►
From YouTube: Kubernetes SIG API Machinery 20180718
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
B
B
C
There
has
been
some
some
active
comments.
It
is
a
detailed
dock.
It
is
certainly
worth
reading
who
are
interested
in
in
both
using
it
and
perhaps
help
developing
it.
If
there
such
a
person
is
out
there,
there
is
I
think
time
to
finish
the
review
before
we
get
to
the
feature
freeze
date,
so
that
would
mean
we'd
be
able
to
pursue
this
as
an
alpha
feature
in
112
if
met
I.
Think
Metis
is
gonna,
be
the
primary
implementer
right,
yes
yeah.
C
C
B
C
We
let
the
no
conversion
come
in
as
beta,
because
the
implementation
of
it
and
the
semantics
around
it
were
extremely
simple.
The
implementation
of
semantics
around
this
will
not
be
as
simple
so
I
think
that
this
should
come
in
as
alpha
and
then
the
other
thing
that
stands
out.
Our
prereqs,
orko,
recs,
I,
suppose
around
pruning
and
defaulting.
C
E
So
David
yeah,
if
you
look
at
how
people
are
using
CR
DS
right
now,
I've
done
that
on
github.
They
overwhelmingly
don't
use
validation,
there's
about
84
percent
that
when
I
checked
a
while
back
but
after
validation
have
introduced
that
were
not
using
validation,
so
users
have
shown
that
they,
like
the
convenience
of
what
I'm
called
dynamic,
typing
sort
of
the
same
way.
People
like
to
use
Python
and
JavaScript.
So
I
think
we
have
to
keep
that
as
an
option
for
users.
So
we
need
to
find
a
compromise
where
you
can
support
validation
and
not.
F
E
F
Have
a
proposal
in
my
document
I
will
share
for
the
second
item
in
the
agenda.
The
plan
is
to
have
an
an
opt-out
option
like
just
filled
in
the
validation.
Where
you
can
say
is
this.
Subtree
is
not
pooled,
but
this
must
be
explicit.
So
what
we
are
talking
about,
that
we
have
validation,
including
pruning
by
default,
but
you
can
opt
out
for
a
subtree
or
even
for
the
setter
for
the
house
Yeti,
which
is
also
fine,
so.
E
G
I
is
also
important
to
note
that,
if,
if
the
conversion
is
done
by
an
external
system,
that
external
system
has
the
ability
to
deal
with
unrecognized
fields
in
a
source
version
by
dropping
them
when
converting
to
a
target
version,
so
while
I
agree
that
I
think
we
want
to
help
people
not
get
in
trouble,
the
system
responsible
for
conversion
still
has
the
power
to
make
decisions
about
dropping
fields.
The
same
way,
our
native
types
do
so.
B
There
is
only
one
one
common:
there
is
yes,
we
can
do
this
running,
but
it
would
be
a
different
stage
as
what
our
standard
types
are
doing
so
I
have
a
diagram
at
the
end
of
the
document.
That
shows
what
where
things
happens
and
pruning
is
doing
at
the
first
step,
even
before
defaulting
and
validation.
E
E
C
F
F
G
I
I'm
it's
a
little
unclear
what
should
happen
if
you
have
existing
data
and
the
validation
schema
changes.
If
the
next
time
you
read
that
existing
data,
if
the
fields
that
got
removed
from
the
schema
should
be
stripped,
which
means
that
what
you're
reading
doesn't
actually
match.
What's
in
it
CD
we
technically,
the
same
thing
would
happen
with
our
native
types,
but
we
don't.
G
E
F
C
It's
not
that
it's
every
client
that
goes
to
create
something
level,
so
you
know
we
got
away
with
adding
some
pieces
to
pods
init
containers
as
annotations.
Yes,
it
was
a
mistake,
fine,
but
that
mistake
was
only
possible
because
we
did
a
very
careful
upgrade
at
least
an
overshift,
because
we
didn't
allow
any
any
security
context
like
we
disallow
certain
hands
and
security
contexts.
We
had
to
check
all
of
our
data
ahead
of
time
because
of
the
risk
of
time
bomb
data.
D
C
C
C
A
A
F
Me
to
open
anything
so
I
have
a
half-finished
document
for
that
which
I
just
showed
David
for
now,
but
I
will
share
the
next
day's.
It's
about
the
first
sentence,
maze
of
schema
and
value
medication.
It
was
a
short,
out-of-date
Daniel
in
some
comment
where
he
was
surprised
that
we
validate
values
while
doing
the
schema
validation.
