►
From YouTube: Kubernetes SIG Testing 2017-08-15
Description
Meeting notes: https://docs.google.com/document/d/1z8MQpr_jTwhmjLMUaqQyBk1EYG_Y_3D4y4YdMJ7V1Kk
A
Hi
everybody-
this
is
the
kubernetes
state
testing
weekly
meeting
today
is
Tuesday
August,
the
15th
we
are
being
publicly
recorded
and
I
will
post
this
to
mr.
Singh
testing.
Yes,
I
am
mr.
Singh
testing.
It
will
be
posted
to
YouTube
as
soon
as
we're
done
here
and
and
with
permitting
today.
I
just
wanted
to
talk
real
quickly
about.
We
we're
here
at
the
new,
my
class
them
I'm
going
to
share
screen
real
quickly.
Let's
just
see
the
whole
desktop.
Why
not
so
code
freeze
is
coming
code.
A
A
Basically
I
don't
want
to
hash
this
out
in
this
meeting,
but
I
think
that
we
should
probably
take
a
look
at
some
of
the
items
that
we
have
this
milestone
and
decide
what
they're
not
they're
relevant
I've
been
catching
up.
I've
been
underwater
of
like
hundreds
of
notifications,
I've
been
catching
back
up
and
I.
Think
a
lot
of
great
stuff
has
happened
over
the
past
couple
of
weeks,
some
of
which
has
been
reflected
in
this
milestone,
some
of
which
might
obviate
things
in
this
milestone,
and
some
of
which
may
have
made
things.
A
A
smile,
spend
less
important.
I
noticed
that
Eric
and
a
couple
other
folks
have
been
good
about
circling
back
and
closing
out
issues
lately,
just
as
we're
making
progress,
I
think
some
really
cool
things
have
happened.
I'd
like
to
spend
a
little
bit
of
time
next
week
after
having
gone
through
this
and
sort
of
cleaned
it
up,
making
sure
that
we're
ready
for
code
freeze
and
the
reason
I
bring.
A
That
up
is
because
at
some
point,
there's
probably
some
amount
of
stabilization
we're
going
to
want
to
apply
to
the
actual
infrastructure
used
to
run
tests.
I
think
there
were,
there
has
been
some
in
the
past.
We
have
continued
to
change
the
way
that
tests
are
run
and
deployed
and
continue
to
change
things,
while
there's
a
code
freeze
going
on
and
that
can
cause
some
confusion
for
people
who
are
looking
for
sense
of
stability
to
burn
down
through
the
release.
A
So
I
just
like
us
to
sort
of
keep
that
in
as
we
move
forward
any
questions
on
that:
okay
cool.
A
That's
it
from
me
by
the
way,
quick
shout
out
to
the
fact
that
we
no
longer
have
clusters
gke
directory
in
kubernetes,
and
all
of
that
has
been
totally
ported
over
to
to
test.
Also
shout
out
to
the
fact
that
cops
testing
is
now
bone
under
the
same
kubernetes
e
to
e
scenario
has
the
rest
of
our
HIV
testing.
There's
been
a
lot
of
really
awesome.
A
Work
done
that
maybe
doesn't
sound,
amazingly
cool
to
the
community,
but
it's
reduced
a
lot
of
technical
depth
for
us
and
its
really
helping
us
aim
at
reducing
their
play.
Crate.
So,
thanks
for
all
that,
hard
work
with
that
I
will
hand
over
to
Quinton
who
just
posted
a
cool
doc
in
the
chat
button.
B
B
B
B
D
Answer
is
maybe
perfectly.
B
Lot
of
ideas
have
gone
around
on
how
this
can
be
on
giving
local
vm's
is
a
popular
one,
running
just
fair
processes.
Mating
the
colon
is
a
popular
one.
So
what
I'm
proposing
is
that
we
do
net
to
soccer
soccer,
so
the
idea
is,
is
you
spend
up
original
level
a
stain
running
a
document
and
money?
Do
it
Wow?
This
is
just
aparment.
B
It
didn't
I
rather
either
as
a
node
for
its
master
and
when
you
dive
in
a
neck
and
going
to
initials
everything
and
all
of
these
containers
running
in
one
single
chicken
stock
in
the
yurt
on
Walker,
this
doesn't
make
sense.
Yeah.
A
A
D
Okay,
but
now
I.
B
I
feel
this
robot
reboot
volume
of
data
all
right,
there's
a
digital
matrix,
yeah
I,
don't
know
what's
going
on.
