►
From YouTube: Meet the Gears Behind Kubernetes APIs: Introduction to SIG API-Machinery - Federico Bongiovanni
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
Meet the Gears Behind Kubernetes APIs: Introduction to SIG API-Machinery - Federico Bongiovanni, Google
It will be a big overview of the SIG. We will go through several sections, including: - the charter of the SIG, - current SIG structure, - the areas of ownership and the different components that fall under the SIG domain - the regular meetings - places where you could get involved - plans for onboarding programs
https://sched.co/Uajg
A
Hello,
everybody
thank
you
for
coming
today.
This
is
going
to
be
an
introduction
to
seek
a
bi
machinery.
My
name
is
federico.
Vaughn
Giovanni
I'm,
an
engineering
manager
at
Google
I,
was
a
user
of
kubernetes
because
before
becoming
involved
in
the
other
side
of
the
coin
and
helping
cornetist
grow,
so
it's
been
a
really
interesting
journey.
For
me,
I'm
very
proud
to
be
part
of
this
community
I
was
making
fun
with
my
fellows,
seek
members
here
that
the
deep
dive
for
API
machine-
it
just
happened
before
this
one.
A
A
A
A
What
is
our
roadmap,
which
is
super
packed?
And
then
hopefully
you
will
learn
about
some
opportunities
where
you
can
join
us.
Just
maybe
quietly
join
our
meetings
here.
What
we
do,
what
do
we
talk
about
and
if
you
want,
you
know,
start
helping
us
contributing
and
learning
from
the
experts
very
good,
so
high
level
overview?
Everybody
knows
what
a
kubernetes
cluster
is.
Hopefully
there
is
a
part
of
the
cluster
that
is
called
the
master
nodes.
There
is
another
part
of
the
cluster
that
is
called
the
user
nodes.
A
There
is
a
persistence
layer
that
is
HCD
in
the
case
of
Cornette
is
master
node
talks
to
the
nose
master,
node,
aka,
kubernetes
control,
plane
talks
with
the
persistence
layer,
and
today
we
are
going
to
talk
about
these
guys
and
a
little
bit
about
these
guys.
This
is
what
capi
machinery
is
about.
What
runs
in
the
masternodes
usually
is
also
known
as
the
cornetist
control
plane.
A
So
first
question
is:
is
API
machinery
equals
to
all
the
kubernetes
api?
A
common
misconception?
No,
it's
not!
So
when
you
have
a
question
about
a
particular
API,
we
might
know
about
it
because
we
use
kubernetes,
but
you
know
we
definitely
are
not
the
owners
of
that.
We
don't
test
it.
Usually
individual
api's.
We
don't
know
everything
about
those
api's.
A
So
following
our
charter,
sig
API
machinery
responsible
for
the
development
and
encasement
of
cornetist
control
plane,
the
scope
covers
api
server,
persistence,
layer
at
CD,
controller
manager,
cloud
controller
manager,
custom
resort
definition
and
web
hooks.
This
is
linked
to
our
Charter.
In
other
words,
API
machinery
is
the
generic
technical
stack
that
allows
you
to
interact
with
a
kubernetes
cluster,
talk
to
all
the
individual
api's
and
extend
the
kubernetes
api.
A
A
You
can
go
to
our
Charter
again
and
get
a
you
know
very
detailed
list
of
things
that
we
own
with
component
names,
in
other
words,
trying
to
make
it
a
little
bit
more
readable.
Basically,
we
own
API
server.
Obviously
we
own
the
controller
framework,
but
we
don't
own
all
the
individual
controllers
very
important
with
on
some
of
them,
because
they
are
part
of
the
machinery.
But
then
you
know
every
kubernetes
resource
or
kubernetes
and
other
areas
that
they
build
the
controllers
they
use
our
framework.
A
We
don't
own
those
we
own
the
client,
generators
and
specifically
client
go
line.
Go
is
an
area
where
a
lot
can
be
done.