So
value
is
something
like
like
expert
on
like
in
the
example
below
we
check
for
one
and
two.
In
this
case,
those
two
things
are
mixed
in
a
nope
API
relation
schema.
So
it's
not
a
real
schema.
It's
more!
F
F
Basically,
we
have
what
open
api
also
has,
but
without
all
those
doom
to
us
or
without
end,
explore
or
not,
and
the
proposal
is
about
and
for
things
that
you
specify
your
schema
in
a
complete
ready
without
those
logical
head
or
extra,
but
not
so
I'm
and
you
can
do
you
can
do
putting
nearly
finished.
We
can
do
pruning
NT
faulting
without
going
into
this
logical
complexity
and
do
the
complete
validation
is
a
normal
validation
step
of
our
history.
F
Yeah,
so
this
does
not
include
to
use
them,
but
the
proposal
is
if
you
want
to
have
a
property
AE
inside
of
such
any
of
you
have
to
specify
it
as
well
outside
of
it
mean
so,
basically
I
just.
Why
did
here?
Yet?
What's
the
properties
here
as
well
as
the
same
so
also
with
ASB
and
the
see
something
like
that.
F
H
F
H
F
B
F
G
Even
just
in
terms
of
using
an
API
like
this
I
think
it
would
be
a
really
difficult
API
to
use
like
this.
The
types
of
api's
that
would
require
the
same
deal
to
default
different
ways
conditional
on
other
fields,
I
think,
would
be
problematic
for
I'm.
Just
thinking
about
other
objects
where
we've
tried
to
do
like
conditional
defaulting
and
how
that
interacted
with
apply
and
updates
and
things
where
you
bring
in
and
an
object
and
we
default
to
field
based
on
end
of
the
field.
G
F
Not
sure
I
couldn't
follow
completely
so
what
you
think
it's
related,
because
it's
complicated
if
you
have
negative
sub
formulas
like
this
and
not
in
sex
or
so
we
also
wants
only
allow
ratcheting
below
and
not
below
XO,
and
not
so
only
below
positive
sub
formulas.
So
it's
related,
but
it's
not
the
same
thing.
F
F
B
I
was
always
discussing
this
with
Stefan
actually
last
week
that
we
are
we're
doing
stuff
on
ser
de
different
people,
different
companies,
different
groups,
even
different
states,
I,
was
thinking.
This
is
a
good
opportunity
to
have
a
work
group
or
whatever
we
call
it.
It
doesn't
matter,
but
I
want
to
get
out
of.
This
is
at
least
meet
the
people
that
works
on
C
or
D
once
a
week
and
discuss
things
that
we
are
doing
and
what
what
I'm
saying
about
what
I'm?
What
I
mean?
B
C
B
One
example
is
cute
builder
that
is
actually
an
interesting
tool
that
helps
us
developing
CDs,
and
we
actually,
even
if
referenced
it
in
the
in
the
document,
I,
guess
or
I,
was
thinking
about.
It
is
in
the
conversion
that
we
think
that
some
tooling
should
be
there
to
make.
For
example,
that
was
easy
for
people
to
use.
C
C
They
are,
you
know
rapidly
experimenting
prototyping
and
trying
to
do
that.
I.
Look
at
this
we're
developing
it
in
master.
We
are
making
fairly
significant
progress
with
the
current
approach
of
of
discussing
here.
Creating
alpha
features,
gating
them
merging
them
at
the
master,
and
we
have
graduated
several
from
alpha
to
beta.
B
So
what
I'm
thinking
is
not
adding
that
much
to
it
actually
just
meeting
with
people
that
working
on
it
and
make
sure
that
everything
is
going
as
smooth
as
it
want
it
to
be,
it's
not
about
deciding
on
anything.
Actually,
if
you
want
to
decide
on
anything,
is
ever
designed
or
anything
like
that
or
just
bring
it
up
to
the
capi
machinery,
it's
just
adding
a
little
project
management
as
it,
and
it.
F
A
B
Yeah
mostly
I
discussed
it
Jordan,
for
example,
directly
or
discuss
with
step
one
directly
or
hopefully
Antony
here.
So
it
would
be
nice
if
we
have
a
channel
of
communication
that
even
if
I'm
not
sure
if
I'm
Tony
has
a
you
know,
V
on
this
would
be
nicely
if
it's
in
that
channel
that
Dante
conceded
and
maybe
has.
H
So
many
who
talk
to
me
about
this
the
other
day
and
the
main
thing
I
was
excited
about-
was
the
idea
of
having
a
meeting
regular
where
the
top
agenda
item
every
time.
