►
From YouTube: Kubernetes SIG API Machinery 20180606
Description
For more information on this public meeting see this page: https://github.com/kubernetes/community/tree/master/sig-api-machinery
A
A
B
In
Cobb
collector
we're
on
resync
mechanism
that
can
cause
talk
contractor
and
on
the
test.
Grades
has
to
itself
all
the
test
period
that
involves
contractor
failure,
has
sent
deadlock
shares
and
the
fixed
force
merge.
The
two
days
ago,
okay,
Jordan
secure
I
copy
the
issue.
Member
thanks.
So
much
I'll
see
how
it
goes.
Okay,
cool.
C
So
we're
seeing
two
kind
of
categories
of
test
flakes
one
is
failures
at
the
beginning
of.
If
you
have
machinery
owned
tests
that
are
just
in
the
setup
phase
like
create
a
replica
set
and
wait
for
it
to
create
a
bunch
of
pods
and
then
we
were
gonna
go,
do
garbage
collection
tests
on
it,
but
the
test
is
actually
failing
in
the
setup
step
and
I
I've
spent
a
fair
amount
of
time
trying
to
figure
out
what's
going
on
here
and
haven't
had
much
success.
C
It
doesn't
there's
nothing
obviously
related
to
garbage
collection
or
API
machinery
at
all,
really
it's
and
similarly
for
the
aggregated
test.
It's
running
a
deployment
and
waiting
for
it
to
complete
and
then
getting
into
the
body
of
the
test
and
we're
seeing
fairly
get
numbers
of
flakes
in
the
setup.
Stefan
I
don't
know
if
that's
something
that
we
should
reach
out
to
see
gaps
on
or
if
it's
a
systemic
like
watches,
are
slow
and
it's
like
a
scalability
thing,
I'm,
not
sure.
C
The
next
two
bullet
points
are
ones
that
are
actually
actually
appear
to
be
things
we
need
to
drive
and
so
I'd
link
to
the
triage
boards
I've
dug
into
both
of
those
a
fair
amount.
We
tried
to
add
I,
know
Jenny
added
some
more
logging
output
to
the
first
and
I
added
more
logging
output
to
the
second
and
we're
still
not
having
a
lot
of
success
in
driving
mister
ground.
But
those
are
ones
who
stay
on.
Ok,.
A
A
D
Yes,
I
see
someone
actually
answered
the
question
yesterday.
I
will
surprise
to
select,
didn't
notify
me,
but
anyway,
the
other
half
of
my
question
is
really
about
ASD
and
I.
Don't
know
if
anybody
knows
the
answers
to
these
questions,
but
it
looks
like
my
questions
are
now
for
its
D,
so
I
post
about
as
deep
dev
like
and
putting
a
slight
channel.
My
question
I
don't
know
I'll
just
try
Sweeting
it
here
see.
D
If
anybody
thinks
that
this
is
a
useful
forum
for
discussing
it
yeah,
it
seems
like
something
that
the
users
are
basically
would
want
to
know
and
I
didn't
find
it.
Maybe
I'm
just
missing
something.
But
this
is
you
know
a
planning
exercise,
alright,
so
I'm
planning
for
a
workload.
Suppose
I
can
estimate
a
stream
of
Rights
there's
elite
rays,
characterized
by
a
revision
number
a
key
and
a
value,
and
presuming
that
all
the
so
my
question
is
I'm
planning
for
database
size.
D
A
D
D
E
Yeah,
it's
actually
principal,
so
it's
the
size
of
the
objects
in
it,
multiplied
by
the
combien
right
effects
of
your
compaction
interval.
So
it
kubernetes
case
you
have
to
it
would
be
do
it
unless
you
have
a
dominant
object
I'm.
So
in
one
use
case
where
we
have
like
yawns
that
are
large
and
those
would
be
written
in
a
pretty
high
rate
as
heartbeats
those
dominated
space,
and
we
can
basically
calculate
it
based
on.