If
you
want
to
help
us,
for
example,
basically
because
client
go
is
used
internally
for
internal
communication,
so
whenever
you
have
to
use
play
and
go,
the
commentation
is
very
poor.
A
The
examples
are
not
super
clear
or
super
basic,
so
I
have
a
funny
story
about
somebody
that
is
not
here
but
works
in
the
office
with
us,
and
he
was
you
know,
coming
like
four
or
five
times
a
day
and
speaking
for
like
half
hour
with
members
of
my
team,
just
wondering
and
understand
how
to
use
the
client
goal
right.
So
after
that
we
did
some
follow-ups,
we
did
some
improvements
in
the
documentation,
but
there
is
a
long
list
of
things
that
can
be
done
there.
A
If
you
are
interested
so-
and
you
know
a
list
of
things,
I
think
one
of
the
most
important
ones
that
everybody's
talking
and
I'm
going
to
mention
a
little
bit
is
the
extension
mechanism
which
in
116
went
GA.
So
you
heard
about
series
you
had
about
admission
webhooks
super
important
and
we
will
talk
a
little
bit
about.
Why
is
so
important.
A
This
is
a
schema
that
Stefan
here
created
for
a
cap
explains
a
little
bit
how,
with
the
life
of
request
coming
through
our
machinery,
you
see
there
is
a
complex
diagram,
a
lot
of
layers
and
I.
Think
one
of
the
things
that
is
very
important
to
remark
is
that
kubernetes,
through
API
machinery
allows
you
every
user
to
extend
the
kubernetes
api
first
year.
A
A
A
This
is
a
code
that
I
really
like,
and
you
know
I
think
illustrates
a
little
bit
that
the
theme
of
this
talk
kubernetes
is
not
just
API
driven
is
API
centric
Brian
grant
for
those
of
you
who
don't
know
him
is
one
of
the
creators
of
cornetist
together
with
everybody.
That
was
the
other
six
that
were
yesterday
in
the
mainstage,
so
more
or
less
I
get
you
get
an
idea
of
what
do
we
own?
A
A
It's
an
interesting
topic.
If
you're
you
know
involved
into
these,
because
there
is
a
number
of
discussion
discussions
in
the
open-source
community
about,
for
example,
api
compatibility
between
kubernetes
minor
versions
and
who
is
responsible
for
guaranteeing
that
we
don't
break
backward
compatibility
or
even
worse
forward,
compatibility
or
even
worse,
which
is
probably
one
of
the
most
scary
things
that
could
happen,
not
changing
compatibility
in
the
api's,
but
changing
behavior
in
the
objects
that
you
are
using
right.
A
A
We
don't
own
most
of
the
controllers,
as
you
can
deduct
from
the
previous
statement.
We
usually
don't
own
the
operational
characteristics
of
your
cluster.
You
know
we
get
a
lot
of
questions
and
I'm
going
to
put
in
the
last
slides
a
number
of
links
to
the
slug
Channel
and
the
mailing
list.
You're
always
welcome
to
come
and
connect
and
ask
questions
but
yeah
the
individual
failures
of
your
cluster,
and
you
know
the
reasons
why
is
failing.
We
can
try
to
help.
A
Next
one
is
a
hot
topic:
we
don't
own
your
particular
API
review,
we'll
be
imaginary.
Some
members
of
API
machinery
are
kind
enough
to
be
API
reviewers,
but
API
reviews
is
something
that
ABI
C
architecture
owns
and
they
have
their
own
process.
You
can
connect
with
Jordan
if
you're
interested
in
learning
more
about
that,
but
we
are
not
the
owners
of
that.
You
can
feel
I'm
pretty
defensive
on.
This
is
not
the
intention.
A
Okay,
so
this
is
a
little
bit
of
learning.
Why
is
so
central
and
important
and
hard?
You
know
picture
of
a
cluster
in
the
center
of
everything
there
is
an
API
server.
You
know
that
in
the
control
plane
we
I
just
list
it
like
two
minutes
ago.
