►
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
Today, Kubernetes is the defacto standard if you want to run container workloads in a production environment, though that wasn’t always the case.
We had have a fair amount of monolithic code for our products and to make sure we could build Central Repository in a way that would be easier to manage at the scale we needed to make changes. For us, that meant changes to the way we write code (more Go), how we build apps (using Docker and Kubernetes), and how we deploy (using Helm). We had to figure out a whole bunch of things as we looked at building proper microservices that we could deploy to Kubernetes.
During the session, we will pay special attention to things like persistence, resource limits, affinity, and probes because containerizing an app is not simply putting your existing app into a docker container.
A
So
I'd
like
to
thank
everybody
who
joined
us
today
and
welcome
to
today's
sincere
webinar
building
a
cougar
and
his
power
central
modules
repository
and
my
name
is
Mauro
I'm
CN,
CF,
ambassador
and
head
of
developer
advocacy
with
Jeff,
Ron
and
I
will
be
moderating
today's
webinar
and
today
with
me,
our
presenter
and
my
colleague
from
Jeff
rogue
rimas
hi
rimas,
and
it
will
introduce
yourself
better
a
similar
solution
developer
at
a
frog
just
couple
of
housekeeping
items,
as
it
is
a
webinar.
You
are
on
mute,
but
we
do
want
your
feedback.
A
B
Hi
so
welcome
and
change.
My
name
is
Remus.
Don't
worry
about
pronouncing
my
last
name:
I
work,
a
senior
solution,
developer,
Jeff,
Rock,
a
really
big
supporter
and
fan
of
open
source
I.
Think
it's
a
good
future
still
ahead
of
add
a
wild
or
co-founded
helm,
wrote
correction
chose
book
and
form
a
sweet
side.
B
I
love,
coffee,
green
tea,
Calvin,
cheesecake,
so
icky
frog,
we
didn't
love,
go
lonk
and
kubernetes
and
we
can
be
using
go
long
since
2015
and
when
ago
models
were
announced
he
got
excited
but
at
the
same
time
soon
realized
that
using
a
model
doesn't
really
guarantee
us
to
have
a
burger
producible
bones.
So
whatever
you
stretch
for
github
could
be
a
different
version.
Every
time
get
it
go,
cuz,
sometimes
not
your
improperly
and
everything,
but
the
two-person
version
but
updated
code.
So
you
cannot
have
reproducible.
B
So
that's
how
or
was
motivation
to
build,
go
center
for
go
along
community
and
of
course,
why
not
use
kubernetes
as
a
platform
as
it
takes
care
a
lot
of
a
platform
later
level
requirements
and
at
the
same
time,
insurance,
which
we
designed
our
application
in
a
cloud
native
way.
So,
basically,
we
are
building
their
apps,
a
trend.
B
B
Of
course,
we
started
from
shell
factors,
stateless
applications,
because
we
don't
want
to
store
a
state
in
local
file
system
across
the
set
of
store
in
databases,
measures,
trees
and
etc,
and
this
which,
of
course,
makes
it's
very
cloud
native
friend
application
because,
for
example,
your
application
restarts
has
no
state
just
can
resume
a
job
and
Doink,
of
course,
that
you
have
to
handle
graceful
shutdown,
for
example,
if
your
applications
processing
something
it
should
stop
and
then
sent
back
to
before,
reports
are
gone.
So
the
next
part
can
resume
and
continue
processing
data.
B
Firstly,
pay
Google
cloud
and,
of
course,
Google
cloud
kubernetes
engine.
Why?
Because
Google
offers
mature
raise
offering
everybody
knows
Coonskin
from
Google,
so
we
can
trust
them.
We
have
a
good
choice
of
managed
services.
In
our
case,
you
need
clothes
girl,
pops
up,
so
you
can
concentrate
on
our
job
so
and
even
building
other
apps
and
a
lot
of
teams
just
could
could
focus
on
deliver
value,
not
read
any
Grabeel
fixing
patching
was
grass
rabbitmq
and
so
etc.