This
is
unfortunate.
I'm
going
to
go,
get
a
headset
I'll
be
right
back.
We
haven't
solve
our
problem.
Okay,.
F
A
Antis
worked
on
that
I
did
kind
of
a
variant
about
that
I
kind
of
abandoned
for
now,
anyway,
called
Ncube.
It
actually
runs.
The
clusters
on
kubernetes
itself
should
be
on
mini
queue,
but
the
idea,
then,
is
that
you
kind
of
have
an
automated
way
to
manage
your
individual
clusters.
You
can
have
as
many
as
you
want
on
a
given
cuber
Nettie's
cluster,
and
you
can
scale
the
nodes
by
scaling.
Changing
the
deployment
replica
count
that
kind
of
thing
but
I'm
kind
of
concerned.
A
If
he
said
he's
done
this
independently,
because
this
is
his
prior
out
there.
That
probably
has
gone
a
lot
further
I've
been
using
the
the
MIRR
antis
cubed
min
in
cluster
as
well
and
I've
been
trying
to
modify
it
to
work
with
ipv6
so
that
we
could
do
some
ipv6
testing
I
can
either
be
too
hard.
I
mean
we
had
something
similar
for
open
chef
and
I
mean
supporting
ipv6.
As
long
as
your
host
stack
supports,
that
I
mean
supporting
a
nested
is
mainly
a
rocker
and.
B
Far
so
good
okay
cool
switched
over
to
a
headset.
So
let's.
C
A
B
Yeah,
okay,
well
I'm,
just
gonna
I'm
not
going
to
start
talking
and
if
it's
intolerable
tell
me
I
think
it's
intolerable.
If.
H
B
Right
so
so,
today,
with
testing
kubernetes,
we
have
two
main
strategies:
we
have
unit
tests
and
in
the
end
tests
and
the
unit
tests,
you
know
they're
pretty
fast.
Unfortunately,
they
don't
give
a
whole
lot
of
assurance
and
the
reason
this
is
is
because
kubernetes
is
fundamentally
a
series
of
side
channels
by
design
and
the
unit
tests
have
optimized
have
a
hard
time
having
confidence
in
the
system
as
a
whole.
B
B
So
what
I
want
to
do
is
I
want
to
create
a
form
of
testing
that
is
as
fast
as
unit
tests,
but
gives
many
of
the
assurances
of
the
into
end
tests,
specifically
with
all
of
the
different
kubernetes
components,
talking
together
all
right.
So
the
way
that
I
have
been
prototyping.
This
is
using
nested
docker,
and
so
what
I
do
is
that
I
create
a
top-level
docker
container
and
I
run
that
container
this
container
runs
docker
and
docker,
and
it
does
docker
runs
for
kubernetes
nodes,
so
it
will
do.
B
It
will
then
start
a
node
that
it'll
start
running
a
container
that
runs
both
a
runtime
which
is
docker
and
a
cubelet
talking
to
dr..
So
it's
a
container
running
more
than
one
process
and
then
on
this
container
after
the
qubit
comes
up,
I
run
cube
admin
and,
depending
on,
if
it's
master,
I
run
an
it
where
I
run
join
and
at
the
end
you
have
a
fully
functioning
cluster
Solms
overlay
Network.
B
A
They
basically
so
originally
there
was
kubernetes
on
docker
and
I.
Think
that
was
dr.
Stephan
did
that
and
then
morontia
sort
of
took
that
ran
with
it
and
modified
its
support,
cube
ATM.
So
it
does
basically
everything
you're
talking
about
and
it's
been
working
for
over
a
year.
So
a
lot
of
the
issues
that
you're
probably
encountering
has
colors
and
fix
because
people
are
using
it
productively,
both
in
testing
and
for
development.
So.
B
I'm
not
sure
what
they've
done
I've
seen
that
other
people
like
just
a
couple
days
ago,
someone
someone
else
had
done
something
very
similar
are
doing
this,
but
one
important
thing
here
is
that
this
is
tied
into
the
kubernetes
build
system.
So
when
I
build
kubernetes
like
today,
I
could
say
you
know,
do
I
just
make
a
release.
I
also
want
to
be
able
to
just
make
integ
tests
right.
A
But
that's
kind
of
porcelain
on
top
I
know
it's
just
effort
actually
does
build
kubernetes
like
in
a
docker
container
and
can
redeploy
itself.
