►
From YouTube: Kubernetes kops office hours 20200228
Description
Recording of the kops office hours meeting held on 20200228
A
Hello,
everybody-
this
is
cop's
office
hours
today
is
February
28
2020
I.
Am
your
moderator,
facilitator,
Justin,
Santa,
Barbara
I
work
at
Google,
a
reminder
that
this
meeting
is
being
recorded
and
we
put
on
the
internet
and
to
please
be
mindful
of
our
code
of
conduct
which
boils
down
to
being
nice
to
each
other
me
being
a
good
person.
I
am
pasting
a
link
to
our
agenda
in
the
chat.
Please
do
feel
free
to
add
your
name
and
any
agenda
items
you
would
like
to
discuss.
A
B
I've,
actually
so,
click
on
apply
if
you're
updating
a
service
which
has
two
port
entries
with
the
same
port
number.
It
applies,
it
I
think
I
found
the
solution
while
researching
the
agenda
item
and
that's
I
think
to
use
server-side
apply.
So
assuming
that
works,
does
it
make
sense?
The
switch
show
Kumari's
118
to
be
using
suicide,
apply
for
our
channels.
A
B
A
I
mean
the
other.
My
concern
is
I
I,
don't
miss
air
want
to
be
the
first
ones
do
things
and
that
it
is
not.
It
is
so
it
is
not
I.
Call
it
service
out
of
play
v2,
because
there's
two
things:
it
is
not
a
plot,
it
is
not.
The
coup
can
apply
logic
put
on
to
the
server.
It
is
not
a
rest
endpoint
you
can
host
to
that
then
cause
Cooper
it'll
apply
in,
like
you
know,
fisher-price,
like
engineering
the
there
is
a
totally
different
way
of
doing
apply
with
totally
different
semantics
and
critically.
A
Once
you
apply
v2,
you
cannot
go
back
so
once
you
once
you
pop
you
can't
you
can't
stop
and
it
would
be
I
think
it's
that
sort
of
per
object
level.
So
it's
not
the
end
of
the
world
to
say,
like
you
know
that
certain
objects
and
coop
system
namespace
are
now
going
to
use,
apply,
v2
or
with
the
cops,
and
you
can't
use
apply
v1,
but
it
is
a
it
is.
It
is
a
decision
we
should
not
take
lightly.
So.
B
Yeah,
the
thing
is,
we
want
to
remove
or
want
to
get
cored
email
start
run
as
root,
and
this
is
kind
of
blocking
that.
So
so
there's
a
desire
to
change
and
I.
Don't
can't
think
of
any
other
way
to
do
it
other
than
if
we
replace
the
service
and
didn't
use
a
plywood
that
work.
Oh
you'd
have
to
figure
out
how
to
get
channels
to
replace
the
service.
Yes,.
D
A
A
B
A
A
A
B
B
A
I
think
that's
fair,
I,
think
I
think
in
terms
I
think
we
have
made
good
progress
on
the
other
side
on
releases
and
getting
that
like
into
a
good
state.
As
of
yeah.
Yesterday
we
have
like
a
stable
release.
That's
not
too
old.
We
have
a
fader
and
an
alpha
and
that's
I
think
where
we
want
to
be
and
we're
like.
The
Alpha
is
like
the
one
that's
coming
up
in
the
next
weeks:
I
guess
so!
That's
roughly
we're
gonna
be
I.
A
Think
we've
made,
as
you
say,
like
the
ability
to
merge,
pr's
and
stuff
rapidly
has
been
that's
good
and
I.
Think
we
should
see
that
night
I
actually
had
a
look
I,
we
talked
it
like
I,
think
was
84
46,
I
camera,
which
ones
which
but
like
the
one
with
a
surge
upgrade
I
think
is
really
yeah
great
and
a
nice
trick
around
the
things
that
worked
off
loss
before
and
I
is
any
a.
A
Cool,
so
8313
I
think
is
we've
if
you're
interested
in
surge,
which
I
think
a
lot
of
people
are.
Please
do
have
a
look
at
that.
I
think
it's
a
nice
a
way
to
get
around
some
of
the
problems
we
had
before,
where
we
weren't
sure
what
words
to
a
state
and
if
Tullius
touring
the
state
john
doing
jump
in
by
like
later.
B
A
A
It
out
and
so
this
and
then
you
don't
count
those
instances,
and
that
means
that
the
the
state
is
effectively
in
the
cloud
provider
and
we
don't
have
to
mess
with
the
tart
with
the
sizes
of
the
instance
group,
which
is
sort
of
where
we
were
before,
and
that's
that
got
sort
of
challenging.