B
Google
Code,
of
course,
offers
a
good
choice
of
troubleshooting
tools
like
extract
Arabic
bigquery,
and
we
love
using
regional
kojiki
clusters,
especially
for
staging
and
production,
because
you
keep
you
know
to
get
spread
between
different
regions.
You
do
your
masters,
cheeky
master,
sir
AJ
as
well,
so
we
can
do
an
upgrade
of
your
kubernetes
cluster.
At
the
same
time,
you
can
roll
your
applications
as
well.
B
So
this
is
no
time
no
time.
Of
course,
there
are
some
glitches
Tama
time
in
Google
as
well,
but
channels
good,
though
so,
currently
running,
total
e5
Cabrini's
clusters
and
each
of
these
environments
has
a
complete
installation.
Oh
go
center
running.
We
have
CI
dev
test
staging
environments,
every
different,
different
projects,
totally
separation
from
each
other
and
staging
production
very,
very,
very
similar
to
each
other.
So
because
how
else
you
can
test
before
running
to
production?.
B
This
is
how
well
clusters
look
like,
so
you
see
on
that
bottom
left
the
use
Google
Cloud,
so
basically
our
back-end
SQL
pops
up
afterwards
views
artifactory
with
our
SAS
at
the
factory
for
to
store,
go
models
so
call
center
is
basically
a
client
of
j-rock
as
well,
and
consent
runs
in
kubernetes
cluster.
We
have
many
different
micro
services
most
of
services.
We
try
to
use
single
container
per
pod
for
for
best
practices.
B
Of
course,
there
are
some
init
containers
which
start
first
initialization
after
so
main
pod
start
and
but
for
UI
microservices,
the
implemented
sidecar
cause.
We
need
some
engines
to
run
in
front
of
you.
I
so
can
into
tactical
traffic,
apply
some
secret
rules
rewrites
caching,
because
we
are
using
engines
interests
which
has
limits
with
caching
so
having
a
sidecar
for
teaching
code
as
a
ninja
container
solve
that
problem,
and
we
are
big
on
pipelines
and
promotions.
I
think
start
in
our
case
for
a
future
branch
where
right
developers
release
a
code.
B
B
Every
can
accessed
be
happy.
We
merge
to
master
a
branch
afterwards
the
ice
system
kicks
in
with
dock
image
and
come
chart
because
TV
version
everything
and
birch
ink
docker
image.
We
have
to
use
so
surely
that
in
some
charter
held
charts
versions
as
well
and
we
start
promoting
I
think
one
by
one
afterwards
then
CI
passes
just
in
say
it
passes
it
deploys
to
dev
and
test
clusters,
so
we
can
do
again
different
tests
on
two
different
environments.
B
Thousands
of
clients
accessing,
at
the
same
time,
how's
the
house,
testing
and
solo,
etc.
When
we
happy
with
the
staging,
of
course,
it
goes
to
production.
So
at
the
end
we
have
six
different
environments.
It
took
a
lot,
but
the
that
still
us
allows
to
do
maybe
two
two
times.
If
you
want
to
today
different
releases
because
so
exci
a
server
tests,
they
are
quick
half
an
hour
and
you
either
gets
tested
our
manual
tests.
We
run
afterwards
on
test
staging
even
on
production
way,
not
too
long
as
well.
B
So
we
try
implementing
I
think
with
the
best
practice
like
killers,
film
secrets,
namespaces,
persistence,
props,
okay,
so
let's
go
one
by
one,
so
Taylor
scale,
a
j-rock
we
used
to
LS
film
beat
you
if
somebody
who's
familiar
with
it.
You
know
we
have
to
learn
Taylor
in
a
cluster
which,
of
course,
nobody
like
that,
because
that's
the
Kuti
breach,
because
you
have
to
apply
cluster
right
wrong.
