►
From YouTube: Practical Way to Build Kubernetes Native Java Controller - Zibo He & Min Jin, Ant Financial
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
Practical Way to Build Kubernetes Native Java Controller - Zibo He & Min Jin, Ant Financial
The controller pattern has been proven as an effective way for managing Kubernetes workloads. However, for other languages, there are still many challenges remained in developing third-party controllers for Kubernetes. In this talk, we will discuss how to implement reflector, internal store, working queue and leader election in native Java style, and demonstrate controller runtime that makes it easier to develop Java controllers from scratch. We will also discuss different operators that we build to integrate with micro-service framework for cloud native application development.
https://sched.co/UacE
A
Hi
everybody
welcome
to
our
presentation,
and
today
our
topic
is
practical
ways
to
be
able
to
coordinate
his
native
java
controller
and
before
this
I'd
like
to
have
a
quick
survey,
anyone
who
tried
or
considered
to
build
a
java
controller
in
a
cluster.
Please
raise
your
hand
please
great
thanks.
Thanks
I
see
around
10
hands
in
this
room
and
that's
actually
much
more
than
I
expected.
A
Yet
there
many
companies
has
tried
or
actually
deployed
a
Java
controller
yet,
and
but
we
did
and
we
bring
it
all
the
way
to
the
production
and
to
manage
the
underpin
infrastructures
into
the
center
in
we
hope
by
this
session.
We
can
share
you
with
some
more
facts
and
experiences
about
programming
and
running
java,
kubernetes
controllers
and
operators
like
pros
and
cons,
and
also
the
lessons
the
painful
lessons
we
learn
from
it.
A
Wait
before
we
dive
into
the
content.
Let's
do
a
brief
warm-up
into
self
introduction
and
it's
gonna
be
start
by
Tony
and
hi
Tony
he's
a
criminalist
maintainer
and
we
are
working
in
the
same
team
in
the
on
financial
past
platform
and
he's
a
marriages
co-lead
of
the
six
COI
focusing
on
COI
multi-tenancy
and
secure
container
runtime,
and
as
for
me,
my
name
is
Ming
and
some
contributors
in
the
community
might
have
known
me
as
Yui
and
yeah.
A
A
So
if
you're
interested
in
API
server
flow
control,
please
visit
a
schedule
home,
but
the
open
API
facilities
in
kubernetes
FS
server
is
tightly
connected
with
our
topic,
because
all
these
Java
controller
framework
are
build
up
on
kubernetes
official
Java
clients
and
the
DES
client
is
actually
generated
from
the
published
open,
API
spec
from
kubernetes
and
for
folks
who
are
interested
in
open,
API
schema
and
we
will
expand
the
expand
more
about
this
shortly
after
ok.
This
is
the
outline
or
a
general
of
what
we
are
talking
about
today.
A
It's
going
to
be
start
by
general
approaches
to
build
controllers
module
English,
and
this
is
where
we
are
going
to
inspect
the
kubernetes
operators
in
a
high
level,
regardless
of
their
languages,
and
this
will
help
us
to
understand
what
Java
control
of
room
work
is
doing
for
you
and
the
the
number
two
is
merits
of
developing
controllers
in
Java.
In
this
scandal
we
will
have
a
vibrant
comparison
with
go
long
operators
so
that
we
can
better
understand
why
we
are
doing
this
like
yeah.
A
So,
let's
start
by
sorting
out
the
existing
control
of
remotes.
Let's
take
a
look
at
the
representatives
of
the
existing
controller
from
works
today.
The
first
one
is
controller
builder,
okay,
sorry
computer.
It's
not!
No
doubt
that
is
the
most
popular
one
currently,
and
also
the
recommended
approach
provided
by
the
community
and
computer
is
actually
a
binary
cutter
tool
that
helps
you
initialize
the
new
projects
and
new
API
resources.
In
even
a
mission
Web
books
by
templating,
the
source
code
is
source.
