►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Provisioning the Distributed Database TiDB on Kubernetes - Shuan Deng, PingCAP
I will give a short introduction for TiDB, and then how we manage this distributed database on Kubernetes. For SQL database, local SSD disk is preferred or required. I'll present how we extend Kubernetes to support local persistent volume for TiDB. And how we use the operator pattern for complicated stateful applications like distributed database.
To learn more click here: https://sched.co/FuKf
A
A
I'm
I
will
do
give
the
first,
since
some
some
of
you
may
not
often
be
firm
in
away
so
tidy
our
do
a
short
introduction
about
tidy
and
then
about
how
to
talk
about
why
we
combine
tidy
and
communities
together
to
then
and
nasta.
We
I
will
talk
about
how
we
make
it
possible
to
provision
such
a
complicated
distributed
database
and
kubernetes.
A
Okay,
first
I
will
be
the
heavy
architecture
tile
V
is
the
components
is
composed
of
three
core
components:
the
PD
the
placement
driver,
it
man,
it
stores
the
clusters
metadata
and
it
way
it
has
the
card
view
of
the
whole
cluster.
They
can
it
slow,
the
the
hot
the
how
to
schedule
at
eight
hub
between
the
different
between
nose
and
and
the
taka
biggest
store,
the
cleave,
a
storage
engine
it
is
based,
it
uses,
rocks
TV,
Internet
and
without
their
rocks.
A
Tv
is
written
by
a
facebook
engineers,
and
but
it
is,
it
is
high
performance,
go
machine,
kV,
storage,
the
engine
we
pack
away
is
tacky.
We
combines
the
rocks
to
be
and
the
rough
consensus
algorithm
it's
a
little
like
easy
city,
but
tacky
way
uses
much
ear.
After
glue,
so
it
can
store
large
amounts
of
data
and
the
the
taiga
B
is.
This
is
a
distributed.
A
Keep
keep
keep,
are
you
in
storage
and
you
enter
now.
It
is
hosted
by
the
scenes.
Ef
and
the
tiny
component
is
the
circular
secure
layer
of
the
whole
database.
It
it's
fully.
It's
almost
a
fully
compatible
with
my
secure
protocol.
Users
can
use
my
single
drivers
to
connect
to
to
talk
to
Atari
B
sameness
me
and
for
the
for
the
they're
tied
Ibiza,
its
hybrid,
a
transaction
or
analyzing
database,
the
HTTP,
but
it
also
for
the
AP
party.
A
A
that
the
around
we
have
the
we
have
used,
the
open
source
monitoring
and
locking
systems
like
Ramona,
and
the
promise
for
monetary
and
the
elastics
elastic
a
stash
stack
for
logging,
the
that's
the
tiny
hole
cluster
for
no
tabbies
since
tell
me
it's
not
alone,
no,
not
a
standalone
solution,
users
most
of
our
users
and
have
already
have
my
store
their
data,
my
seho
and
for
but
my
signal
is,
is
a
single
machine
single
order
tell
base
it
for
a
large
data.
It
we
have
most
of
used
most
there
are.
A
A
A
A
The
we
all
know
that
kubernetes
has
become
the
de-facto
of
de-facto
container.
Our
sewage
system.
Actually
kubernetes,
we
can
think
of
a
covenant,
is
as
disability
distributed
cloud
provider,
agnostic
operating
system,
for
example.
We
can
think
of
we
can
we
know
that
communities
has
supports
their
device
plugins.
So
the.
A
A
We
user
can
write
drivers
directory
for
kubernetes
is
not
just
a--
for
the
traditional
operating
system
and
for
for
container
or
pod.
We
can
think
of
it.
A
comparability
to
with
the
traditional
operating
system
process,
or
rather
and
for
the
dr.
Matthias.
It's
just
like
the
traditional
executable
artifacts
and
the
same
pod
and
poured
on
containers
are
just
as
no
single
process
or
single
process
or
Matteson
or
stress.
If
the
we
are
stalled
by,
if
the
stop
about
any
reasons,
it
will
not
restart
it
or
recreate
it
with
deployment
instead.
