►
Description
Kubernetes Enhancement Proposal (KEP) Reading Club is an initiative by sig-architecture.
KEPs covered in this session:
- https://github.com/kubernetes/enhancements/pull/2640
- https://github.com/kubernetes/enhancements/tree/master/keps/sig-node/2000-graceful-node-shutdown
A
B
A
A
A
Okay,
sorry
about
that,
let
me
just
give
like
a
brief
overview,
so
what
happens
is
every
week
every
two
weeks
we
have
this
session
and
people
such
as
caps
that
are
of
interest
to
them
or
if
they
are
an
author
of
the
cap,
they
suggest
the
cap
for
us
to
like,
go
through
and
read
and
understand
and
provide
feedback
on
if
we
can.
Otherwise
it's
just.
If
you
find
a
cap
interesting,
you
suggest
that,
and
we
all
read
it
together
and
discuss
it
later
on
and
kept
this.
This
has
been
going
like
said.
A
A
lot
so
kept
stands
for
kubernetes
enhancement
proposal,
it's
the
new
mechanism
for
getting
new
features
in
or
deprecating
older
features,
or
even
deleting
all
difficulties
or
updating
current
features.
So
that's
the
whole
process
of
getting
those
things
into
the
kubernetes
project,
and
the
purpose
of
this
is
basically
to
like
familiarize
ourselves
with
the
whole
process,
and
things
like
that
and
kepler
is
like
the
reading.
Our
character
is
the
best
way
to
know
familiarize
ourselves
with
the
future.
Okay.
A
E
A
B
A
A
E
E
C
C
E
I
have
a
very
general
question
not
like
before,
starting
with
the
cap:
what
do
what
do
we
mean
by
feature
gate
because
yeah
there
is,
we
are
talking
about
protecting
secrets
and
config
map
in
two
cases
in
the
entry
and
the
outre
resources,
and
one
of
them
is
enabled
through
a
feature
gate
and
the
other
one
is
through
a
flag.
So
what
do
we
really
mean
by
feature
gate
in
general.
A
I
can
probably
take
like
what
is
a
feature
gate,
but
I'm
not
too
sure
about
the
distinction
between
the
flag
enabling
through
the
flag
and
enabling
to
the
feature
gate
for
the
entry
and
others
so
feature
gate
in
general.
Is
it's
like
a
mechanism
for
you
to
enable
features
that
are
probably
like
alpha
or
beta?
So
if
a
feature
is
an
alpha,
then
the
feature
gate
for
that
would
be
by
default
set
to
false.
So
you
won't
be
able
to
use
that
feature
in
the
binaries
that
come
out.
A
And
if
a
feature
is
in
beta,
then
that
feature
gate
by
default
is
set
to
true,
meaning
that
if
you
want
to
disable
it,
then
you
would
have
to
explicitly
go
over
and
disable
it.
And
when
a
feature
is
in
beta,
it's
usually
meant
to
like
get
further
feedback
from
the
community
and
like
continually
improve
on.
A
So
that's
why
it's
by
default
set
to
true
and
it's
at
a
relatively
stable
state
now
when
it
finally
reaches
ga
after
getting
feedback
and
increasing
consistent
consensus,
and
things
like
that,
when
it's
in
ga
you
don't
really
need
a
feature
flag
over
there,
it's
it's
there
in
kubernetes
by
default,
you
can't
really
disable
it
and
yeah
it's
there.
Basically,
you
want
to
see.
A
D
A
E
A
A
E
E
So
there
is
a
situation
there
and
we
are
considering
this
as
one
of
the
blocking
conditions,
because
here
the
user
would
have
to
go
manually
and
delete
the
finalizers
to
delete
the
secrets
up
so,
like
our
design,
details
actually
takes
care
of
this
particular
situation
or
it's
it's
a
known
issue.
At
the
moment.
It's
a
broker.
D
E
A
A
D
D
D
E
A
A
A
A
C
E
So
we
are
talking
about
adding
a
new
new
config
filled
shutdown
grace
period
so,
and
we
are
talking
about
having
a
graceful
shutdown
for
pods,
no,
not
at
the
moment.
So
my
question
was
like
if
I
am
understanding
it
properly,
we
are
talking
about
adding
a
new
film
in
the
cubelet
conflict
right,
yeah,
okay
and
I
have
a
follow-up
question
which
is
here
we
have.
We
are
talking
somewhere
like
how
we
are
going
to
calculate
the
time
like
for
how
long
it
will
wait.
E
B
E
A
E
C
So
what
I
wanted
to
ask
was
like
it
mentions
that
the
reason
this
needs
to
be
done
is
so
that,
like
the
pods
can
be
gracefully,
shut
down
and
like
to
have
some
buffer
time.
What
I
wanted
to
know
was
like
does
anyone
know
of
any
examples
like
where
pods
like
need
to
do
something
before
they
shut
down
and
like
which
is
essential,
which
is
what
we
are
like
trying
to
provide
here.
A
So
one
thing
that
I
can
think
of
is
let's
say
you
are
running
some
application,
which
sort
of
opens
or
like
creates
certain
resources,
but
once
you
want
it
to
shut
down,
you
want
to
make
sure
that
those
resources
are
cleaned
up
in
the
way
that
you
define
so,
for
example,
like
I
can't
think
of
a
like
a
very
negative
effect
sort
of
use
case
happening.
