►
From YouTube: 2017-03-16 Kubernetes SIG Scaling - Weekly Meeting
Description
2017-03-16 Kubernetes SIG Scaling - Weekly Meeting
B
Here
that
you
see
it,
so
it's
basically
back
to
our
consideration,
work,
/,
etc.
Joe
I,
don't
really
know
but
actually
slows
the
proposal
down
just
because
it
had
a
constant
with
Daniel.
Anybody
explain
to
me
the
direction
in
which
configurations
of
clouds
are
going
to
future,
how
they're
going
to
removal
by
the
slice
these
sent
to
all
components
and
they're
going
to
be
putting
to
a
file
go
when
it's
a
fake
and
that's
going
to
be
disabled
vehicle
new
mechanism.
B
How
to
update
tissot,
isn't
applied
dynamic
li
et
cetera
and
those
are
going
to
be
plenty
to
you,
config
maps,
so
that
you
can
pull
them
out
as
they
basically
a
strength
because
invert
invitation
on
and
then
you
can
monetize
their
yeah
so
anyway,
because
of
the
direction
of
your
going
and
some
momentum
in
that
direction
and
I
just
it
was
just
going
to
give
it
away
already
bragging
about
it.
Sorry,
I
drugged
you
organs,
it
no.
C
C
For
their
fat
going
on
and
then
they're
still
going
to
be
some
landline
flag
and
depending
on
how
this
stuff
works,
if
it
pulls
it
off
of
discourse,
is
pulling
it
off.
Thank
you
can
I
serve
directly
I
mean
there's.
You
know,
there's
still
some
stuff
we're
going
to
want
to
collect
my
pendant,
but
yeah
I
think
it
plan.
Need
you
to
take
a
step
back
to
make
sure
that
it's
hot,
it's
together
with
everything
else,
is
going
on.
Yeah.
B
So
I
did
it
up
I
closed
the
ticket
and
they're
moving
in
a
different
direction
by
analyzing.
What
we
have
right
now,
where
the
directions
going
cetera,
and
so
one
thing
that
the
group
we
realized
is
that
out
of
all
the
components,
dps
server
seems
to
be
somewhat
of
a
snowflake,
because
it
doesn't
have
a
slash,
couldn't
fix
d
influence.
All
other
ones
do
and
harshly
part
of
the
reason
why
we
suspect
applications
because
up
sensitive
information
that
might
be
exposed
there
right.
B
So
anyway,
we
felt
that
there
was
enough
out
there
with
all
components.
We
felt
that
we
can
control
the
DA
exposing
of
sensitive
data
enough,
that
we
can
go
ahead
and
just
create
something
see.
Endpoint
put
a
lot
of
those
components,
figurations
into
component
of
things,
and
then
exposure
to
fix
you
in
plenty
to
get
server
is
so
I
went
ahead
and
did
that
endpoint
and
I
think
it
works
a
little
while
I
mean
it's
doing
is
nothing
different
from
what
other
components
are
doing.
We're
in
it
worries
is
consistent
and
we
still
kill.
B
Billy
will
be
able
to
get
some
information
from
a
bi
server
that
we
need
for
the
work.
We
need
the
future
going
forward
right,
there's
a
good
side
of
wasn't
you
I
expected
that
gets
direction
so
anyway,
I
just
want
some
advice,
maybe
see
how
I
can
get
another
one
on
board
on
this
and
grew
bored
with
it.
It
really
is
a
point
of
rotational
component.
It's
nothing
new
I,
don't
think
it's
as
controversial
as
what
we
were
rolling
earlier
and
I.
B
C
C
D
Difficult,
what's
interesting
to
me,
is
the
sort
of
life
cycle
between
what
is
considered
component
config
and
what
people
are
trying
to
put
the
config
map,
because
component
config
to
me
is
not
a
sold
story.
Right
like
it
is
at
least
this
current
incantation
of
how
it
right
it
requires
a
lot
of
care
and
feeding
in
love.
In
order
for
you
to
do
is
both
a
like.
D
Adding
a
single
command
line
flag
requires
you
to
regenerate
the
entire
API
no
I'm
like,
and
that
is
a
non-trivial
level
of
work
for
something
which
is
a
trivial
thing
right.
So
if
a
person
wants
to
add
a
simple
flag,
they
need
to
reach
in
everything
and
that
that
is
quite
the
it's
quite
the
super
high
bar
or
adding.
What
are
the
equivalent
of
you
know
a
knob
in
a
file
right,
so
there's
going
to
have
to
be
this
whole
round
trip
process.
I
think
it.
D
It
seems
to
be
critical
massing
in
this
release
like
because
Brian
sent
out
that
email,
I
see
thing
I,
see
many
points
converging.