Codes
are
for
the
developers
degenerated.
A
All
the
templates
controller
is
empowered
by
another
project
named
k-6
controller
runtime,
which
is
awesome.
Project
encapsulate
sand
highs.
The
underneath
mechanisms
like
in
formerly
swatching
in
cash
management
for
the
developers
and
operators.
Sdk
is
used
to
be
really
popular
framework,
and
it's
now
under
active
maintenance
and
I
think
we
really
didn't
have
much
experience
using
these
projects,
but
I
think
the
computer
and
operator
SDKs
looks
really
close
to
us
because
they
are
basically
doing
the
same
thing
from
the
users
perspective.
A
The
third
one
is
a
Passover
builder
alpha
and
I'm.
Sorry,
the
third
one
is
API
server
via
alpha.
The
cupula
project
is
actually
a
precedence
of
these
projects
and
the
API
server
build
alpha
is
developed,
developed
by
Google
Earth
in
the
beginning,
and
its
feedbacks
and
drawbacks
are
successfully
addressed
in
computer,
but
this
project
is
still
under
active
maintenance,
but
this
project
is
mainly
like
aimed
at
helping
users
who
build
new
resources
based
on
API
server
application.
A
Stanza
is
a
really
interesting
project
and
I
think
we
can
call
this
a
callback
based
controller
and
there's
a
interesting
project
named
the
meta
controller.
In
this
project,
the
web
developers
will
have
to
develop,
build
or
compile
a
whole
controller
for
anymore.
What
you
require
to
do
is
just
to
simply
implements
a
web
book
and
wait
for
the
central
color
to
notify
you
when
you
need
to
reconcile
something
so
great,
let's
list
among
the
table
and
have
a
good
comparison
between
the
two,
as
you
can
see
in
the
screen.
A
Let's
start
by
the
first
two
rows
in
the
table.
Builder
based
controllers
are
actually
helping
you
open
users
who,
with
the
step
one
of
building
controller,
but
it
doesn't
handle
how
to
continuously
manage
operators.
Life
cycle
right,
upgrading,
computer
January's
are
working
deployments,
yamo
documents
for
use
and
let
the
native
deployment
controllers
deal
with
all
the
other
stuff
well
in
the
combat
based
it's
taking
over
and
decoupling.
A
A
Like
it's
doing
well
in
scaffolding,
test
solution,
let's
watching
and
cash
management
well
last,
but
not
least,
as
for
mod
Z
lingo,
suddenly
the
arm,
how
does
the
true
person
works
for
multi
programming
languages
for
callbacks
based
controller
framework?
Develop
developers
are
only
required
to
write
a
server
implements
simple
HTTP
or
HTTPS
protocols,
interface,
which
is
really
simple
nowadays
now
and
on
the
other
side.
Sadly,
builder
based
controller
doesn't
work
or
cross
languages,
so
we
will
have
to
implement
a
completes
builder
libraries
per
programming,
language
and
the,
for
example.
A
The
controller
anthem
is
written
in,
go
so
in
nature
requires
the
developers
to
write
their
controllers
in
go
steadily
for
nowadays.
Java
developers
doesn't
benefit
from
that,
so
that
what
we
are
doing
is
to
port
those
mature
controller
libraries
to
Java
world
and
make
use
of
the
merits
of
Java
like
object-oriented
programming
languages
as
much
as
possible.
A
Ok,
here's
another
question.
So
how
are
you?
Why
are
saying
that
you
are
porting
libraries
from
Gollum
to
Java?
Is
there
any
improvement
for
us
to
do
so?
Yes,
I
believe
and
I
listed
the
several
items
on
the
screen:
let's
expand
them
one
by
one.
First,
one
is
the
heap
dumping.
Currently,
the
full
or
complete
heap
dumping
is
not
supported
in
gulag.
