►
From YouTube: Kubernetes SIG Scheduling meeting - 2018-10-11
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
So
all
right,
let's
start
I,
have
a
couple
of
updates,
but
not
that
many.
But
today,
then
we
can
open
up
the
question
of
comments
and
you
guys
might
have
so.
One
of
these
is
regarding
gang
scheduling.
You
know
this
has
been
an
effort
that
has
been
going
on
for
a
while.
The
most
recent
version
of
the
gang
scheduling
for
cause
Ollie's
close
to
final.
A
This
is
a
feature
required
for
some
of
the
HPC
performance
computing,
as
well
as
some
machine
learning
workloads
also
generally
in
batch
batch
processing.
This
is
a
feature
that
is
sometimes
needed
so
anyway,
if
you
are
interested
in
the
feature,
please
take
a
look
at
the
proposal.
Proposal
e
is
linked
to
our
spreadsheet
and
the
spreadsheet
is
linked
or
meeting
proposal
which
also
our
meeting
notes,
which
is,
and
they
in
the
calendar
invite
that's
one
update.
A
B
A
First,
PR
of
this
disk
during
framework,
we
are
going
to
introduce
the
interfaces
for
a
couple
of
extension
points,
particularly
for
pre
bind
and
reserved
extension
points.
If
you
are
familiar
with
the
framework
is
that
proposal.
These
are
two
extension
points
used
after
a
node
is
chosen
basically,
and
once
we
try
to
update
the
scheduler
cache
as
well
as
before
binding
a
pod
to
a
node.
These
two
extensions
are
called
so
we
are
for
the
very
first
draft.
We
are
gonna
introduce
these
extension
points.
A
A
C
Know
one
question
so
rebind
and
mine
stated
more
more
of
a
fYI
thing
to
the
extension
points
that
this
is
the
node
being
chosen.
So
any
particular
reason
that
we're
starting
with
the
like
the
first
implementation,
is
starting
with
these
extension
points
like
in
a
some
sort
of
a
reverse
chronological.
So.
A
Reserve
is
more
like
FYI.
This
is
basically
for
mostly
used
for
by
the
plugins
or
that
want
to
update
some
of,
for
example,
a
cache
for
a
state
that
they're
keeping
pre
mind
is
an
approval
plug-in
basically
so,
prima
all
the
preborn
plugins
must
return
true,
for
the
part
be
able
to
be
bound
or
not.
If
any
of
them
return
false,
it's
not
gonna
be
about,
and
the
part
is
gonna
basically
get
to
reject
it
and
go
back
to
the
scheduling
queue
for
another
attempt
which.
A
D
A
Level
of
checks
and
the
reason
that
we
chose
these
two
particular
plugin
points
to
implement
as
the
first
ones
is
that
there
are
a
couple
of
features
that
would
require
these
one,
which
is
most
important,
is
dynamic
volume,
binding,
so
dynamic
volume
binding
is
today
implemented
as
a
as
sort
of
like
a
hard-coded
part
of
the
scheduler,
and
in
fact
it
was
one
of
the
reasons
that
motivated
us
to
to
think
about
a
scheduling
framework.
We
are
not
very
happy
with
the
fact
that
it's
a
lot
of
logic
is
integrated
inside
the
scheduler
core.
A
So
that's
one
of
the
reasons
we
would
like
to
take
it
out.
Jonnie
scheduling
is
also
gonna
benefit
from
a
couple
of
these
extension
points.
How
we
both
reserved,
by
probably
it
probably
needs
one
more
before
we
can
actually
implement
thanks
Caitlin
using
the
current
scheduler
or
scheduling
framework,
but
anyway,
these
are
also
needed.
Yes,
thank
you.
B
You're
probably
wondering
on
question
on
the
framework
so
I
think
right
now,
the
most
I
recommend
in
the
way
of
the
use
of
its
minor
efforts
to
eat,
is
to
use
the
extender
right
to
rise
on
several
extender
and
implement
their
specific
knowledge.