If
you
want
to
deploy
it
every
two
different
namespaces
or
have
run
multiple
killers,
a
namespace,
so
that's
a
maintenance
burden.
B
Of
course
we
have
to
keep
in
mind
even
running
different
custom
controllers
and
operators.
It's
a
similar
story
with
RAM
as
well,
so
cluster
right
namespace
wide.
Our
back
rules,
though
so
we
use
tiller
as
film
will
be
to
which
face
your
ins,
helm,
client
and
Taylor
together
off
a
cluster
in
the
same
container,
which
does
the
plans.
B
B
B
Sorry,
you
didn't
promote,
promote,
promote
and
promote
their
environment.
Even
our
secrets
are
some
charts,
because
that
allows
easier
to
upgrade
them
and,
of
course,
secrets
are
external
secret.
We
collect
as
external
search
our
applications,
so
instead
she
is
City
pipelines.
So
you
don't
have
to
worry
about
fetching
sequence.
The
deployment
began
because
they
are
a
referent
as
Excel
secrets,
so
we
don't
have
to
worry
systems.
B
B
A
B
As
we
deployment
using
calm,
our
charts
sets
limits
on
all
sellers
if
she
deployed
again
at
some
best
practices
and
we
lowered
some
complex
calculations,
you
come
to
make
sure
we
done
our
appropriate
hour
knows
every
deployment
all
status.
Estimates
on
the
services
are
deployed.
This
is
quite
useful,
auto
scaling,
but
also
guard
against
pot
using
too
much
resources
which
could
impact
our
over
post-ironic.
On
the
same
note,
for
example,
if
you'll
note
is
used
almost
nearly
nine
percent
of
CPU
basic,
it
loses
all
connections
to
database,
to
network,
to
everything
so
basically
becomes
useless.
B
So
if
you
have
to
be,
you
always
have
to
set
not
only
request
by
minimum
question
e
but
limit
the
max.
You
want
to
give
you
application
that
you
of
course
needs
proper
calculations
to
be
done
in
proper
testings,
the
load,
testing
and
etc.
So
we
know
how
your
application
became
behaves
and
how
much
you
give
it.
B
As
well
for
a
che,
you
need
to
use
note,
note
and
pot
affinity,
especially
what
affinity
is
good
using
polyphony
allows
if
your
application
has
microservice
has
five
different,
fiber
Atticus
charm.
If
you
don't
use
Polly
finicky,
it
could
land
on
the
same
node
fiber,
because
they've
had
no
traffic
sources.
If
you
note,
for
some
reason,
gets
killed,
gets
getting
an
upgrade
of
GK
or
something
else.
B
Basically,
application
goes
away
and
all
will
take
could
be
couple
minutes,
a
minute,
a
half
a
minute
to
be
redeployed
on
different
nodes,
so
using
an
affinity
allows
to
spare
them.
So
only
one
part
from
the
same
replica
set
will
be
deployed
on
one
numbers
in
low.
We
will
be
all
spread
around
Amina
one
node
is
gone.
B
B
B
So
you
not
not
overkill
your
application
at
that
time
round.
The
next
pod
gets
deployed.
If
you
said
90
80,
90
percent,
you
could
cut
a
big
and
when
you
port
is
deployed
basically
that
it,
what
could
be
become
too
busy
to
accept
more
traffic
also,
so
we
call
for
computer
TV.
You
have
a
couple
services
using
Grauman's
cpu,
most
of
a
services.
B
We
use
pops
up,
QC
check
with
HP
area
like
casa,
metrics
and
if
Q
is
a
bigger,
rescale,
my
part,
so
basically
using
native
which
be
a
complete
and
spew
the
scanning
and
custom
auto
square,
using
the
pops
up
and
of
course,
so
that's
a
messaging
and
then
pink
is
gone.
Cpu
is
the
less
use.