We
own
a
number
of
components.
Some
of
the
most
important
ones
are
probably
the
controller
manager,
the
cloud
controller
manager,
the
scheduler,
and
then
there
is
others,
but
those
are
the
main
ones.
A
There
is
a
number
of
things
that
you
can
do
to
your
cluster
to
extend
it.
Some
of
them
may
run
in
the
master
nodes.
Some
of
them,
like
you
know,
web
hooks
could
be
running
somewhere
else.
You
know
I
put
them
in
the
cluster,
but
they
could
be
also
outside
of
the
cluster.
Then
there
is
the
nodes,
and
then
there
is
everything
that
wants
to
interact
with
the
cluster
I
forgot
to
mention.
Probably
you
know
with
the
the
fact,
and-
and
this
is
where
I
linked
it
with
the
quote
from
Brian
grant.
A
Kubernetes
is
a
bi
centric,
which
means
that
even
these
components
from
the
master,
node,
local,
the
controllers
or
the
cloud
controllers,
they
are
not
allowed
to
speak
to
each
other.
Nobody
can
talk
to
each
other.
You
know
all
the
communication
is
done
through
API
server.
Api
server
is
the
one
that
knows
what
needs
to
happen
before
a
message
can
or
an
event
can
be
persisted
or
communicated
to
the
rest
of
the
controllers.
A
I
also
didn't
put
in
this
diagram
because
funny
enough
I
run
out
of
transitions
to
do
these
animations,
but
the
reset
CD,
who
also
only
talks
with
API
server,
an
API
server,
only
talks
with
HC.
So
nobody
else
can
talk
to
that.
So,
let's
go
there.
Let's
go
back
to
where
we
were,
and
this
is
users
and
users
are
sometimes
a
human.
Sometimes
is
a
client,
sometimes
is
a
human
doing.
Cube
CTL
get
pots,
sometimes
is
a
Python
client
as
part
of
Jenkins
continuous
deployment
pipeline
trying
to
deploy
something
to
the
cluster.
A
A
If
you
have
run
clusters
in
production,
you
may
know
that
you
know
API
server
could
go
down
and
the
cluster
and
what
is
trying
on
the
nodes
for
some
time
will
survive.
That's
some
time.
Some
time
it's
undefined.
It
depends
a
lot
on
how
your
cluster
is
configured
and
how
it's
running,
probably
will
survive
a
couple
of
hours,
but
as
soon
as
something
goes
wrong
in
the
node,
something
needs
to
be
reacted.
If
API
server
is
not
responding,
you
know
you're
going
to
get
into
troubles
and
this
can
grow
right.
A
So
you
can
have
a
lot
of
nodes.
If
you
go
to
the
talks
from
seek
scalability,
you
may
have
heard
that
you
know
we
are
trying
to
push
the
boundaries
of
how
much
how
big
a
cluster
can
be.
I,
don't
know
that
some
numbers,
but
it's
in
the
tens
of
thousands,
so
you
can
imagine
that
you
know
how
complicated
gets
the
reliability
of
a
cluster
when
all
these
cubelets
are
trying
to
synchronize
the
state,
and
you
know
when
scheduler
is
trying
to
tell
API
server
to
go
on.
A
A
This
is
another
view
from
a
talk
from
a
previous
clip,
come
from
somebody's
here
in
the
audience,
but
it's
the
same
like
API
service
at
the
center
of
all.
These
components
and
I
wanted
to
include
these
to
highlight
these,
like
controller
manager
talk
to
a
big
server,
but
at
the
same
time
Control
Manager
is
not
a
component
controller
manager
is
talking
to
all
the
controllers.
This
is
the
controller
framework
right
so
and
then
there
is
the
custom
controllers
that
you
create.
As
part
of
your
see,
are
these
an
extension
so
yeah?
A
A
Why
we
think
that
we
are
important
because
we
are
basically
a
PI
server-
is
probably
one
of
the
most
critical
pieces
of
the
cluster
okay,
so
dramatic
this
mean
very
good.
