but I want to show the picture
that it works from here as well
and it has ever been in the online which means that
for the chem speech over this
and also be never quite possibly
source code record where do you see that
you is the monster and vast
the sleeve and that's
examine the benchmark a bit too damn
divers part
same suspension limited BL young starting gold clampdown
transactions sec crown and now I'm going
to go here and issued a stretch to pass common
and this geographical stretching this possible because of the car expect on a
act and putting big black *** Baddour
a proxy which is the actually capable of talking
d'amato Demeyer secure a buyer protocol
so we don't need hipster this ritual will not be fine for the connectors
this ritual work today to the other the center and
feel examine it is it in a similar vein
the transactions per second a stroke 20
aunt ritual but is done but he missed
see steer shoulda cue blank check using
and their curriculum movie from this screen
and my question is did the field over *** hand
icon if you think that the figure would rip your disorder
Q which still fear generated in suspense
okay three people banking for people
that the field over 5
okay 6 okay 7
okay so
believe you're not convinced the by the fact that
the Q is growing but you're actually doing transactions
see that you're doing street transactions done in the next second we
have drinks here
and if you take a look at the evidence response time of a transaction it's
almost a minute
this is a default suspense transaction which contains
lacquer planted
of different types and the two or three or four tracks so
it's a lot of grown clips and the limitation br PP
yep you think your disrupting the cluster server which in itself
the limitation you're hitting cured is the science of the planet and the speed
of light
because we are trying to use a Best Coast databases
from the EU rid Servite order on trip times and more
and witness demonstration I wanted to
wanted to show to you that despite that term
you are capable of
BRB cross continent
across continents cluster which real
be transferred into your application you probably don't want to do it
at this not could cross continent this your beeper factors under more
close to each other this is not to thus it's not
that much finished and if returned a benchmark damn you receive and the
switchover happens
because suddenly the latency will be OK again
okay and did started to do 50 transactions per second and
big delicious the number the latency
started to the keys and it started to the
is because the transactions we're currently processing were
rethink in the queue for the a hundred and
50 seconds and now it is
sketching up and
i'd give it back to robert thank
was a great day
up so let me see if I can myself up in the last dance
switchover okay
let me know
when you can see my screen up
to tell Microsoft switchgear are you able to see the demo architecture so I
peter
I steer this is at least a
up okay great thank you on good
so we're almost done here we have a number questions coming up I don't wanna
hold back on those
let me just talk about a couple of to practical items
up before we conclude arm so
the first thing I think peer are really showed this
are very well so we don't have to harp on it too much but
the the the tungsten installation is done through
a a procedure called TPM sense for tungsten package
I'm manager and this is a really critical feature for setting up these
large-scale systems and the key here
his you develop that tungsten any file the Peter showed it's very simple
I'll you then download the software any just 20 p.m. install
and it its run on this it works on the same principle is something like sheffer
or puppet where you have a specification that
tells what that machine should be doing and TPM will
will read that any file and do correct configuration so that it takes up the
role that specified
in that up be a connector your a master
slave whatever and that's how you that's how you get these systems
are set up very efficiently moreover it allows you to do things like
upgrades automatically I you know you can just change
for example you can change the any file and then you can run a commerical TPM
update which will then make appropriate configuration changes
on each host so it's a very powerful feature
that allows you to have the same single file describes the overall system that
you want to build
and that have the right things happen each host that level of automation is
really critical to get systems
set up the other thing I'd like to talk about is
is perhaps a more interesting issue and that is deciding
when to go with the composite data service versus multi master
and so the composite data services have is you seen a number
a big benefits they don't generate conflicts and this is just master/slave
you only a
the reason that happens to you only update on one place as a result
you don't have to change your applications just about every
application
that runs with my sequel to also go on with my sequel master/slave with
master-slave replication as Peter shaky the failover is transparent
arm so you can in the case about
a heart fail over you you get broken connections thats
not really easy that's not really possible to avoid but
I am your applications don't have to be reconfigured they don't have to be