A
A
Okay,
kubernetes
also
has
at
first
the
communities
is
a
deployment
is
state
of
stateless
up,
Natalie's
application
deployment,
deployment,
a
platform,
and
now
nowadays
the
kubernetes
is
becoming,
has
more
has
much
powerful
extensively.
It
has
the
standard,
interface
and
I
could
seeing
I
CRI
CSI
and
etc.
For
the
scheduler
we
have,
we
can
also
extend
their
schedule
via
user,
can
specify
their
own
schedule
or
just
write
a
scheduling
center
to
extend
that
the
buting
scheduler
and
for
the
controller
manager
we
have.
A
We
can
write
the
CRT
controllers
to
manage
our
own
customized
resources
and
for
API
server.
There's
aggregate
API
server
to
extend
the
the
API
server.
The
cooperate
also
has
a
virtual
community
which
can
use
apply
the
we,
which
is
the
serve
service
infrastructure
and
and
for
the
for
the
club
provider.
You
know
we
have
the
Justin
need
to
implement
the
loadable
answer
or
presentable
interfaces,
so
users
can
use
this
across
all
cloud.
A
How
do
we
entirely
operator
is
the
tour
we
is
the
tour
connect,
Kylie
and
kubernetes
it
it's
open
source.
You
can
check
out
a
here
and
it
has
the
following
features.
It
can
manage
multiple
podcasters.
In
fact,
we
can
in
fact
our
Enterprise
Edition.
We
use
title
reader
to
manage
e
multi-tenant
of
time
crossers,
so
users
can
just
deploy
one
of
one
hardware
reader
or
to
manatee
tank
roster
for
different
users
and.
A
A
A
A
Here's
the
proof
in
architecture
of
that
we
are
either
use
user,
suggest,
create
how
much
art
and
for
in
time
being
criminalized,
then
hiding
orbital.
Well,
what
should
be
our
customer
resource
tab
across
er
for
manager
for
creating
or
the
related
resources
related
the
resources
like
safe,
a
set
of
Harvey
and
the
deployment
of
the
monitoring
system.
A
A
Usually
the
new
stay
for
application
itself
was
in
a
start-up.
You
know
it's
complicated
that
you
different
the
different
graphic
artists
start
with
different
arguments,
and
so
users
have
to
do
some
tricks
to
some
hack
to
make
the
application
to
run,
to
prescribe
correctly
and
the
pass
CUNY
most
applications,
the
most
stable
applications
that
endures
Kansas
Gil
correctly
and
with
the
slave
Isetta.
A
Some
someone
load
the
last
yearÃs
researcher
operator
to
energy.
The
last
research
service,
also,
the
upgrade
some
some
civil
applications
have.
Some
extra
steps
before
upgrading
is
not
supported.
The
native
natively
in
instead
of
asada
and
the
favored
feature
is
almost
not.
We
are
not
working
for
most
of
the
applications.
A
Then,
how
do
we
implement
this?
The
as
we
all
know,
the
cannoli
operator
or
the
customer
controllers,
the
users
can
just
just
impair
the
domain
operation
logic
in
their
code
that
you
automatically
manages
New
York
customized
the
resources.
There
are
two
kinds
of
extensions
personnel
about
sepia
or
CRE.
Tv
are
now
it's
duplicated
and
Ciardi
customer
resource
definitions.
Users
can
just
create
a
customer
resources,
just
like
their
beauty
in
deployment
or
self
assess,
and
users
can
write
the
domain
operation
logic
in
the
CRT
controller.
A
A
It
has
a
meaty
rapke
and
resources
limit
or
request
there
and
other
fails
and
the
container
just
sink
the
when
their
controller
or
what
is
the
spec
changes
either
we
are
to
the
actual
operation
to
keep
the
cluster.
If
looking
exactly.
What's
like
the
spec,
it
compares
the
status
and
the
spec
and
gradually
reconcile
the
resources
to
make
them
match.
A
A
First,
the
community
stands
in
a
support:
local
TV,
a
local
person
volume.
It
uses
ladder
worker
personal
Network.
We
know
that
a
not
the
world
persons
volume
has
has
wrapped
has
a
replica
each
date,
amazing
in
store
the
several
wrapped
castle.