So
let's
get
out
of
the
dark
zone,
and
let's
say
that
you
know,
given
all
these
things,
what
did
we
do
lately,
I
think
is
a
super
good
story
that
was
having
116
CR
DS
were
finally
promoted
to
GA.
A
It's
been
a
journey
for
more
than
two
years,
with
a
lot
of
people
from
the
community
involved,
agreeing
that
seer.
This
was
the
way
to
go.
I
recommend
you
to.
When
you
go
back
home
to
listen
to
Eric's
tune.
Talk
was
earlier
today
he
was.
He
did
a
very
extensive
analysis
about
the
kubernetes
resource
model
and
the
usage
of
seer
DS
across
github,
and
how
the
adoption
has
been
growing
I
can't
there
was
a
graph
that
was
really
promising.
A
So
and
I
wanted
to
highlight
this
because
there
was
a
lot
of
you
know,
it's
a
super,
complex
technical
problem.
There
was
a
lot
of
become
force
in
the
community
until
we
finally
agreed
on
what
was
the
right
technical
direction
to
go
and
what
was
the
exact
way
that
we
were
going
to
allow
versioning
of
your
series
and
conversion
between
versions
and
then
all
the
cascading.
A
It's
not
that
bad.
The
problem
is
looking
at
the
future.
How
does
it
constraint
or
or
helps,
keep
consistent
story
for
the
ecosystem
of
clients
around
kubernetes?
Imagine
that
you
know
if
somebody's
using
cube,
CDL
apply
and
the
validation
and
the
logic
is
happening
on
the
client.
Now
we
have
to
replicate
that
logic
to
all
the
clients
around
the
corner,
this
ecosystem.
So
it's
really
really
needed.
That
said,
Antoine
Palace
is
here
who
is
one
of
the
main
drivers
of
this
feature?
A
Making
it
to
beta
was
really
really
encouraging
and
important.
We
still
have
a
long
way
to
go.
We
are
dealing
now
with
basically
performance,
because
what
was
happening
on
the
Clarion
now
is
happening
on
the
server
and
in
order
to
that,
we
had
to
change
the
model
of
the
objects,
because
you
know
now
the
changes
can
be
applied
concurrently
and
you
know
we
need
to
maintain
clear
an
ownership
of
the
different
fields
of
the
object.
So
nobody's
stepping
on
top
of
each
other,
so
the
objects
are
bigger,
serialization
and
deserialization
changes.
A
So
basically
there
is
a
lot
of
work
to
do
so.
I
encourage
you
to
give
it
a
try.
Give
us
feedback
I'm,
jumping
a
little
bit
ahead,
but
apply
server
side
up.
Life
has
its
own
working
group
as
part
of
a
bi
machinery,
so
it
gets
together
once
every
two
weeks
you
can
come.
I
have
a
you
know
a
special
place
in
my
heart
for
that
group.
I
saw
particularly
in
that
work
stream
at
least
five
people
coming
from
nothing.
A
You
know
I'm
here
to
listen
to
actually
become
an
members
and
owners
of
some
of
the
parts
of
the
repo
and
coach
contributing
to
ABN
machinery.
So
yeah,
it's
a
it's
a
very
good
place
to
start.
If
you
want,
and
then
we
have
a
bunch
of
other
things
that
we've
been
moving
forward:
storage,
migrate
or
network
proxy.
For
those
of
you
who
don't
know
API
machinery
owns
something
that
is
called
SSH
tunnels
that
has
been
duplicated
deprecated
for
a
while.
We
are
trying
to
get
rid
of
it
and
to
get
rid
of
it.
A
We
are
pushing
forward
another
project
that
is
called
Network
proxy
to
replace
them.
We
need
a
bunch
of
work
to
stabilize
matrix
if
any
of
you
runs
cornutus
in
production,
and
you
have
some
kind
of
pipeline
for
your
matrix
and
visualization
charts.