D
E
So
for
that,
if
you
have
like
a
say,
you
have
a
two
kilobyte
note
object
in
your
1000
notes.
Then
that's
roughly
roughly,
you
know
two
megabytes
there
couple
megabytes
there
and
then,
if
you're,
if
you
write
them
once
every
10
seconds
and
you're
compaction
or
bold,
is
five
minutes,
you
just
do
them.
Okay,
well,.
D
Let
me
make
sure
I
understand:
what's
a
multiply
right,
so
kubernetes,
every
five
minutes
in
the
default
configuration
every
five
minutes.
Cabrini
is
going
to
do
a
compaction
to
the
revision
that
was
current
five
minutes
ago.
So,
just
just
before
the
compaction,
you
have
ten
minutes
of
history,
that's
correct!
So
if
I'm-
yes,
let's
just
take,
say
a
nominal
cluster
size-
say
5,000
nodes
there
heart
beating
once
a
second,
so
that's
500
updates
a
second.
D
D
E
Well,
that's
that
the
the
database
file
does
not
automatically
size
itself
down
so
the
way.
Yes,
so
that's
probably
what
you're
mentioning
yeah
that's
an
effect
right,
if
you,
if
you
have
a
lot
of
activity
in
grow
your
database
size
and
then
it
shrinks
back
down.
What
you'll
have
is
a
mostly
empty
database
with
mostly
three
pages,
but
it
won't
actually
size
it
down
unless
you'd
be
fragile.
Now,.
D
D
D
D
E
D
You
don't
know
for
sure
that
might
so.
Let
me
just
repeat
my
example,
and
it's
just
part
of
the
question,
but
if
I,
if
I
just
Sam,
create
a
cluster,
okay
can
create
a
cluster
creating
nodes
10.
A
second
forget
the
heartbeats.
This
is
just
same
creed.
Just
talk
about
creeps
per
minute:
okay,
I'm,
just
creating
nodes
10
a
second;
they
don't
heartbeat,
okay,
so
just
before
compaction,
there
are
6000
nodes
that
had
created
in
the
last
10
minutes
and
there's
another
800
safe
the
end
of
the
day.
B
E
A
A
D
I'm
I'm
not
sure
I,
don't
have
clear
evidence
on
that,
because
I'm
doing
mixed
work,
less
my
cry
started
by
asking
just
but
pure,
creates
I.
Think
I've
got
an
answer
and
it's
plausible.
So
now,
let's
move
on
to
updates
so
suppose
that
I'm,
let
me
make
an
example
here.
Well
is
try
asking
so
I
think
what
so,
if
there
is
an
update
since
the
last
compaction
saying
say
if
we're
given
object,
this
one
update
since
the
last
connection.
D
E
And
they're
all
they're
all
full
copies,
it's
copy-on-write,
so
there
are
full
full
records.
Okay,
now,
let's
do
delete
how
much
space
does
a
delete?
Take
up
a
delete
is
basically
just
a
tombstone.
How
big
is
that
I?
Don't
know
exactly,
but
it's
a
couple
nights:
okay,
very
small,
plus
a
few
lines.
Ism.
C
E
D
Actually,
one
of
the
follow-up
to
my
question,
it
seems
like
the
question
I
was
trying
to
ask
is
something
that
should
be
documented.
Doesn't
it
seem
reasonable
to
be
in
the
SD
documentation.
D
E
A
E
A
Okay,
let's
see
till
did
you
say
you're
into
Goshen?
Do
you
know
that
would
be
great.
A
G
G
A
Personally,
have
any
objection
to
this
and,
in
fact,
I
think
to
continue
to
scale
with
sig
needs
or
needs
to
delegate
more
to
some
projects
and
like
treat
those
as
first-class
citizens,
so
I.
If
you
think
that
makes
sense,
I
don't
have
I,
don't
have
any
like
procedural
objections
or
other
okay
sorts
of
objections.