Is
here's
exactly
the
list
of
things
we
need
to
get
to
see
our
new
GA
my
new
year.
Here's
the
state
of
all
of
the
things
that
are,
you
know,
then,
are
in
progress
towards
that.
Here's.
The
thing
that
remaining
this
meeting
isn't
I,
don't
think
the
venue
where
we
would
have
to
kind
of
focus.
We.
C
Have
not
had
that
goal
so
we
have.
If,
if
we
have
a
goal
and
and
to
my
knowledge,
it
has
not
explicitly
didn't
stated
that
we
would
like
to
get
CR
DS
to
GA
by
the
end
of
the
new
year
so
and
I'm
assuming
you're,
saying
2019,
because
in
2018
the
answer
is
gonna
be
no
I.
Think
because
you
can't
possibly
make
it
from
alpha
to
beta
to
GA
with
something
that
goes
alpha
in
the
middle
of.
C
G
If
we're
talking
about
distinctions
between
working
groups
and
sub
projects,
I
I
think
CR,
these
could
probably
make
sense
as
a
sub
project
of
API
machinery,
I
mean
it
not.
Everyone
involved
in
API
machinery
is
working
on
that
or
interested
in
that,
and
that
is
an
area
with
work
to
do
and
different
features
and
priorities.
G
D
C
B
B
I
just
want
to
be
to
add
some
some
sort
of
in
an
organization
here,
because
there
are
things
that
I
may
know
or
may
not
know
about
that,
going
on
and
series
and
we'll
make
sure
that
everybody
knows
that
they're
working
on
validation
go
work
and
from
ingerborg
and
Duvall
thing
they're
working
on
and
who
is
working
on
on
those
I.
Just
just
want
to
make
sure
that
yeah.
G
As
far
as
a
mailing
list,
like
I
kind
of
agree
with
David
I,
there's
actually
very
little
traffic
on
the
API
machinery
mailing
list
as
it
is,
and
so
starting
CRT
related
threads
on
that
list
seems
like
it
would
be
a
good
starting
point
and
that
way
people
who
are
interested
in
them
can
follow
those
and
people
who
are
not
can
mute
them
and
if
it
becomes
overbearing
how
much
traffic
is
on
the
API
machinery
mailing
list
related
to
C
or
D.
Is
that
most
people
don't
care
about?
Then?
B
Okay,
so
I
think
I
think
that's
that's
fair
to
see
if
we,
if
we
start
using
the
API
missionary
channels,
and
if
we
made
a
lot
of
noise,
then
we
can
start
our
new
channel.
I
was
thinking.
You
know
the
at
least
for
me.
There
many
times
that
I
had
something
and
I
didn't
post
it
to
the
slack
Channel
may.
A
I
Hi
everyone,
as
promised
I'm,
going
to
present
the
average
Russian
proposal
today,
but
today
I'm
going
to
because
there
are
a
lot
of
discussions
in
the
cap
already
so
today,
I'm
going
to
focus
on
the
problems
trying
to
solve
and
the
different
problems
are
garbage
question
in
general,
that
I
observed
and
also
what's
the
next
step
for
this
crap,
so
some
background
for
garbage
collection.
I
I
The
first
one
is
on
a
reference
place,
so
any
resource
that
has
an
owner
its
metadata.
They
can
be
cleaned
up
with
by
generic
garbage
collector
and
the
way
what
I
mean
by
clean
up
is
like.
If
your
owner
is
deleted,
then
the
garbage
collector
action
in
Forza
cascading
deletion.
By
looking
at
the
dependence
on
the
references,
then
we
have
capacity
based
garbage
collection.
I
So
once
the
number
of
terminated
parts
keep
the
limit,
the
GC
controller
party,
C
controller
will
kill
the
pods,
but
users
have
no
control
of
which
pause
are
filters
or
capacity,
and
then
another
capacity
based
garbage
collector
garbage
collection
is
the
controller
history
of
resources.
That's
created
and
managed
by
weblog
controllers,
for
example,
deployment
will
use
REM
cassette
as
its
controller
history
objects,
and
then
you
can
say
how
many
history
I
want
to
keep
in
this
deployment.
I
Then
deployment
controller
will
only
keep
the
number
of
replicas
as
and
if
the
capacity
exceeds,
they
were
delete
them.
So
it's
controlled
by
the
users
can
set
that
limit,
but
only
the
workload
controller.
The
things
that
are
controlled
by
regular
controller
can
benefit
from
this,
and
the
last
one
is
TTL
based.