So
that's
really
nice
and
if
anyone
I
was
inclined
to
merge
it,
but
I
felt
like
other
people,
should
have
a
look
and
currently
the
only
supports
AWS.
A
B
Cloud
riders
can
implement
the
API
right,
there's
a
testing
issue
which
I'll
be
bringing
at
some
point,
which
is
I'd
like
to
be
able
to
have
a
test
that
cloud
providers
could
run
and
make
sure
that
they
implement
the
API
correctly
and
that's
not
quite
an
e
to
e
test.
That
means
run
on
ete
infrastructure.
A
Merged,
okay,
but
you've
seen
the
mocks
and
yeah
this
we
can
talk
about
in
two
weeks
and
yeah.
We
also
have
a
work-in-progress
PRI
put
up
to
like
two
tests
using
Jupiter
which
might
like
to
actually
make
it
easier
to
do
an
integration
test
which
includes
a
rolling
update
with
a
search
yeah.
So
we
have
I'm
optimistic
there.
A
B
Okay,
I'll
call,
there
is
I,
think
the
Brewers
might
want
to
prioritize
reviews
that
have
LZ
TMS,
because
I've
LG
tend
PRS
and
had
them
sit
for
weeks.
So
that's
a
good,
that's
a
good
approach
and
then
yes
kind
of
reduces
the
low
and
approvers,
because
the
reviewers
didn't
take
some
of
the
lower
hanging
fruit.
A
B
E
So
I
think
I
was
the
biggest
poser
of
this.
There
are
a
ton
of
peers
that
affect
the
API.
That's
the
cost
of
writing
code
here.
So
even
now
we
have
a
couple
of
BCD
something
metrics
and
settings
that
apply
to
the
cluster,
and
probably
people
want
to
at
least
back
port
it
in
1.7,
terapy,
so
I
think
either
we
look
at
this
merge
them
or
whatever
and
try
to
get
one
one,
seven
as
fast
as
possible.
Otherwise
we
will
keep
porting
stuff
too,
and
it
will
be
a
pain
because
automated
cherubic
won't
work
anymore.
E
A
E
A
It
does
seem
for
actually
from
that
point
of
view
like
we
are
perhaps
at
the
optimal
position
in
it.
Like
117
has
just
gone
beta,
so
we
shouldn't
be
adding
a
lot
more
API
changes
to
it.
118
is
hopefully
gonna
go
more
stable
soon,
116
is
is
released,
so
we
shouldn't
be
adding
a
lot
of
API
changes
to
it.
I
think
this
could
be
the
perfect
time,
I
think
where
we,
where
we
have
active
PRS
that
make
API
changes.
A
E
E
A
C
A
A
A
Okay,
so
I
think
this
is
a
good
way
forward,
so
I
think
unless
anyone
objects
to
the
idea
of
removing
alpha
one
which
it
sounds
like
we've
told
her
in
the
past,
it
sounds
like
we
had
previously
said
we
could
always
like
see.
If
people
objected,
it
doesn't
sound
like
anyone
is
or
has
objected,
so
we
will
release
it.
G
G
B
A
B
A
And
this
is
another
thing,
I
think.
Eventually
we
should
get
to
the
point
where
no
one
can
do
it
without
accident.
Gonna
accidentally
do
whatever
that,
and
it
is
what
I
made
it,
but
yeah
we're
not
there
currently
alright
sounds
like
there's
something
else
on
that
topic.
So
next
up,
is
it
Michael?
I,
don't
see
you
here.
E
E
E
A
Would
certainly
encourage
the
idea
that,
like
even
if
you're
running
kubernetes
one
one
six,
you
can
run
cops
117
now
I,
don't
I've
lost
track
of
where
we're
on
the
releases
in
about
the
month.
Let's
see
yeah
so
I
think
it's
it's
not
it's
not
ridiculous
to
backward
to
116
if
it's
the
most
recent
stable,
if
it's
the
current
stable
branch
and
this
does
look
fairly
low
risk,
but
I
think
we
can
have
a
look
at
it.
From
that
point
of
view,
I
don't
know.
Other
people
feel
particularly
strongly
either
way.
A
A
E
I
I
I
didn't
get
the
ow
wait
on
Justin
to
build
B
you,
a
city
manager,
image
and
once
that
is
all
done,
then
we'll
figure
out
which
release
we
want
to
land
that
it
I
can
I
can
build
that
today.