Pops
up
is
clean,
so
we
scale
I,
think
it
down,
because
why
you
should
pay
for
things.
They
don't
really
need
to
be
beautiful.
So
you
should
be
scale
for
me.
B
Five
notes,
5
part
2
morality
is
allowed,
set
neon,
auto
scaling
groups,
parts
and
nose,
so,
for
example,
so
this
is
like
our
looks
like
and
she
who
speaks
so.
That
means
like
CPU,
so
today's
much
higher
so
walking
to
this
do
some
little
mats.
So
example,
our
node
can
accept
five
parts.
So
we
have
15
parts
running.
B
B
Note
pose
not
posed
it's
very
powerful
stuff
and
it's,
and
in
our
case
we
have
a
few
different
types
of
null
pools
like
for
persistence
for,
like
default,
nepali
using
torrents,
a
couple
microservice
for
ghosts,
interviews
which
should
be
restarted
because
they
process
traffic
process,
the
data
a
lot,
and
where
is
stunning,
it
will
take
a
while
to
pick
it
up
again.
We
have
a
preemptive
all
now
put
most
of
our
microservices
ironing.
Somebody
is
not
familiar
primitive.
No
poles
is
basically
offering
from
google.
You
pay
only
technique
under
5%
of
a
real
price.
B
If
one
node
gets
killed,
it's
nothing
much
happens.
You
know
what
was
invite
because
they
are
most
important
micro
services.
They're,
like
processor,
we're
running
on
persistent
copper,
vegan
and
parasite
on
PM's
and
making
is
having
six
PPM's
and
paying
only
kinda
wish
for
two
of
them
evenly.
It
allows
us
to
have
more
portfolio
to
be
used.
B
He
was
really
leveraging.
Yes,
though,
as
I
mentioned
pmstwin
it
on
24
hours
and
closed
24
hours,
their
home
output
could
be
killed
at
once,
though
we
don't
want
that
to
happen,
so
we
used
G,
keep
even
cleaner.
The
made
me
buy
a
spectacles
so
that
small
controller
I
get
put
on,
but
it
has
which
runs,
doesn't
need
any
coaster
ride
or
walk
access.
The
trans
inconvenience
quarter,
labels
or
TV
ends.
Sorry,
it
is
obedient
and
after
12
hours
starts
itself,
killing
them
one
by
one,
of
course,
not
at
least
same
time.
B
So
we
investigate
each
other,
much
better
control
when
PBMs
go
down,
go
away
and
come
back
so
far
so
good.
We
are
running
that
way
more
than
about
seven
months,
big
huge
stating
sneaking
billions
on
one
side
and
we
have
more
resources.
We
can
accept
more
traffic,
so
we
shall
now
to
slow
so
default,
multiple,
the
running
most
important
things
for
my
stateless
micro
services,
but
it
should
be
the
killed
mix
of
ours.
B
B
So
it's
not
a
very
easy
time
checks
which
these
mystics
you
can
see
with
spikes,
because
Afghan
checks
are
done
from
all
different
regions,
Google
illusions
and
Tommy
time
regions.
We
have
some
with
spikes
very
latency
and
we
get
modified
by
slack.
Some
take
happened.
We
check
it,
usually
some
network
latency
at
Google
site,
but
it's
good
to
be
notified
about
that
as
well.
B
We
leverage
a
lot
started
logging,
so
why
we
should
use
something
else
if
there
is
a
log
inside
what
any
part
of
that
managed
service,
inaudible,
Google
that
sweet
sweet
up
from
running
Shabana,
we
have
our
rocky
bond
at
all
for
testing
purposes,
but
not
even
much
yet,
but
this
real
cool
one.
You
can
check
a
logs.
We
can
expert
logs
to
betray
you.
Let's
go
case
against
that
and
of
course
we
have
Prometheus
with
her
father,
because
we
like
different
dashboards,
we
can
check
different
things.