You
will
notice
that
sometimes
version
2
version,
the
matrix
break
or
the
charts
break
because
matrix
are,
were
you
know
a
little
bit
of
loose.
A
What
else?
So
we
don't
own
a
CD
if
you're
interested
in
that
CD
a
CD,
it's
a
separate
project
has
its
own
repo,
its
own
community,
but
we
are
very
close
friends
with
that
CD,
basically,
because
we
are
the
only
ones
that
talk
with
that
CD
and
because
at
CDs,
part
of
is
basically
a
guest
in
our
house.
Like
all
the
time
you
know
he's
in
our
guest
room,
all
the
time
I
see
this
related
to
performance
is
related
to
upgrades
and
downgrades
of
the
control
plane.
A
So
we
are
very
close
to
the
HDD
community.
We
helped
the
release
of
a
series
three
for
a
CD.
Three
four
comes
with
a
lot
of
improvements,
especially
in
performance
when
we
were
talking
about
how
big
a
cluster
can
be
and
other
things,
for
example,
allowing
you
to
add
no
non-voting
members
to
a
net
CD
cluster,
so
the
cluster
can
be
upgraded
in
pieces
without
you
know
compromising
the
integrity
of
the
data
and
then,
as
part
of
these,
we
added
a
bunch
of
conformance
tests.
A
A
A
So
how
can
we
make
sure
that
a
malfunctioning
controller
or
a
controller
that
is
you
know,
configured
in
a
awkward
way,
is
not
taking
all
the
bandwidth
of
API
server
and
then
bringing
it
down
and
causing
an
outlet
for
the
cluster?
So
this
is
a
project
that
is
complex
enough.
We
are
working
on
this
one
and
then
the
world
for
C,
R,
D's
didn't
finish
with
GA
that
also
there
is
a
lot
of
work
that
still
needs
to
happen
there
and
one
of
the
other
working
groups
that
belongs
to
a
PA
machine.
A
Here
is
Q
builder,
Q
builder.
If
you
don't
know
it
is
the
framework
or
runtime
that
helps
you
use,
C,
R
DS
and
generate
scaffold
all
the
controllers
based
on
the
resources
that
you
are
defining,
so
they
get
together
once
a
month.
You
can
also
participate
in
that
one,
and
then
we
have
Stefan
here
and
Joe.
A
We
are
working
on
immutable
fields.
Also,
you
can
learn
more
about
that
in
the
deep
dive
and
in
the
end
we
have
28
open
caps
right
now
that
are
tagging
API
machinery.
So
you
can
see
how
much
people
is
getting
interested
in
the
in
the
center
of
cornetist
and
trying
to
do
things
which
takes
a
lot
of
our
bandwidth.
We
do
it
with
a
lot
of
joy.
Sometimes
you
know
we
wish
we
were
faster.
We
wish
we
had
more
people
to
help
us.
We
don't
so
we
do
the
best
we
can
yeah.
A
How
can
you
get
involved?
I
wish
I
had
a
better,
more
organized
answer.
You
know,
like
we
have
this
program
for
beginners
this
for
intermediaries.
You
know
we
don't
yet
I
have
the
commitment
of
working
on
that
next
year.
Next
year
is
you
know
we
still
have
a
couple
of
weeks
this
year?
What
I
can
suggest
for
now
is
if
you
are
interested
any
of
these
topics
or
things
that
I've
been
saying
rings,
a
bell
awakens
your
interest.
You
know
this
meetings
that
I'm
going
to
this
here
they
are
open.
A
A
There
is
an
agenda,
you
can
add
your
topics
or
you
can
see
the
topics
that
people
is
going
to
discuss
and
then
you
know
see
if
it's
interesting
for
you
or
not.
These
meetings
are
recorded.
Also,
there
is
a
nexus
slide.
Will
this
the
YouTube
channel,
where
all
these
meetings
are
there
a
next
one
for
me,
is
super
interesting,
it's
very
low
bar
to
green.