restarted as long as they know how to retry
a transaction they will be able to to
handle any kind reconfiguration and it really it works very well
statement replication an that's a big benefit because state replication
is quite efficient particularly if you do large transactions
that affect a lot rose thanks to consider
well the first thing is the highly see writes the primary master is Peter says
if your spray if you're going between from from Ireland
the west coast which is the devil you just saw that's not good for
applications to have to have take that hot
when they're talking between application between applications
arm between application database it follows from that
that when you do a switch which tungsten
reduces to a single like command that's the easy part
you actually may need to switch the locus if your application activity
over to that side as well that's the part the that we'll handle it
involves reconfiguring DNS records
activating applications so that they begin you know maybe firing up
additional
ap capacity on your Nutr site and as a final issue which is that
that having this type of a bomb of configuration
does introduce coupling which can
be a problem if you have flaky network connectivity between the sites
so as we looked at the arm he does we looked at in the AM a psyche case
you don't wanna do things that introduce unnecessary coupling so it's something
to think about
arm you know when you when you deploy this
that said this is the most common way to spread it across sites by far
I'll the we dealt with many many many applications and the reason is
requires no app changes on the other hand
multi master this is really where we all wanna be I
I think that if if we all had our druthers be great to run completely
independent applications to completely independent
databases and have them magically reconcile
each other and the reasons are obvious there's there's no coupling between
sites
that's a very very powerful benefit for for high availability
there's no fail if things stop and this isn't in fact in
in in failures up entire site is pretty rare delusional
data but even with with it whether with Hamas are on premises
it's very common to have the other inaccessible for
a day or more that's just not a common feature so
you just get the site fixed when it comes back everybody exchanges
transactions that were missed because they're stored it locks
can you back in you back in business your
no changes require and then of course them the right waits is a huge benefit
the thing about multi master yes the there are some things you have to
consider and you have to test carefully in this is true for every single multi
master solution
is obviously potential for conflicts arm
tungsten does not magically resolve conflicts it gives you a lot of ways to
avoid them and make them less likely
we can do this to filtering but you definitely have to look for conflicts at
the site had managed
as a result it's quite common to arm
require ap changes and it happens that
also with multi master just to avoid a drift
we recommend using replication is just gets much better results
because there's there's no ambiguity arm
about worrisome causes things like that that
that can cause data drift in multi master situations
that said it's actually surprising how many applications
do work well with multi master just give you three examples
applications primarily at data you know sir test results
or or something like that they work very well multi master
another thing that tends to be very helpful is if your application is
largely or entirely based on
autoincrement keys and doesn't use I for example natural keys
which have the possibility of them duplicating across sites
and finally arm you a sass applications
tend to support multi master quite well
in many cases in the reason is that many sass applications
have their customers primarily in one region
therefore you can assign a particular customer a single region happened do all
their rights their
arm and help other customers and other regions in this case is exactly the case
that
its exact situation Peter just demonstrated which
yes that arm that there's just no conflicts because the applications are
ready two different schemes
so with the hat I'm will leave you with
I Justin link somewhere today get information
art cluster replication documentation is
available on the Continue website our software
%uh the replicator is out on Google Code UCC Google to calm the clustering
software
is from downloads and then of course check out the common
continue web sites is a wealth of information including blocks
up and all kinds of other I'm consulting other things like that
that can help you so
with that holy site turn it back t you i think we're ready to answer some
questions
great thank you test and Matt 222 invite people to continue to after questions
and the guy web in our quest to spell well take as many as we can in the next
few minutes
and like I said before if we don't get you don't despair
we will get to you after the webinar through biker specter reaching out
directly
alright cell guys ready for Special Care cast
what are the benefits up tungsten in comparing with
Calera cluster she compare the two
Peter you wanna go for that and all of the law
okay so the
to all things are completely different
soul it for