G
Are
good
we
we
are
creating
like
a
name
of
a
repo
is:
does
anyone
from
the
sake
feel
that,
like
they
want
to
review
like
the
naming
conventions,
maybe
we're
establishing
stuff
for
these
repos
first
easy
enjoy
the
reclose
themselves?
Yes,
what
what
name
are
you
thinking,
we're
thinking
about
platform
as
a
suffix,
so
starting
with
controller
platform,
I.
G
A
A
G
A
H
I'm
Patrick,
what's
up
to
you,
we're
trying
to
push
through
a
dynamic
auditing
as
little
to
dynamically
configure
your
the
advanced
auditing
features.
This
would
be
primarily
owned
by
sig
off,
but
we
are
looking
for
participation
from
you'll,
hear
I
and
that's
been
one
of
the
asks
from
sig
arch
as
well
as
that
you'll
are
involved
so
I'm
just
trying
to
bring
awareness
to
the
topic
and
see
if
we
could
possibly
get
some
reviewers
on
the
documents
and
approvers
potentially
from
the
sig
as
well
all
right,
you're.
Looking
for
our
reviewers
and.
A
H
B
I
A
That
makes
them
I
I
think
the
project
overall
is
so
there's
there's
two
things
going
on
here.
There
is
just
getting
api's
reviewed
in
general
I
think
the
project
is
super
bottleneck
down
there,
I
think
honestly,
Jordan
has
been
doing
more
far
more
of
those
than
is
fair,
so
yeah,
that's
that's
not
helping
for
the
project.
The
other
thing
that
if
there
is
a
web
book
it
is
this
adding
a
net
new
web
hook.
No.
C
A
C
And
also
I
think
adding
the
possibility
or
like
not
multiplexing
but
I
believe
today
you
can
only
convince
a
single
audit
webhook,
and
so
this
would
be,
presumably
maybe
that's
part
of
kind
of
getting
back
to
the
use
cases,
letting
you
create
one
more
than
one
with
different
verbosity
and
focus
and
filtering
and
and
yeah.
F
A
C
A
B
B
A
H
A
B
A
C
C
B
A
A
C
I
added
some
logging
to
try
to
dump
information
about
what
was
going
on
inside
the
pods.
But
if
someone
who
has
access
to
the
stack
driver
stuff
where
the
container
logs
get
scraped
could
take
a
look
at
it,
that
might
be
more
helpful.
I
know
Walter
had
done
that
a
few
times,
but
I
think
that
might
be
google
internal
still.
Yes,.
A
B
C
Then
the
thing
hitting
the
service
end
point
was
just
getting
a
404
back
and
explicably
I
added
a
fair
amount
of
logging
around
that
and
we're
still
seeing
the
failures
and
the
logging
didn't
turn
up.
Anything
that
seemed
useful
Jenny
saw
some
denial
errors,
but
those
aren't
those
are
consistent
on
good
ones
and
bad
ones.
They're
not.
C
A
I
have
seen
there
is
a
issue
floating
around
with
somebody
complaining
that
if
the
metrics
API
server
isn't
healthy
well,
the
condition
that
I'm
worried
about
is
if
the
metrics
API
server
isn't
healthy.
Then
the
aggregator
can't
aggregate
the
opening
can
that's
somehow
prevents
of
conserving
requests
from
API
servers
that
are
healthy,
at
least
that's.
That
seems
to
be
the
claim
and
that
issue
I'm,
not
sure
if
that's
really
the
case,
but
if
it
is,
we
should
fix
that.
That's
not
good
I
failure.
C
Think
it
was
during
registration
at
one
point,
there
was
a
synchronous
path
that,
like
on
the
first
hit,
could
block
aggregation,
so
existing
things
would
keep
serving
fine.
But
if
you
registered
the
new
API
service
Wow,
the
open
API
aggregator
was
trying
to
do
the
first
hit
on
a
1-1
that
was
failing.
Other
things
would
get
blocked
from
registering
as.