Okay
sounds
good
once
that's
then
I
came
back
for
it
and
landed
on
118,
and
then
we
can
cherry
pick
it
so
117
or
not,
but.
I
A
A
Thank
you
for
the
ping.
Sorry
for
not
doing
so.
I
think
this.
We
can
group
us
under
the
previous
discussion
of
the
need
for
timely
reviews.
I
think
the
doctor
health
check
I
think
I
circle
back
I
did
promise
a
circle
back
up
with
the
people
that
are
in
for
the
doctor
health
check
and
did
update
the
issue
with
what
I
found
I
think,
which
is
that,
yes,
you
did
it's.
E
I
think
I
addressed
your
last
comment.
There
great.
Thank
you.
One
question
regarding
the
health
check
this
week.
I
had
a
bit
of
time
and
look
at
the
note
problem.
Detector
installed
it
on
some
clusters
to
see
how
it
works.
I
see
that
it's
a
pretty
easy
to
install
head
on.
It
has
even
a
help
chart.
You
were
saying
that
we
may
want
to
have
it
in
cups.
So
I
was
curious.
Do
you
still
feel
that
it
should
be
in
cups?
A
A
We
have
cops
controller.
We
can
do
that
sort
of
thing.
We
can
build
a
separate
piece,
but
that's
that's
where
maybe
like.
We
want
to
treat
it
as
more
of
a
core
piece
or
we,
or
we
just
put
that
second
piece
in
a
separate
pieces,
the
act
to
it.
No
problem
fixer,
you
know
the
node
fixer
in
a
separate
piece
as
well.
D
A
A
E
A
H
We
use
no
problem
detector,
you
know
via
Elm,
so
I
also
think
it
could
be
a
worthwhile
thing
to
find
a
good
space
in
our
dots
to
say
they
also
suggest
you
run
this
and
you
know
point
them
to
the
helm
chart
because
we
could
add
it
as
an
add-on.
Now,
but
personally,
like
I
kind
of
think,
it's
little,
you
know,
I
know
not.
Everyone
uses
helm,
but
at
least
is
an
easy
way
than
it's
one.
Less
thing
we
could
worry
about
until
we're
ready
to
help
suggest,
employed
and.
E
Yes
right
now
it
we
tests,
have
a
pretty
simple
ready
detection
system
they
check
if
all
nodes
are
ready
and
then
that's
if
they
go
forward.
I
added
after
that,
a
cops
validate
cluster
command
with
wait.
Five
minutes
and
I
was
proposing
to
wait
for
the
validation
a
few
times,
not
sure.
If
that's
the
best
way,
I
had
some
discussions
in
the
PR
with
John
about
it.
E
So
do
we
want
to
replace
that
now
the
status
check
completely,
because
we
can
even
now
say,
hey,
wait,
15
minutes
the
same
way,
it
waits
check
for
50
minutes.
If
it
succeeds,
if
not
consider
it
that
it's
broken,
do
you
want
to
keep
it
disease
and
don't
do
anything?
I
was
proposing
to
do
this
because
I
was
hoping
that
our
checks
are
more
frequent
and
it
could
go
to
the
testing
phase.
E
B
E
Of
that,
so
it's
already
there
we
can.
The
thing
is
I
would
start
with
15
minutes.
So
that's
totally
separate
things
from
the
PR,
so
we
can
do
it
even
now.
Just
delete
that
section
and
with
what
we
have
in
current
master.
We
can
do
it,
but
it
will
validate
only
one
time.
So
if
it
gets
all
pods,
probably
in
status
running,
it
will
go
forward
with
the
tests,
not
sure
if
it's
more
complicated
the
check.
But
in
my
experience
you
could
have
a
pod
that
says
running
and
10
seconds
later.
E
B
B
E
The
power
e
to
e
o'clock
e
to
eat
pests,
don't
run
just
cops
with
the
let's
say,
cube
net.
They
run
cops
with
whatever
flavor
of
network
clogging
or
whatever
else.
We
decide
core
dns
configured
in
some
strange
way.
Maybe
and
it
could
behave
unpredictably.
This
is
why
waiting
longer
before
saying
hey
it
passed
a
few
times.
The
validation
could
catch
cases
when
these
plugins
misbehave,
like
start
early,
say
running
and
then
just
fail
a
bit
later.
B
E
B
D
B
F
B
A
Suggestion
on
the
coop
testing
I
was
liking.