A
Even
though
golem
provides
really
useful
toolings
to
do
the
profiling
on
the
running
processes,
but
yeah
we
can
analyze
the
overall
tracing
of
memory
allocations
or
CPU
consumptions,
but
there's
no
food,
that's
not
complete.
Instead
in
Java
because
of
a
memory,
many
more
management
is
running
on
TV,
a
higher
layer
on
the
operator
system.
So,
with
the
help
of
that,
we
can
like
simply
checking
out
what
is
happening
on
the
heap.
What
is
happening
on
a
stack
with
the
like
toolings,
like
Eclipse,
Matt
yeah?
This
is
really
useful.
A
As
you
can
see
in
a
screen,
the
left
side
is
Java
and
using
the
Matt's
toolings,
you
can
easily
find
out
what
was
the
status,
the
details,
the
status
of
every
object.
This
will
help
us
to
do
some
fine
query
in
the
deeper
games
like
when
you,
when
you
want
to
check
out.
What's
talking
in
The
Informers
work,
you,
like,
oh
sorry,
the
controller's
Q
and
for
running,
go
process,
it's
almost
impossible
to
finding
out
what
is
happening
over
there
and
also
you
can
find
out
what's
missing
in
the
cache
entry.
A
A
A
As
for
the
performance,
the
second
one
is
performance
diagnosis,
so
I
listed
the
the
Prometheus
metrics
on
the
screen
on
the
left
is
the
Prometheus
DMX
exporter
yeah.
As
you
can
see.
Obviously,
the
number
of
Java
is
more
than
go
but
yeah.
Let's
compare
the
supported,
Prometheus
metrics
on
there
under
around
town
yeah.
A
In
conclusion,
Java
does
better
in
capturing
what
is
happening
in
GC,
but
on
the
screen.
There's
like
a
smiley.
Okay,
there's
like
only
two
metrics
over
here
in
the
screen.
That's
for
GC
yeah,
that's
sad
for
go-go
developers,
but
for
Java.
You
know
that
you
can.
We
can
easily
replace
the
GC
engine
and
there's
how
lot
of
metrics
for
for
debugging,
Java
GC
processes
and
but
the
other
ways,
both
in
Java
and
go
long.
A
You
are
required
to
you
are
required
to
have
a
deep
understanding
of
the
memory
management
model
of
the
both
languages
and
yeah.
Okay,
let's
jump
to
the
third
the
thread
model.
The
Java
thread
is
running
as
a
state
machine
driven
by
external
cause,
and
a
thread
can
be
interrupted
immediately
by
another.
By
sending
a
signal,
while
in
:
is
tough
to
shutdown,
a
goroutine
per
actively
interfering,
the
states
of
a
running
go
routine
is
almost
impossible.
A
Currently,
if
I
have
to
do
something
like
a
select
select
channel-
and
we
like
like
stop
quits
asset,
a
Kokoro
team
and
they
when
a
stop
channel
returns,
but
practically
unlike
the
go
applications
the
java
application.
Won't
you
really
have
a
happy
ending
because
when
it
comes
to
starting
it
takes
operators
more
times
to
shut
down,
and
it's
in
the
cord.
A
It's
a
common
thing
to
have
operators
loop,
crushing
sometimes,
but
in
Java,
is
almost
in
embargo
like
if
you
develop
a
java
application,
you
will
easily
allow
allow
it
to
like
keep
crashing
in
the
kubernetes
cluster.
That's
also
a
bearable
in
our
company
and
biingo
long,
especially
in
Nha
setup
operators
are
normally
working
by
locking
all
resource
yeah
in
kubernetes.
We
call
this
a
leader
election
I,
don't
pretty
much
believe
it?
That's
not.
Actually,
a
leader
election
in
from
my
view,
yeah,
but
:
operators
are
really
good
at
crushing.
A
It
doesn't
matter
how
many
times
it
just
crash,
never
mind
there.