You
know
like
there's
a
proposal
on
the
scheduling
sig.
To
also
do
this
right
to
basically
have
a
single
configuration
file
to
rule
them
all
for
the
scheduler
and
it
will
be
a
config
map
whether
or
not
they
will
push
all
the
details
through
two
component.
Config
is
still
it's
still
a
TVD
item
right,
okay,
so
it's
this
release
cycle.
There's
going
to
be
a
lot
of
churn
and
I.
D
B
D
D
B
D
B
A
A
This
topic
for
a
really
long
time
and
most
it
started
and
I
think
we
still
have
this
issue,
which
is
how
do
we
compare
the
performance
of
different
different
configurations?
We've
spent
a
lot
of
time,
trashing
around
on
config
tweaking
for
things
like
QPS
limits,
and
the
issue
that
we've
had
for
quite
some
time
here
is
that
the
the
overall
system,
performance
or
truth,
but
can
be
massively
affected
by
some
of
these
configuration
things.
D
Yeah
so
I
think
the
idea
is
like
you
want
to
have
one
tool:
you
plunk
it
down.
It
grabs
enough
information
not
just
to
get
you
performance
data,
but
to
also
be
able
to
say,
like
here's,
the
configuration
plus
the
performance
data.
So
that
way
you
don't
just
have
results.
You
have
the
the
full
config,
along
with
the
results.
Okay,.
E
F
D
B
A
D
Better
I
mean
there's
been
results,
pushed
back
I
think
we
found
a
bunch
of
tiny
artifacts
in
the
code
and
we
fixed
a
number
of
them.
I
think
the
the
second
step
to
that
folks
want
to
do
is
to
we
talked
on
sig
testing
is
to
segregate
the
tests
so
that
the
ownership
is
clear
right
because
right
now,
it's
just
a
model
right.
The
first
step
is
segregation.
The
second
step
is
to
start
to.
D
So,
at
least
on
the
six
scheduling
side,
we
did
all
three
okay,
so
we
we
segregate
'add,
we
found
the
timing
artifacts
and
we
refactored
it
to
get
rid
of
them.
Now.
I
would
like
to
take
this
forward.
You
know
it's
sort
of
spread
the
gospel,
so
to
speak,
that
this
is
the
easiest
way
for
everyone
to
benefit.
So
that
way
you
get
instant
assignment
and
you
get
less
flakes.
D
You
know.
If
you
follow
these
set
of
patterns,
your
be
on
a
better
path:
I
don't
want
to
own
it,
but
I
think
what
I'll
do
is
start
to
do
some
roundabout
gosling,
while
I,
try
to
get
pieces
of
the
tests.
The
testing
infrastructure
cleaned
up
to
support
that
notion
of
what
we
eventually
want
to
do
is
to
be
able
to
have
plunk
down.
D
You
know
a
single
end-to-end
binary
at
one
point
and
then
be
able
to
you
know,
gather
all
the
data,
because,
right
now
we
have
this
parceling
of
tools
right
everything
else
and
I
think
I'm
going
to
work
along
those
lines
to
get
this
moving
forward.
So
I
went
I
gave
you
the
full
round-trip
circle,
but
I
talked
about
why
some
tests
are
flaking
and
how
some
people
have
gone
about
fixing
them
is.
D
I'm
writing
up
a
proposal
to
basically
say
like
hey:
let's
follow
this
pattern,
any
other
cigs,
that's
underneath
my
purview,
you
know
like
I,
will
talk
with
Merrick
and
why
tech
and
push
the
performance
ones
into
their
own
performance
location
then
do
a
similar
pattern
there
after
they
left
after
that
right
at
the
proposal
and
submit
it
to
community
stove
perfect.
So.
B
Maybe
this
race,
what
you're
talking
about
and
if
it
isn't
sorry
I'm
eating
it
I'm
noticing
a
lot
of
place.
Our
timing
issues
like
something
is
being
created
or
be
brought
up,
and
then
you
know
num
sideration
series
you
later
psychic
things
I
can
take
time
and
then
you
expected
to
be
there.
But
your
your
feedback
is
not
there
and
you
know
we
killin.
Error
n
is
not
available.
Yes,
mentioning
like
yeah.
D
So
the
big
changes
that
we're
making
or
that
made
inside
of
the
scheduling
tests
is
instead
of
us
having
an
explicit
wait.
You
know
which
has
done
all
over
the
testing
infrastructure
is
to
basically
have
an
informer,
the
former
than
is
looking
for
specific
events,
and
it
will
timeout
after
a
law
much
longer
time
off
integral.
Like
two
minutes,
you
know
basically
means
that
it
never
saw
the
event
at
all
right,
mom
and
instead
it
is
basically
looking
for
this
wit,
very
narrow
window
of
exact
behavior
right.