So
it's
like
in
terms
of
getting
working
cluster
out
of
a
source
tree.
They
pretty
much
have
that
locked.
If
there's
you
know
end-to-end
automation,
you
need
on
top
of
that.
Maybe
you
could
adapt
your
work
to
their
stuff,
but
I
would
be
concerned
if
you
were
just
duplicating
a
lot
of
the
cluster,
bring
out
and
build
stuff
that
they
already
have
in
place.
A
A
My
name
is
guru,
infrared,
okay,.
B
Cool
all
right,
so
if
they
have
something
working,
that's
better
than
what
I
have
I
would
absolutely
just
pull
it
into
a
kubernetes
mainline
and
run
it
there.
I'll
have
to
look
into
specifically
what
they
have.
A
G
B
So
what
I'm
doing
look?
Let
me
explain
the
point
that
I'm
trying
to
get
across
here
is
not
look
at
this
amazing
thing.
I
ran
kubernetes
as
docker
and
docker
right
like
that's.
That's
not
what
I'm
getting
at
at
all
here,
I'm,
not
saying
everyone
should
go
use
this
thing,
I
just
hacked
up
recently.
What
I'm
trying
to
get
at
is
I'm
trying
to
say
we
need
a
better
standard,
consistent
way
of
testing
these
things
locally
right.
B
B
And
so
like
I
said,
I
was
not
familiar
with
mer
antis
when
I,
when
I
was
looking
into
this
I
did
not
find
there
and
I'm
still
actually
having
trouble
finding
it
right.
Now
there
there's
a
link
posted
in
Jah
and
there's.
B
B
Yes,
I
saw
a
cube.
Spawncube
spawn
I
saw
just
a
couple
days
ago
and
Moran,
just
okay,
I.
A
B
B
A
I
mean
when
you
say,
merge
into
kubernetes
I
thought
that
that
sort
of,
like
I,
think
when
I
mean
because
he's
presented
it
at
the
community
meetings
he's
presented
in
cig
testing
and
I'm.
Definitely
in
your
camp
like
I,
think
this
is
the
preferable
way
for
the
majority
of
people
to
do
testing,
but
I
think.
Maybe
one
of
the
caveats
is
that
a
lot
of
people
simply
don't
care
about
multi-node
very
much.
I
know
that's
kind
of
sorry.
B
A
Much
force
agree
with
you
I'm,
just
trying
to
say
that
my
experience,
because
I
come
from
a
networking
side
of
things
where
you
need
multiple
nodes
and
so
I
built
tooling
before
any
of
this
for
open
ships.
That
did
something
similar
and
outside
of
our
networking
group.
We
didn't
get
a
lot
of
interest
and
I
kind
of
saw.
The
same
thing
happened
with
darker
and
darker
cluster
people
who
cared
about
you
know,
simulating
multiple
nodes
and
running
tests
in
that
way
cared
about
it,
and
that
was
a
very
small
segment
of
the
community.
A
So
it
didn't
get
a
lot
of
traction
like
if
you
could
convince
people
that
they
should
do
this,
and
this
should
be
the
default.
I
mean
I'm,
certainly
in
agreement.
Ocean
short
in
is,
and
many
other
people
in
the
sig.
But
the
challenge
is:
it
does
have
a
certain
out
of
complexity,
and
it
does
have
a
certain
amount
of
sort
of
caveats
that
it
seems
that
I
don't
know
if
they're
insurmountable,
but
they
seem
to
have
been
kind
of
they've,
slowed
or
prevented
people
from
really
buying.
B
Into
it,
I
mean
III
think
that
I
think
the
important
thing
is
if
people
can
run
a
command
like
a
make
local
integration
test
and
spend
of
a
cluster
and
run
a
test
or
make
local
cluster
and
spin
up
a
local
cluster.
I
think
that
is
the
bar
that
the
application
developer
wants
and
if
a
testing
product
does
not
provide
such
a
simple
way
of
using
it,
then
you're
always
going
to
suffer
with
adoption.
B
A
B
Be
nested
arbitrarily
deep,
if
you're,
if
you're
doing
the
correct,
if
you're,
taking
the
correct
precautions
and
as
long
as
as
long
as
you're
only
focused
on
testing
cluster
orchestration
and
not
testing
container
runtimes
with
the
qubit,
and
you
can
just
use
one
container
runtime
like
docker
I.
Don't
know
why
the
people
adopting
the
testing
framework
need
to
know
or
care.