There
will
be
a
leader
they
are
holding
to
doing
all
this
stuff
yeah
in
in
Java,
because
the
thread
life
cycle
can
be
managed
by
external
cause.
We
can
do
something
like
you
can
completely
shut
down
a
Java
controller
without
just
restarting
the
container.
You
can
like
manage
the
worker
thread
in
a
vine
growing
manner.
A
Okay,
the
next
one
is
inherent
generics
apparently
go,
doesn't
know
anything
about
it
and
there's
a
custom
extra,
but
in
Java
there's
a
lot
of
Java
is
an
object-oriented
language
and
it
has
the
benefits
to
program
like
that.
You
are
a
you
are
like
implementing
a
Lister
or
Informer.
You
will
just
have
to
pass
the
like
know
that
and
pawed
some
resources
as
a
type
parameter,
and
you
you
don't
have
to
like
you
know
we're
using
a
code
generator
to
generate
other
stuff.
Those
are
those
generated.
A
Code
are
really
duplicated
and
that's
all
because
go
long
doesn't
fully
supports,
doesn't
doesn't
support
any
hybrid
inheritance
and
generics
and
as
for
the
open
source
resources,
yeah
a
patch
has
been
running
for
years
and
there's
many
many
successful,
open
source
projects
resources,
although
even
though
they're
some
of
them
has
been
lost,
lost
the
maintenance
there.
They
are
not
actively
maintaining
but
yeah,
but
still
up
his
resources.
Much
more
than
this
in
this
sense,
ever
since
he
ever
I
did
it
for
now.
A
So,
let's
back
to
the
business,
what's
our
Java
controller
from
work
is
providing
is
many
libraries
which
is
parted
from
:.
We
are
not
planning
to
events
or
reinvents
the
existing
client
libraries
like
Informer
in
the
foreseeable
future,
and
what
we
are
trying
to
keep
it
is
just
to
make
it
the
same
as
the
gold
libraries
as
much
as
possible
so
that
Java
developers
can
use
it
auto
paths
and
do
anything.
Just
:
developers
can
already
do
by
the
way.
A
I
think
there
is
a
barrier
between
the
two
language
for
now
on
Java,
actually,
like
Java
developers,
can
think
that
they
may
be
good
at
redefining
the
things
go.
:
a
breweries
have
already
yeah.
Vice
versa,
but
at
least
we
start
by.
We
can
just
start
by
acquiring
the
same
capabilities.
Gophers
Gophers
can
already
do,
and
it's
gonna
be
a
quick
demo
from
Toni
about
how
to
actually
use
our
Java
control
of
frameworks
and
heighten
e.
B
B
B
So
can
see
so
here
in
my
laptop
I
have
a
local
co
blend.
His
cluster
yeah
also
I
have
initialized
Java
product
with
some
necessary
dependencies,
as
you
can
seen
here,
you
can.
You
have
to
either
turn
Java
and
Java
API
and
Java
extended
those
three
modules
to
your
project
memory
dependencies.
So
for.
A
A
second,
so,
as
you
can
see
in
the
screen,
the
palm
dependencies,
the
care
that
there's
a
IO
kubernetes
group
in
artifact
ID
is
Java
client
Java,
that's
the
official
Java
clients
we
are
talking
about
and
on
the
screen,
there
version
is
601.
That's
the
latest
for
know
about
we'll
be
releasing
a
seven
really
soon,
probably
next
this
week.
So
next,
okay,.
B
B
A
B
B
Just
the
finishes
reconsider
all
right
and
this
here,
the
it
without
force
means.
Then
you
have
no
need
to
rig
in
queue
the
object
at
some
time.
Okay,
so
with
just
a
finish,
you
reconsider
or
the
next
step.
You
need
to
write
or
controller
to
go
back,
reconsider
and,
of
course,
before
you
write
your
controller,
you
have
to
create
the
background
Informer
and
you
may
notice
that
we
have
already
added
s
here:
ok,
which
is
a
little
bit
different
from
goal
and
controller
eyes.