it's like a app Apple Store interest compared to standard application
works for absolutely
absolutely differently in so
under different circumstances you will have
you will have different benefit so
that civilian the in thanks fam
you have to avoid conflict your sound for
configure a conflict avoidance
at the application level to make sure that to
conflicting data is not replicated between sites
bout but it doesn't clear and q2
to write better to err to aside which doesn't belong there
in there extra be clustered the sisters old
the read the certification process so if you have a marked a master conflict
you end up having to certificates of India richer surface transection
riches are all back our commitment but unification still has to handle
throwback and also the if you have a mock the master
conflict drawback is really a class from a petition
it involves hope being pregnant for from undue
so if you have a lot of conflict anyways
which means that you will have trouble
the you won't have the same kind of trouble with extra to be faster but you
will
job have kind of trouble because your book have
pretty bad performance investors 30
work but this river Kira
it ever created minestrone but you know the the total products are like
completely different practically so that application is through that applications
make did it's not the same burden on the same
yen and I just add to that the that I wanna Peter to go first because he's
he sorta looking at this and has doesn't have so the natural vendor bias
a untouched and that idea but I think the two things that really stand out for
me
are first of all because we're based on a sec
master/slave underneath we can handle very
lobby workloads and this is something that we deal with our customers people
doing
huge transactions that I'll you know maybe
for example in Clare would end up blocking the queue for substantially
time I'll those are handled quite well
up by tungsten I am you know we can also handle
the on my *** answer all this dirty stuff that you tend to find them
in real applications I master/slave and hence tungsten Hills
very well the second thing that tungsten does
is that it allows you to build systems that are fully
couple between sites in the case the Calera
you can't bill truly independent systems using glare ample
replication is cool air every time you commit art recommend one location it's
gonna wanna go ahead certify with all the members
so the arm tungsten allows you to build these systems which
I'm are completely independent and where you could cut the links between them
for you in a couple days and then just lessons links to restore the
the joint Cup at work so spear says it's definitely it's definitely apples to
oranges and there's
there's there's some fairly big I'm differences in the benefits
these offer okay great
and several people have asked not the questions and I'm going to try to get
and all one question at least per person in there
it's the top of the hour Robert D and Peter to have a couple extra minutes me
to ask a few more questions
absolutely okay and audiences you have to audience if you have to leave
please know that I will send you a week warning %uh this webinar
I'm so you won't miss a thing and they'll be a follow up plot per se you
can ask additional questions once that is
up and running to alright so on to the next question
if we want to replicate from multiple masters 21 server hardware
for example fannon which oppressed purchase the best
using text text in fannon or native MySQL replication
with multiple texts instances under different port
on the same hardware
oh that's a Saudi Lisa could you repeat that one so that question it can
at because it's a complex question yeah I'm in a that can pop up the questions
too
take a look am alright so it is if we want to replicate
from multiple masters 21 server hardware
which fannon which approaches past using tanks 10 with fan and or native
MySQL replication with multiple instances under get reports on the same
server
arson car hire okay I got your text well
I think a one thing I should point out is if you're running I am
you know any the production versions a might seek
fannon isn't supported I so fannon
I Peter correct me if I'm wrong that's gonna come out Anna Maria
arm ready be is going to
support this I don't know 0.56 supports I think it's coming out in 57
so there's but one cop his choice that if you want to work on the service you
have in production now times that she only choice
I'll I think any that any other I've seen a few things discussed before that
involved
are you know kinda clergy switching between I am
you know between mast reports and they just don't work
I'm you definitely want you something like constant to do it
AP did it have anything at onto that
now so if you have fannon approach
and man I have TB support for
the that know you were able to join
and between your different sources speeches
good which is something you have to do with their application sure using
multiple instances
hope question is do you want to have apps for
what is easier for you if you don't want to have it but this beautiful you're
managing
and in on managing mop instance some
okay and next question if multi-site
cleft not master is a requirement
in what circumstances in Cinci is continuing text into better solution