We
have
so
event
is
created
with
a
TTL
set
and
it's
implemented
with
@
çd,
TTL
and
so
I
think
by
default.
I
I
So
the
the
main
problem
that
I
see
from
these
three
types
of
garbage
questions
is
that
for
capacity
based
and
the
TTL
based
they're,
not
generic
and
you
can
see
from
their
name
because
they're
not
called
something
like
generic
garbage
collection
and
also
they're,
not
configurable
by
users,
because
you
can
only
set
it
with
a
component
thought
flag
only
cluster
and
we
can't
control
that.
And
if
you
want
to
change
the
number
you
need
to
restart
those
component
and
for
a
capacity
based
of
what
I
mean.
Is
that
especially
for
the
pod
GC
controller?
I
There's,
no
fairness,
because
you
can
have
one
net
base
that
creates
lots
of
pods
and
the
other
namespace
has
only
a
few
pods,
but
it
may
be
affected
by
the
other
and
larger
the
dump
space.
With
a
lot
of
pots,
and
also
it's
not
related
to
resource
cora
and
also
you
cannot
undelete
the
for
pot
GC,
you
cannot
delete
the
pods
in
the
order
of,
for
example,
their
termination
time.
I
I
Another
problem
is
that
we've
seen
a
pattern
that
I
call
immutable,
so
I'm
not
talking
about
that.
The
resource
cannot
be
mutated,
I'm,
just
talking
about
the
resource
that
users
use
usually
keep
creating
them,
and
they're
short-lived
and
users
never
update
those
resource
if
they
want
to
change
something
about
the
resource.
They'll
just
create
a
new
one.
I
The
example
of
this
is
our
jobs
or
config,
Maps
or
secret,
so
for
these
type
of
resources,
users
need
to
keep
track
of
whether
they
still
need
this
resource
or
not,
and
they
need
to
figure
out
how
to
clean
them
up,
and
if
the
users
don't
do
this,
and
it's
very
easy
to
overload
the
cluster
and
slow
down
the
master.
We've
seen,
we've
seen
something
like
a
small
cluster
with
a
small
master
PM,
but
with
like
100
K
of
code
jobs
that
users
probably
don't
even
need,
but
because
they
never
clean
up
those
jobs.
I
The
FC
and
API
server
starts
coming
out.
So
you
may
think
that
how
about
using
on
the
reference,
but
because
those
jobs,
if
they
are
not
created
by
core
chrome
jobs,
then
they
don't
have
any
owners.
If
you
don't
have
any
owners,
then,
on
the
reference
won't
help
you
with
that.
I
The
last
problem
is
that
we
cannot
centralize
the
cleanup
of
multiple
resources
with
on
the
references.
What
I
mean
is
that
let's
say
you
have
a
bunch
of
manifests
and
you
want
to
deploy
them
into
kubernetes
and
you
want
to
specify
the
owner
reference
relationship
in
the
manifest
before
you
deploy
them
so
that
you
can
easily
keep
clean
them
up
without
doing
anything.
You
can
just
delete
the
root
resource
and
have
everything
else
be
cleaned
up
with
it.
I
But
you
cannot
do
that
today,
because
owner
UID
is
required
in
all
no
references,
and
you
can
only
know
the
UID
until
you
create
a
resource.
So
then
users
need
to
first
create
a
bunch
of
things
and
then
figure
out
the
u
IDs
and
then
ensure
the
owner
references
into
each
dependent
resources,
and
even
if
you
make
UID
optional,
then
you
will
hit
a
problem.
When
you
create
a
dependent
before
the
the
owner,
then
your
dependant
might
be
killed
prematurely.
I
G
G
It
seems
like
creating
several
resources
declaratively
and
then
deleting
a
single
root
resource
isn't
symmetric.
That's
not
how
I
see
most
people
doing
declarative
management.
They
passive,
complete,
manifest
when
several
objects
to
create,
and
then
they
pass
the
same
complete
manifest
with
those
objects
to
delete.
I
You
mean
I
just
delete
all
my
many
funds
with
something
like
Q
control.
Do
it?
Yes,
the
whole
folder
well.
G
I
Can
do
that,
but
some
user
may
not
want
to
do
it
off
the
things
that
are
in
the
folder.
You
may
have
not
one
tree,
but
several
trees
I
mean
several
routes.
G
Yeah
I,
if
we're
talking
about
like
declarative
management,
it
seems
a
little
strange
to
have
declarative
management
with
overlapping
things
like
I
have
a
manifest
and
you
have
a
manifest
and
there's
a
shared
resource
between
our
manifests.