It
like
if
we
ran
today's
valve
a
cluster
before
we
waited
for
the
ready
notes,
so
we
just
swapped
the
order,
the
mm-hmm
with
that
that
would
be
that
would
get
around
the
wait,
Freddy
nodes
going
in
to
back
off,
because
we
would
expect
ready
notes
to
be
ready
once
father
de
Coster
had
succeeded
so
we'd,
never
we
took
away
for
he
knows
to
be
to
pass
first
time
effectively.
So
it's
never
gonna
touch
we're,
never
gonna
back
off
and
then
and
then
I
think.
A
The
the
function
that
you've
implemented
in
terms
of
count
I
think
makes
a
ton
of
sense
to
run
in
ete,
and
then
we
could
also
optionally,
with
the
flag
perhaps
or
not
like
like
we
could
run
it
twice
and
we
could
say
if
it
validates
and
then
fails.
That
is
a
test
failure
of
cops
validate
cluster.
It's
a
little
weird
because
we're
failing
validate
not
like
you
know
not
that
BRS
but
but
yeah
I
think
that's.
If
there
was
a
bug
in
validate
that's
how
we
catch
it.
Yeah.
B
A
Have
to
make
sure
we
have
those
on
there,
but
ok,
yes,
yes,
but
I.
Think
so.
I
think
the
the
thing
I
think
I'm
aware
of
is
if
a
deployment
is
slow
to
sketch
if
their
parts
haven't
been
scheduled,
yet
haven't
even
created
for
whatever
reason
like
you
could
have
deployment
at
0
of
3
no
pods
yet
theoretically,
and
it
would
pass
validation
because
we
haven't
checked
that
the
deployment
has
created
the
pods
as
creating
pods,
possibly
if
there
are
more
things,
but
no
any
cluster
needs
to
check.
I
would
like
to
know
about.
A
H
E
A
A
E
G
A
D
A
A
G
After
some
testing,
it
turns
out
that
there
is
effectively
a
breaking
change
in
it
in
the
that
previously,
you
provide
a
config
file
in
the
container
via
I
amount,
and
they
had
its
support
for
also
recognizing
config
files
through
CR
DS,
as
well
as
like,
detecting
and
config
Maps,
and
so,
rather
than
having
the
default
behavior
still
look
for
that
config
file.
They
it's
now
trying
to
look
for
CRTs
and
ignoring
the
config
file,
even
if
we
specify
it.
So
that
means
without
changing
any
of
the
command
arguments.
G
It's
now
no
longer
functional
and
so
I
was
looking
at
possible
solutions
to
add
the
new
beckoned
mode
argument.
It'll
only
work
on
0
5
0,
so
we
would
need
to
and
we
allow
through
the
API.
We
allow
users
to
specify
an
image,
so
that
allows
them
to
use
older
versions
if
they
want.
So
that
made
me
cautious
of
just
hard-coding
the
new
back-end
mode
argument.
G
Another
option
would
be
trying
to
do
some
sort
of
semantic
version:
personing
parsing
of
the
image,
but
because
people
can
provide
their
own
I,
don't
think
that
would
be
sufficient
and
so
another
option
I
talked
with
Rodrigo
about
was
adding
a
new
API
field
that
would
allow
users
to
specify
it,
but
I,
don't
for
that
same
reason
about
hard
coding.
It
I
don't
know
if
we
could
effectively
set
a
default
value
for
the
API
field.
A
A
A
A
In
the
API
field,
and
then
probably
in
the
release,
notes
as
well,
that
doesn't
mean
we
can't
read
upgrade
their
if
they're
running,
if
they're
running
an
existing
image
and
haven't
specified,
it
means
I,
messed
up
yeah
we
need
to.
We
need
to
think
this
through
very
carefully
I
guess
in
terms
of
yeah.
A
Yes,
the
approach
we
took
the
island
is
how
we
basically
dealt
with
this
so
far
in
terms
of
like
existing
clusters
create
a
field.
The
optional
value
is
the
legacy
behavior
and
then
create
cluster
will
set
the
new
value
on
created
clusters
so
that
or
a
newly
created
cluster,
so
that
users
don't
have
to
like
do
anything.
D
A
G
D
A
G
A
A
B
So
currently,
our
duck
site
is
publishing
the
dots
that
are
checked
in
on
masters.
So
when
someone
adds
a
feature
and
master
and
then
they
document
it,
and
someone
sees
that
and
wonders
why
it
doesn't
work
because
they're
using
AG
a
version
of
cops.