We
get
together
twice
a
week
for
30
minutes
and
we
go
through
everything
that
was
created.
A
Tagging
API
machinery
from
the
last
meeting,
so
this
includes
pull
requests
open
issues
we
go
through
them.
We
try
to
understand,
you
know
what
is
this
pull
request?
What
is
this
person
trying
to
merge
into
the
repo?
Sometimes
there
is
interesting
discussions.
Sometimes
there
is
follow-ups
for
the
people
if
you
ever
join
one
of
those
meetings,
sometimes
people
you
know
it's
like.
Oh,
this
issue
is
interesting
to
me,
or
this
happened
to
me
or
I
want
to
learn
more
about
I.
A
A
We
also
try
to
be
good
and
tell
the
ones
that
are
you
know
in
the
kubernetes
community.
That
is
this
good
first
issue
or
something
either
is
the
tag.
So
you
can
also
try
to
query
what
is
label
APA
machinery
and
it's
labeled.
First,
good
issue
take
a
look
to
them
nobody's
going
to
bite
you
you
can
come
to
the
channel
and
ask
you
know
if
you
think
you
have
a
solution
or
not.
My
humble
recommendation
is
probably
at
the
beginning.
A
What
is
going
to
happen
is
somebody
of
these
guys
is
going
to
say:
hey,
you
know,
looks
like
you
have
a
good
intention
figure,
but
we
need
a
test
or
what
is
this
fixing
or
what
is
the
motivation?
Sometimes
people
is
not
very
good
at
providing
context,
but
although
things
are
workable,
so
don't
be
afraid
and
then
there
is
the
working
group
meetings
those
are
recorded,
so
you
can
look
at
them.
You
can
see
that
we
don't
bark
or
bite,
so
it's
very
friendly
meetings
can
come
and
join
yeah.
A
Basically
that
and
I
really
hope
that
the
next
year
we
are
going
to
have
something
more
formal
and
we
have
an
issue
with
API
machinery,
which
is
the
ramp
up,
is
going
to
take
you
some
time
depending
on
your
dedication,
but
can
be
from
three
months
through
six
months
and
we
are
committed
to
help
people
run
pub.
What
we
need
is
commitment
on
the
other
side
that
people
is
not
going
to
walk
away,
because
you
know
it's.
This
is
basically
open.
A
Source
people
invest
from
their
heart,
their
effort,
their
knowledge
to
help
in
every
way,
run
Bob.
And
then,
if
you,
you
know
disappear
like
you
know
that
doesn't
go
anywhere
right.
So
we
are
a
little
careful
of
you
know,
being
overly
open
to
say:
yeah
I
come
and
you
know,
sit
with
me
for
four
hours
and
you
know
I
will
explain
you
everything,
but
it's
doable,
and
you
know
we
are
committed
to
that.
We
have
a
big
seed.
A
If
you
go
to
the
meetings,
you
will
see
that
in
the
bi-weekly
meetings
we
have
around
I
know
between
15
and
30
people.
Depending
on
the
topics.
The
list
has
more
than
500
people
subscribe.
The
deep-dive
was
like
packed
and
before
this
meeting,
but
we
need
more
people
to
help
us.
There
is
a
lot
to
be
done,
and
these
are
a
couple
of
useful
links.
A
We
have
two
chairs
I
recently
became
one
of
those
we
have
two
tech
leads.
One
of
them
is
here
the
other
one
is
David
who
is
like
in
the
East
Coast
right
now,
and
then
you
can
see
our
homepage,
our
Charter,
our
slack
channel.
It's
very
active
you
can
ask
questions
all
the
time,
somebody's
going
to
answer
them,
the
YouTube
playlist
and
the
mailing
group.
If
you
subscribe
to
the
mailing
group,
you
will
get
an
automatic
invite
to
all
these
meetings.
I
was
saying
so
you
know
easy,
PC
right
and
before
I
forget.