It's
not
generally
gonna
work.
Well,
if
we're
both
sort
of
assuming
creation
deletion,
lifecycle,
ownership
of
the
same
things,
but
anyway,
I
just
wanted
to
point
out
that
you
can
pass
a
complete
manifest
to
delete.
So
that's
all.
I
So
an
overview
of
this
proposal
is
that
I
want
to
introduce
CTL
mechanism,
but
the
TTL
is
not
the
time
it's
not
set
when
you
create
something
it
is
set.
When
you
want
to
clean
up
things
that
are
already
either
finished
or
not
used
by
anyone.
So,
for
example,
actress
the
job
finishes,
you
set
a
CTO,
for
example,
five
minutes.
I
It's
because
sometimes
you
need
to
observe
the
results
from
the
job
somewhere
else,
so
users
can
set
that
time
for
them
to
observe.
The
result,
then
clean
up
in
another
use
case
is
for
config
maps
and
secrets.
When
you
create
a
config
lab,
the
users
can
set
the
TTL
to
something
like
30
minutes,
and
then
it
can
be
cleaned
up
after
it's
not
used
by
other
resource.
For
example,
it's
not
a
reference
in
deployment,
template
or
pods
or
staple
science,
etc.
And
then
the
user
can
just
deploy
the
applications.
I
So
I'm
not
going
to
talk
about
the
details
of
the
proposal
today,
but
based
on
the
feedback
that
I
get
in
general
to
implement
the
the
conflict
lab
secrets.
Type
of
cleanup,
which
I
can
I
will
call
it
the
usage
reference
cleanup.
I
I
I
So
the
next
step
is
that
I
want
to
introduce
a
TTL
controller
that
can
clean
up
the
finish
jobs
and
completed
parts,
and
but
I
also
want
to
tackle
other
things
like
custom
resources.
That
can
also
finish
and
can
be
cleaned
up,
and
the
use
case
of
the
of
the
TTL
controllers
is,
for
example,
the
cube
flow
jobs.
I
They
keep
creating
jobs
and
they
want
a
better
way
to
clean
up
those
jobs
and
for
a
completed
pods
things
like
spark
controller
I
mean
spark
operator
and
like
Jenkins,
that
using
kubernetes
plugins
they
will
create
pods
that
will
complete,
and
sometimes
they
don't
know
how
to
clean
up
those
pods
and
if
they
can
set
a
GTO
to
those
things,
you'll
be
useful
for
them
and
I'm
also
thinking
about
making
it
general
so
that
we
can
put
a
GTO
in
the
metadata
that
all
resources
can
set
it
instead
of
putting
it
in
every
resource
bag.
I
So
if
you
have,
if
you
are
interested
in
this
problem
or
you
have
feedback
on
the
usage
reference
and
you
can
come
to
the
can
go
to
the
cap
to
take
a
look-
and
the
cap
is
right
here-
number
two
to
eight
seven
in
community
repo
and
I
also
opened
another
design
doc
for
the
clean
up.
The
finished
objects
here
and
you
can
link
it
from
the
cap.
I
put
it
in
the
very
bottom
of
the
page.
So
that's
all
thanks.
C
Does
a
different
thing
right,
so
this
controller
isn't
actually
responsible
for
trying
to
chase
links
when
a
thing
has
already
been
deleted,
which
is
what
GC
does.
This
would
be
a
piece
of
a
controller
that
is
responsible
for
deciding
to
issue
that
delete
and
then
later
GC
would
be
used
to
apply
the
delete
properly.
I
meant.
K
J
K
So
we
are
able
to
create
an
object
using
qcg
over
try
sending
these
new
endpoints.
It's
actually
not
a
new
NIS,
neutrons
I'm.
Sorry
from
the
fashion
trend,
so
you
can
create
out
of
Pioneer,
add
objects
on
these
endpoints.
They
will
ask
the
bride
is
going
to
be
in
the
object.
It's
going
to
be
mulched
and
obviate
and
stalling
I
see.
So
this
is
what
we
have
today:
well,
not
completely
sure
that
we're
going
to
keep
the
existing
Mouse
logic,
which
has
been
written
by
a
field
a
while
ago.
K
So
we
don't
know
it
just
emotional
energy.
We
want
to
keep
in
the
end.
We
have
tremendo
different
scenarios
now,
like
first
I'm
good
at
saving
intent.
We
want
to
save
the
intent
of
mutating
webhooks
and
we
don't
have
a
thoughts
option
for
controllers
by
example.
So
if
you
have
a
conflict,
controllers
are
going
to
fade
on
that.