But
I'm
sure
there
is
something
that.
A
Yeah
yeah
like
when
you
want
things
to
like,
and
also
most
importantly,
when
you
want
the
bot
to
follow
the
life
cycle
that
you
expect
it
to.
That
is
like
scheduled,
running,
then
terminating
and
so
on.
So
you
wanted
to
follow
the
normal
life
cycle
rather
than
it
being
abruptly
terminated
so
and
this
this
the
next
thing,
I'm
about
to
say
I'm
not
sure
of
but
another
implication
of
that
I
think
could
also
be
and
like
anyone
can
hear,
can
correct
me
right
wrong.
A
But
if
you,
if,
like
a
note
shuts
down
and
all
the
pods
they're,
also
shut
down
abruptly-
and
let's
say
you
didn't-
want
some
pot
there
to
shut
down
then
to
recreate
that
pod
and
reschedule
it,
you
would
have
another
additional
check
over
there,
rather
than
it
rather
than
like.
If
you
follow
the
normal
life
cycle
of
it,
then
through
the
normal
life
cycle,
you
would
get
know
that
something
is
shutting
down
and
then
in
the
normal
workflow
of
things
you
could.
A
E
I
I
have
just
one
one
question
or
a
one
open
question
or
one
point
to
make
so
number
one
like
doubt
an
open
question,
so
here
by
adding
a
graceful
not
not
shut
down,
we
are
making
sure
that
the
pods
have
time
to
shut
down
properly
or
something
like
that
right.
We
are
trying
to
help
the
at
the
pod
level.
With
this
with
this
node
graceful
shutdown.
E
E
Okay,
okay,
so
I
I
was,
I
was
saying
like
there
is
something
known
as
poor
disruption
budget,
which
does
the
same
thing
so
where
poor
disruption
budget
is
used,
one
of
the
examples
could
be
when
we
are
upgrading
a
cluster
or
in
those
cases.
What
really
happens
is
we
need
to
up
like
we
are
upgrading
the
platform.
So,
for
example,
we
have
to
upgrade
the
kubernetes
version
itself
on
a
production
cluster.
E
So
what
we
do
is
we
we
try
to
gracefully,
do
it
node
by
node,
so
we
we
try
to
drain
one
node,
and
then
you
like
put
this
put
the
ports
on
different
node,
but
in
that
situation
this
kind
of
having
graceful
board
shutdown
helps
because
we
don't
want
to
right
away,
delete
everything.
That's
scheduled.
We
need
to
first
check
like
what
is
important
and
what
needs
to
be
scheduled
on
a
different
node
first
before
just
deleting
it
right
away.
E
C
B
A
A
E
E
Yeah
yeah
yeah.
That
was
my
question
like
when
I
was
reading
it
that,
in
this
case,
when
we
are
when
we
are
implementing
this
feature
like
that
feature,
would
take
care
of
the
port
disruption
budget.
If,
if
any
is
set
like
it
would
take
care
of
that
or
it
was
just
right
away.
Wait
for
that
particular
minimum
seconds
that
we
are
introducing
and
then
right
away
yeah.
A
A
A
E
A
A
A
A
C
A
B
B
D
A
A
E
And
and
I'm
just
reading
in
the
official
documentation
under
the
pre-stock-
and
there
is
a
very
interesting
line
there-
it
says
the
pods
termination
grace
period
countdown
begins
before
the
pre-stock
hook
is
executed.
So,
regardless
of
the
output
of
the
outcome
of
the
handler,
the
container
will
eventually
terminate
within
the
ports
termination
grace
period.
So
it
means
like
the
the
pre-stop
logic.
C
I
think
it's
like
we
are
trying
to
trigger
this
pre-stop
hook.
If
so
that,
like,
if
there's
like
an
immediate
something
happens
in
the
note,
then
the
pre-stop
hook
won't
get
triggered
right,
because
the
note
didn't
know
that
it
had
to
shut
down
stuff.
At
least
that
is
what
my
understanding
was
that
this
would.
A
So
this
graceful
shutdown
is
to
delay
the
shutdown
of
the
node,
as
in
like
the
actual
box
or
the
machine
by
some
amount,
so
that
in
that
time
frame
you
can
actually
terminate
these
pods
and
those
containers
in
those
forms.
So
that's
why
you
have
that
graceful
shutdown
minus
something
as
the
grace
period
of
the
pod.
So
that
is
the
gray
sphere
that
is
stopped
above
in
the
cockpit.
C
E
So
so
we
are
saying
at
the
moment,
even
though
there
is
a
there
is
a
graceful
domination
like
grace
period
at
the
moment,
but
because
the
node
doesn't
have
any
at
the
moment,
so
it
doesn't
like
it.
The
poor,
dis
port
termination
doesn't
happen
properly,
so
we
are
trying
to
help
board
at
the
moment.
It's.
C
E
Yeah
I
mean
that
that's
what
the
impression
I
got
after
reading
like
we
are
switching
off
the
machine,
the
power
button.
So
I
was
thinking
in
my
like
when
the
power
button
is
off
to
me.
Really
everything
is
gone.
So
what
like
you
ask
what
what's
the
use
case
we
are
trying
to
help
her.
That
would
be
interesting
to
know
like
what's
the
example.