B
B
B
A
B
B
Okay
for
the
future
enhancement,
they
are
already
in
some
work
items
in
your
roadmap.
I
think
in
Maine
is
working
on
open,
API
version
3,
which
brings
more
powerful
features
like
multi
returning
onboard
enlist
or
stream.
Once
this
is
done,
the
Java
controller
will
switch
to
land
in
the
old
version.
Open
API
schema
also
with
the
notice
some
memory
or
performance
issue
in
large
size
cluster.
B
So
we
are
working
on
some
customer
object,
civilizer,
which
can
chop
the
fields
which
you
are
not
interested
in
to
make
the
operator
more
memory,
training,
ok,
and
about
the
controller
deployment.
We
are
working
on
some
gallery
release.
We
try
to
integrate
the
gallery
release
strategy
with
your
operated
deployment.
So
then
you
can
control
how
you
can
deploy
your
operator
to
the
cluster
in
case
there
is
some
issues
with
your
leo
version
operator:
ok
mohawk,
as
in
more
work
to
do
so
contribution
is
WorldCom.
Ok,
I
think
we
just
finished
any
more
questions.
A
C
Thank
you.
It's
actually
more
of
a
statement
and
a
question.
It's
the
first
of
all
great
work,
I
very
welcome
this
and
we're
definitely
not
alone,
and
we
also
are
working
a
company
at
a
phosphorus
phosphorus
operator
and
we're
building
it
in
Java,
because
that's
what
our
expertise
is
so
and
I've
already
talked
to
some
other
people
who
are
willing
to
commit
some
efforts
to
a
kind
of
a
job
operator
framework
and
so
far
we
have
included
most
of
the
things
that
you
mentioned,
including
up
in
open,
API
or
including
web
cool
web
hook.
C
A
Thank
you.
Thank
you,
so
our
our
project
is
is
always
opening
to
the
community.
So
if
you
got
any
better
implementations,
we
are
always
welcome,
like
you
meant,
you
just
mentioned
some
advances
or
focuses
for
a
second.
Do
you
completely
implemented
The
Informers
or
the
controller's,
or
just
like
the
plain
reactive
programming
in
Java,
the.
A
C
Thank
you.
I
just
want
to
say
that,
first
of
all,
it's
a
work
in
progress
and
and
second
is
following
a
more
reactive
programming
model
right
now
and
has
some
differences.
It's
based
on
fabric
aid
instead
of
the
official
Java
client,
it's
not
targeted
to,
but
it's
a
using
growl
vm
for
native
compilation
as
by
default,
because
that's
what
we
wanted-
and
you
know
is,
there's
many
different
approaches
to
building
these
things.
C
A
Yeah
there
is
many
successful
reactive
programming
frameworks
in
Java
today,
but
just
as
you
say
that
it's
not
actually
a
complete
equivalence
of
the
go
long
informers,
that's
probably
a
problem
here,
because
if
you
check
out
the
implementation
of
the
under
underlying
formers
in
the
indexers
in
colon
they're,
like
really
they
are
designed
really
well
and
it's
complicated.
It's
not
just
like
pushing
events
and
execute
something
in
notification
or
all
the
controllers
and
operators
are
designed
to
be
running
in
a
stateless
way.
A
The
events
can
be
duped,
I'm,
sorry
for,
for
example,
if
you
have
multiple
update
events
in
the
queue
that
can
be
squashed
into
one.
So
many
stuffs
like
this,
so
there's
a
lot
of
extensions
in
Golan's
Informer,
comparing
to
the
plain
reactive
programming
frameworks,
yeah
we'd
love
to
see.
If
you
are,
if
you
are
willing
to
contribute
these
features
to
the
community,
we
always
welcome.
We
have
the
families
for
reviewing
and
merging
your
pools.
I
think.