B
So
how's
engineer
he
used
house
of
ginika
on
basic
cleaning
nodes,
killing
parts
trap
see
checking
how
it
behaves
but
good
thing
as
well.
We
have
House
South
engineering
by
default
because
using
PDF,
sorry
cousin
generic,
because
Todd's
get
killed
and
now
not
get
killed
every
day.
So
this
in
a
real
time
in
production
which
Apple
will
find
cause
engineering,
but
we
of
course
do
extra
things
with
just
in
advance.
We
protesting
the
Deaf
in
test
court
load
testing.
B
We
do
long
different
load,
testings,
Janica,
api
jamming,
you
I
thousand
users
different,
go
models,
maybe
checking
how
to
behave,
maybe
a
new
version.
Maybe
it
needs
more
CPU
around.
Maybe
more
resources,
so
I
think
has
to
be
very,
very
well
tested
on
that
side,
because
if
you
don't
pretest
law,
even
infrastructure
class
applications
very,
very,
very
hard
ways
and
of
course
we
wanted
to
keep
in
touch
without
no,
it
doesn't
think
so
on
me
he
was
a
small
cube,
slack
application,
which
checks
or
what
killed.
B
What
would
he
started
and
when
we
deplore
the
new
version
being
notified
well
in
slack
because
he
kept
a
support
center
support,
slack
channel.
So
that
get
modified,
the
new
version
got
released,
guitarist
getting
released
so
support
guys
they
won't
freak
out
and
it's
finished
or
training
release
is
finished
because
an
innovation
is
not
caught.
You'll
see
video
messages
post
part
30.
We
killed
an
air
it
covered
so.
B
A
A
You
have
on
your
on
your
webinar
explain
and
while
you
are
typing,
your
questions
in
I
want
to
remind
you
that
cube
con
plus
cloud
90
of
North
America
coming
upon
us
relatively
soon,
and
if
the
flagship,
CN
CF
event,
and
this
year
it's
in
San
Diego,
which
is
the
best
location
that
you
can
find
for
conferences
in
November,
because
it's
it's
nice
and
shiny
and
obviously
obviously
do
not
means
for
for
many
years
cube
colonies
is
sold
out.
So
now
will
be
good
time
to
grab
your
tickets
before
before
before
it
sold
out.
A
A
That's
that's
good,
and
there
is
another
question
and
my
what
are
one
of
the
results
I
mean:
we've
been
running
NGO
center
for
more
than
more
than
half
a
year,
now
Yeah
right
from
January,
it's
like
nine
months
and
and
I
think
that
it's
it's
about
time
to
to
question
ourselves
and
and
see
if
our
architecture
actually
works.
So
what
is
what
is
the
uptime?
How
did
it
behave?
What
do
we
see
after
abatement
for
production.
B
B
B
A
A
B
A
And
and
just
to
I
guess,
my
point
is
thank
you
and
this
for
allowing
us
to
to
provide
a
service
which
has
zero
downtime
for
four
four
eight
months
of
production,
service.
I.
Think
that's
remarkable
and,
and
a
lot
of
the
crane
goes
to
goes
to
kubernetes,
here's
another
and
another
question:
how
to
measure
appropriate
resource
limits
for
your
applications.
B
B
You'd
add
a
bit
on
top
of
it.
That's
a
limit.
Adding
will
always
thinks
if
you
have
test
again
and
see
how
the
became
basically
it's
it's
testing
and
testing
and
test.
It's
not
like
just
pure
maths
after
you
adjust
and
adjust
and
adjust
experiments
experience.
And,
of
course
we
don't
do
that
before
Duan
death
with
you
on
test.
Do
that
on
staging-
and,
we
repeat,
repeat,
repeat
and
be
happy,
we
only
die.
Then
we
apply
to
production
because
productions
production
just
has
to
be
no
downtime
so
far,
so
good
for
that
yep.