in what circumstances something like pecan
extra TV cluster better solution
out ok of first on this one so I think that
so what are the things that i think is a big benefit of
the extra TP cluster is it doesn't let you data getting consistent in this just
applications that that need to be
that cannot permit data to become inconsistent and one
copies of data in multiple places what in that case
glare or for comedy be extra actually be clusters your only solution
on the other hand if you're army if you're going over geographic distance
and what you're looking for is being able to deal with Lucky workloads are
high-performance or being able to do things that really involves
complete decoupling up the at the servers and I think tungsten
better solution at that would be sort my take on
and hear what you have to say about that came so the
my take on this is better
my first question *** came to my mind just do you have an hour or more
barter but turn
what it boils down to is is your application
good it is synchronous replication or or better
with synchronous replication so the answer to that the
did sturdy very application dependent
what district loaded more to try a truck stop daughter
protracted for example for
you know gramma Matt thinks stop now
so and the denounced what is it depends on the application
but you know the corner concert for nineteen ninety percent of their
complex data distributed questions is overrated bans
so this is one of them but Robert federated
some guy I agree completely it really does and
and will do that will do that our talk on some other occasion
her thank you guys alright I think I'll take am
to my questions and then the the rest of them will be answered
as I said before after work alright
Seth let's see here and is a possible in a master-slave set up to automatically
switch
reads to the master when the slain slacking
yes that's a in fact that a great connectivity feature rate
unfortunately like like the difference between this an extra TV cluster we
couldn't show everything
there's by there's a level of service where you can specify
I what's the maximum latency tolerate on a slave lets the set the sixty seconds
will either find your slave that's within that late see when you when you
form your connection
if we can find you go back to the master same thing applies if
if no slaves are available because they all failed err
down for maintenance you automatically go to the master
okay on and then the last question and for now
is 12 attempts to connect your work across hybrid database
replication well work to replicate between or call
in MySQL
I love yeah that's an interesting question any the answer is No
and arm I packages put in a quick plug here for open source
we can't really did there's really two questions under the air
I am what is I I think years
can you use the connector to have arm it
to have transactions automatically sent to two places perhaps different
databases
are in in that case the answer is No but it's because
that's just a very difficult system to get to work we have previous product
that it that it didn't work right
the other question now is whether the connector in general work with
the arm with Oracle the answer is No
and that's because Oracle has a proprietary wire protocol
the connector magic that you were seeing today depends on us being able to
inspect the wire protocol
and arm and for example correctly identify a transactional boundary so
that we can all
connections reform connections and other locations
up that kinda magic is very difficult to do
on Oracle and in fact which find if we were sitting there in the middle
a you would probably not be able to: local support as a pilot's that you
later
to look at the protocol so for now your best bet is to use open source database
okay a.m. Peter did you have anything to add to that
to that no I have Mar think ladno looking
through the question sure we had a question that can you correct robber
that it is possible to write to have been across the but it's not the entire
09 so while mark roberts talk about what we talk about
is that if you happen with failure backs %uh
blog on till her in Quetta for numbers for
most similar on feared that no she's the suspect and mot
so we didn't talk about that you when you have an ulterior tractor
stopped completely on company thirty until you
recovered but rather until it reaches
suspect I'm alt I just and its toll
last night just wanted 20 gratified that on okay
thanks Peter okay perfect ok everybody thank you so much
and brother and Peter curry top have lotsa good feedback here
i dnt thank you so very much for your time today
said everything we're getting email on next 48 hours with
E link to the recording enter the slides
to be sure to read take a look at that a shirt with here and coworkers sir piers
I'm just want to remind everyone that the Connell I
if the next for kinda lives in Santa Clara April 134 and you'll be able to
meet Breivik
Peter and as well what people for Canada
continue its tail it's good chance for you to Palo them and ask them additional
questions in person
I'll be there too see if hi to me and with that said
thank you everyone again have a wonderful evening
afternoon or morning depending on where you're calling in 10 min we hope to see
you again and future kept webinars
everyone have a wonderful time