So
now
people
are
putting
in
the
documentation
that
this
feature
is
now.
You
know,
as
of
118,
and
the
in-game
of
that
is
that
yet
your
Doc's
ridiculously
document,
which
version
each
feature
got
added
in
so
which
is
kind
of
hard
to
read.
B
H
You
probably
mostly
meaning
yeah
I,
talked
to
a
couple
groups
on
this.
I
know
that
for
kubernetes
kubernetes
they
actually
have
independent
sites
every
time
they
do
a
release.
They
spit
up
a
new
site.
That
way,
then
you
know,
and
then
it's
linked
to
the
release
branch
of
that
site.
I
think
that's
a
little
too
much
overhead
for
cops
personally,
and
you
know
I
think
that
that's
I
mean
it
would
be
simple.
You
know
tying
things
to
a
release
branch
and
just
living
off
of
a
specific
URL
for
that.
H
H
H
H
A
A
That's
true
mortuary
thinks
that's
the
you
did
describe
I,
don't
know
how
other
people
feel
you
did
describe
the
from
version,
this
as
being
a
sort
of
anti
pattern.
I
I
personally
find
it's
a
little
nicer
than
when
I
click
through
onto
like
a
certain
C
and
I
providers
website
and
like
land
on
some
random
version.
That
was
whatever
I
guess:
Google
decided
to
index
at
that
time
and
then,
like
it's,
not
the
latest
version,
but
you
can.
A
H
A
Yeah,
like
some
sort
of
block
like,
but
and
so
that
we
can
also
drop
them
off,
as
we
do
start
deprecating
like
like
to
say
like
available
in
cups,
one
six,
it's
like
that
shouldn't
be
in
the
release,
notes
anymore,
right
right,
we're
in
the
docs
anymore.
That's
just
like
okay
read
but
like
available
as
a
as
a
cops
1:18.
That's
interesting
information.
Now.
G
A
We
could
probably
start
by
not
not
dangling
the
fancy
shiny
features
of
118
in
front
of
the
users
that
are
not
on
my
team
by
pegging
to
a
release,
branch,
I,
guess,
yeah,
that's
a
good!
That's
a
good!
First!
First
idea:
we
have
one
item
left
and
five
minutes
left
I
will
move
us
on.
If
that's
all
right.
A
C
I'm
I'm
working
on
some
of
the
and
n
testing.
The
goal,
which
is
not
important
to
this
particular
question
I,
have
is
to
get
GCE
and
and
tests
like
enabled
and
get
that
going
for
PRS.
However,
I
realized
that
all
of
our
antenna
tests
are
in
sync
cloud
providers,
AWS,
and
there
are
GCE
tests
in
there,
which
is
a
little
weird
I
proposed
a
change
but
I,
don't
think
anybody's
gonna
do
anything
until
Justin
thumbs
up
or
thumbs
down
on
it.
C
It's
a
it,
basically
just
the
lift
and
shift
right
now
to
move
it
into
safe
cluster
lifecycle,
which
I
think
more
properly
describes.
What
we're
doing
at
this
point
there
that
that
one
I
think
isn't
easy,
is
I,
don't
think,
there's
going
to
be
much
competition
to
it,
but
I
could
be
wrong.
My
second
question
is
about
consolidating
the
presubmit
checks
and
I.
Don't
know
if
we
can
do
that
into
thee,
because
we
now
have
tests
in
kubernetes
cops
and
we
have
them
in
kubernetes
sync
cloud
provider
AWS
in
the
testing
for
every
phone.
C
A
C
A
A
A
C
M
C
C
A
I
don't
fit
I
put
it
I,
don't
think
it's
letting
us
on
the
release
plan.
That's
important
I
think
we
might
actually
got
to
create
some
buster
a
mice.
This
week,
I
told
that
image,
people
and
I
will
do
another
180.
Now,
if
there's
anything
but
otherwise,
I
don't
have
anyone
has
any
last
topics
or
releases.
They
would
like
to
see.
E
Not
releases,
but
maybe
in
the
same
topic
is
testing
from.
Maybe
someone
from
let's
say
this
party
could
get
added
to
the
owners
file
where
cops
go
is
in
testing
front
me
and
Peter
have
been
doing
some
changes
there
and
it's
pretty
annoying
creeping
the
week
or
so
for
an
approval
or
something
that's
only
related
to
cops.
Okay,
yes,.