One.
H
Of
the
things
we
run
into
an
origin
with
the
doctor
and
doctor
stuff
that
Maru
built
was
there's
a
lot
of
really
hairy
interactions
between
slightly
versions,
cute
Dockers
and
so
you've
got
a
doctor
I'm
the
host
here
in
slowing
docker.
In
the
container.
We
ran
to
a
whole
slew
of
problems
that
came
out
of
that.
H
A
Are
things
that
you
can
probably
work
around,
but
the
takeaway
is
it's
not
like
it?
Just
works
right
I
mean
right.
That
was
our
conversion,
a
docker.
You
need
a
certain
version
of
Dakin
and
docker
if
they
skew,
for
whatever
reason
say
the
developer
happens,
to
have
a
version,
you
don't
expect
unexpected
things
can
happen,
and
so
it's
like
it's
kind
of
a
support
issue.
It's
if
you
have
enough
resources
enough
people
willing
to
fix
it.
So
so.
B
A
A
Container
run
time,
it's
that
certain
capabilities
that
you're
trying
to
use
may
break.
If
you
have
SKU
and
the
darker
version
yeah.
If
they
have
a
lower
version
of
docker
up
above
you
mean,
or
they
have
a
newer
version,
it
could
be
either
way
like
darker
the
way
that
they
validate
like
they
don't
support.
Darker
and
darker.
They
run
their
tests
and
dr.
docker
actually
sure.
But
if
you're
running
a
configuration
that
they
don't
test
themselves,
then
you're
doing
it.
You're
just
you'll
find
the
problem.
You
have
to
fix
it
and
darker
may
not
be.
B
A
Yeah
I'm,
how
could
you
do
that?
I
guess
like
from
from
my
experience,
getting
the
solution
in
place
and
testing
it
and
CI
and
proving
it
reliable,
I
think
that
would
be
sort
of
the
starting
point
for
saying
that
developers
should
use
it.
If
you
don't
run
into
you
know,
if
you
can
consistently
run
tusks
and
you
know
give
an
environment
using
darker
and
darker,
then
awesome,
then
let's
try
to
put
that
in
developers
hands
and
make
them
more
productive,
because
I
absolutely
think
that
it
would
make
a
large
number
of
people
more
productive.
A
A
A
But
that
has
led
people
to
believe
that
need
to
be
linking
you
to
be
tests
are
the
only
things
that
we
should
be
paying
attention
to,
and
so
one
thought
is
that
perhaps
by
using
the
soccer
and
doctor
mechanism,
whichever
implementation
you
choose,
we
could
still
basically
run
the
same
e2e
tests.
But
again
something
that's
significantly
faster
and
we
generate
a
lot
of
or
lose.
B
These
the
second
part
of
what
I
want
to
get
at
is
not
just
a
that.
Conformance
test
should
be
run
against
this
because
that's
fairly
obvious,
but
rather
that
the
vast
majority
of
our
tests
outside
of
a
couple
of
specific
cases
like
cloud
providers
and
storage
providers
which
are
moving
out
of
the
repo
anyway,
could
be
run
against
such
an
environment
right.
A
Yeah
now
I,
don't
think
that
the
goal
of
you
know
making
tests
more
reliable
or
faster
is
contentious
in
any
way,
I
think
I'm
kind
of
hesitant
to
say
this
is
like
a
silver
bullet.
I.
Think
our
economy
test
suite
a
lot
of
those
tests
you
shouldn't
be
in
to
end.
It
should
be
targeting
like
an
integration
environment,
because
they're
only
really
validating
API
interaction
and
doesn't
really
require
nodes.
A
B
B
A
A
B
B
A
I
guess
I'll
have
to
agree
to
disagree
with
that,
because
it's
arpeggiate,
I'm
proposing,
seems
to
work
fairly
well
in
the
Federation
case
and
I've
seen
similarities
and
other
tests
like
I
said
it
depends
on
the
test
I
but
I
the
idea
that
one
style
of
test
just
solves
all
problems,
and
it
will
agree
with
that.
This
is
a
component
in
a
more
comprehensive
strategy.
A
We
correct
me
if
I'm
paraphrasing
incorrectly
here,
but
we
basically
don't
want
to
expose
doctor
on
the
host
to
prowl
Jones
for
positive
run
on
behalf
of
cron
jobs,
because
that
could
be
seen
as
a
security
vulnerability.
So
the
issue
then
becomes.