With
with
when
combined
the
tacitly,
with
Network
her
network,
a
TV,
the
air
as
there
as
extra
overhead
for
the
theta
for
the
IO,
and
here
we
consider
the
problems
with
local
TB.
A
If
we
want
to
maintain
or
the
Loden
know
to
be
just
as
fails,
how
well
we
do
with
no
copay
for
city
for
sidwa
cetera,
as
for
deployment
at
the
know,
when
the
load
of
fields
in
the
deployment
controller
is
just
crazy,
another
pod
on
other
node
and
also
for
the
no
for
the
network
giving
and
we
can't
just
create
critter
and
not.
We
can
just
stop
the
node
and
create
another
pod
and
then
pound
their
network
give
a
to
another
node,
but
for
local
people.
A
We
can't
just
do
like
the
answer,
so
we
we
just
create
another.
We
just
increased
the
stables
at
replicas
and
to
make
it
an
act
just
for
so
the
stiffest
at
a
controller
will
create
another
tacitly
pod
and
the
take
away
is
a
self
can
in
turn
a
internally
sync
synchronized,
the
replicas
to
the
new
newly
created
potter.
A
So,
but
if
we
and
deployed
more
than
three
tackles,
we
can
actually
have
this.
This
is
acceptable
since
tight
tackle
yourself
has
labeled
just
like
criminalists
label
it
can.
Users
can
attach
physical
topology
in
neighbor
to
the
attack
of
a
store.
So
we
for
this
to
tackle
base.
We
it
has
no
de
a
label,
and
it
has
a
lot
of
is
notice
a
when
when
PD
is
when
fitness
schedule
date
has
the
same,
the
same
rapper
castle
with
a
not
scheduled
around
the
same
no
same
notice,
the
attack
of
a
powder.
A
A
We
we
implement
this
schedule
extender
to
to
such
as
advanced
tournament.
This
is
not.
This
is
impossible
with
the
community's
beauty
in
scheduling
policy
and
there's
another
problems
we
encountered
during
our
deployment
nest.
Considerations
such
the
following
scenarios:
if
there
are
the
the
target
that,
if
let's
assume
that,
if
the
node
has
the
tensile
term
memory
and
ten
storage,
ten
local
storages,
if
we
and
the
one
of
it
in
the
one
of
the
fattest
clusters,
a
tacky
way,
consumers
afford
and
secure
a
memory
and
for
storage
when
they
tackle
a
is
upgrading.
A
The
stiffest
said
will
tune
in
the
attack
of
its
Rapada.
Then
the
resources,
then,
when
the
lag
step
or
the
schedule
in
the
resources
would
be
become
1070
memory,
so
the
not
earn
their
Mellon
might
be
a
chance
that
another
attack,
another
class
or
strike
away
or
other
reason
podo
would
with
a
schedule
and
in
this
mode
because
it
can
still
assign
just
still
has
eight
CPU
and
five
memory.
A
Six
storage
on
this
load,
but
when
the
newly
created
a
high
kv
pod
should
either
since
either
has
the
local
PD
bound
on
this
node
in
the
master.
Be
scheduled
on
this
node
then
so,
but
it
doesn't
has
the
required
resources,
so
this
jacobean
would
be
always
pending.
How
do
we
solve
this?
We
have
advanced
the
scheduling
based
considering
the
virtual
resources
for
this.
A
Rate
we,
we
would
not
wait
with
a
reserve,
at
least
the
force
a
few
and
a
three
memories
for
the
when
other
port
is
the
schedule
on
this
load,
so
the
there
and
they're
still
have
four
CP
1
a
3
memories
occupied.
We
when
the
attack
heavy
another.
Another
Porter
is
scheduling
on
this,
so
we,
which
should
only
say
six,
six
appeals
and
seven
memory,
so
we
we
can't
schedule
this
other
other
pod.
A
On
this
note,
so
this
this
is
actually
because
the
local
CIA
and
the
kubernetes
a
beaut
beauty
index
beauty
instead
of
ascites
many
for
the
its
first.
What's
the
first
designed
for
not
whoever
key
bay,
so
it
doesn't
Atkins
consider
such
a
scenarios.
So
there
still
has
many
scheduling
problems
with
this
with
no
coffee
bay,
and
we
do
massive
effort
on
this.