How
could
we
leverage
this
approach
while
still
allowing
crown
so
you
take
this
sort
of
stuff
off
I
mean?
Maybe
maybe
we
want
to
think
about
balance
so.
B
So
that
then
the
answer
there
is
I
believe
nested
docker
right.
So
so,
when
you're
running
nested
doctor
you're
not
ever
talking
to
the
host
docker
instance
you're
talking
to
your
own
doctor
instance,
you
may
have
to
be
spun
up
with
special
privileges,
which
you
know
creates
other
other
problems.
But
you
don't
actually
have
to
talk
to
my
host
doctor
ever.
A
Coletti
Nessa
talker
as
a
general
solution
for
this
class
of
problem
for
proud,
would
be
interesting.
You
know
the
other
big
thing
that
people
kind
of
look
to
this
project,
for,
but
don't
yet
have
is
automated,
builds
of
all
of
the
images
for
the
various
add-ons
and
test
components.
When
you
push
a
commit,
you
generally
have
to
go
chase
after
some
Googler
to
actually
fill
the
image
and
push
it.
The
GCR
dot,
IO
and
it'd
be
great.
If
we
can
automate
these
as
crops
being
able,
urges
g-cloud
docker
build
geoduck
push.
A
A
So
if
I
were
where,
as
this
is
perhaps
an
attempted
taking
existing
corpus
of
tests
in
test
e
to
e
and
finding
a
way
to
run
them
the
slightly
faster
case
against-
and
it's
not
to
reproduce
as
much
as
the
fidelity
as
possible
of
a
total
extra
Nettie's
cluster
and
those
are
alternate
approaches
at
finding
a
faster
way
to
get
the
same
level
of
confidence,
bubbling
less
set
up
and
I.
Think
that
is
an
alternate
approach
site
that
can
be
explored.
A
Yeah
I
was
really
hoping
for
both
I
mean
for
a
lot
of
the
tests.
I
mean
we've
had
we
had
discussions
in
the
sig
before
I'm,
not
sure,
specifically,
who
has
been
kind
of
spearheading
the
discussion,
but
about
trying
to
rein
in
the
growth
of
e2e
and
leave
it
for
the
golden
power.
Testing
of
you
know,
interdependencies
between
components
and
trying
to
focus
controller
testing
on
the
integration
side,
which
is
currently
it
seems
most
tests
or
other
unit
or
e2e.
Any
of
this
huge
gulf,
in-between
I,
don't
think
running
in
dentists.
A
You
know,
you
know
faster,
better
environment
is
the
answer,
at
least
not
the
sole
answer,
because
I
mean,
if
I
think
about
like
the
testing
pyramid,
it
doesn't
consist
of
a
bunch
of
unit
tests
and
a
whole
bunch
of
e2e.
There's
a
chewy
like
integration
layer
in
there.
That
actually
does
will
provides
a
lot
of
value.
It's
certainly
my
name,
my
experience
and
other
projects
and
a
limited
experience.
I've
had
writing
integration
tests
for
tube,
not
saying
it's
the
only
answer,
but
I
don't
think
we
should
be
ignoring
it
either.
A
B
A
Just
I
would
note
that
I
mean,
as
Caleb
mentioned
in
the
chat
Miranda
is
actually
proposes
for
incubation,
not
inclusion
in
the
cube
repo,
for
which
there
would
be
a
higher
bar
as
far
as
I
know,
that
request
was
denied.
So
the
idea
that
you
can
put
anything
of
this
nature
into
the
cube
repo
I
think
to
me
is
kind
of
in
question.
It
doesn't
seem
like
people
think
it
has
enough
value,
and
maybe
these
people's
minds,
but
I
would
certainly
try
to
raise
the
visibility
of
this.
You
know
in
the
dad.
B
A
B
B
A
Think
the
argument
could
also
be
made
that
you
know
this
is
basically
another
mechanism
for
spinning
up
the
cluster
and
it
may
rely
on
virtualized
infrastructure,
but
he
doesn't
necessarily
have
to
be
tied
to
the
tests
right.
They
can
retarget
any
cluster,
and
this
is
just
another
cluster,
but
how
deep
void
in
a
very
specialized
way?
Okay,
gentlemen:
I-
have
to
hop.
Unfortunately,
ecology,
yeah
I,
think
the
suggestion
that
we
carry
this
discussion
up
in
kubernetes
steps
in
front
of
a
wider
audience
sounds
like
a
good
idea.