So
basically
in
the
timeline
perspective.
So
when
do
you
think
it's
mature
for
a
user
to
use
the
framework
as
the
most
recommending
way
to
customize
schedule,
yeah
sure.
A
B
A
B
A
We
can't
possibly
think
about,
and
maybe
moving
some
of
the
extension
points
as
beta.
So
114
is
probably
gonna
be
the
earliest,
but
you
know
again
depends
on
the
Preferences
of
our
users.
Some
of
the
companies
that
we
work
with,
for
example,
never
use
any
of
the
alpha
or
beta
features.
They
want
only
GE
features
so,
depending
on
your
criteria
and
restrictions
that
you
have
remained,
you
may
want
to
wait
until
it's
GA
and
that
could
possibly
be
a
year
from
now.
A
The
reason
that
I
did
this
is
a
lot
of
our
peers
are
assigned
to
these
folks
and
need
to
be
reassigned,
and
sometimes
the
authors
of
the
these
PRS
are
not
fully
aware
of
the
fact
that
some
of
these
folks
are
not
very
active
and,
as
a
result,
their
PRS
remain
unattended.
Basically,
nobody,
nobody
looks
at
them,
so
we
wanted.
We
wanted
to
avoid
this
situation,
so
I
actually
sent
a
PR
to
change
our
reviewers
and
I
hope.
A
Folks
are
okay
with
that.
I
know
that
particularly
I
actually
wanted
to
bring
this
to
our
wishes.
Attention
Ravi's
here,
which
is
not
here
but
August.
If,
if
you
see
a
badge,
please
tell
him
that
actually
moved
him
from
the
from
the
reviewers
list
of
schedule.
If
he
thinks
that
he
has
the
time
to
help
us
more
with
the
reviews,
I
will
be
happy
to
add
him
back.
A
A
F
Actually
I
did
not
get
time
to
work
on
them
and
there
is
one
more
feature
that
way
and
I
have
started
them
paying
based
evictions
I
have
increased
the
coverage.
Luckily,
I
did
not
create
a
PR
for
that
when
I
was
testing
it
locally.
The
other
thing
that
I've
noticed
is
it's
not
as
fast
as
normal
evictions
are
meaning
the
way
we
are
calling
the
deletion
from
the
node
controller
code.
We
were
previously
directly
making
calls
to
delete
API
for
the
pot
on
the
no
it's
not
as
fast
as
that.
F
F
F
A
E
D
Curavi
yeah
I
also
have
a
question
related
to
what
Bobby
just
asked
is
why
I
was
expecting
like
a
race
condition
or
like
a
an
issue.
Was
the
attained
based
controller
with
putting
notes
on
there
and
the
couplet
would
studying
evictions
and
the
documentation
says
that
there's
a
throttling
implied
here
so
in
order
to
avoid
kind
of
a
massive
epic
eviction
of
pots
on
a
note,
but
I
would
also
strive
for
the
simplest
solution
with
the
couplet,
which
seems
to
be
a
little
bit
more
easier
to
ferment
from
an
operative
perspective,
configuring
those
things.
D
F
A
A
So
if,
for
example,
in
a
cluster,
you
suddenly
decide
to
take
down
a
whole
zone
with
I,
don't
know
thousand
nodes,
then
you
will
have
like
a
thousand
cubelets
on
thousand
different
machines
that
can
work
in
parallel
to
evict
all
the
parts,
as
opposed
to
like
a
single
controller,
and
the
master
know
that
is
that
has
to
take
down
tens
of
thousands
of
parts.
That's.
F
The
job
of
note
controller,
to
tell
I,
mean
the
eviction
could
happen
from
the
cubelet
side,
but
the
node
controller
is
the
one
that
is
responsible
for
telling
okay
I'm
going
to
play
a
taint
or
I'm
going
to
set
the
instead
available,
but
I'm,
not
100%
sure
if
cubelet
is
going
to
do
the
elections
based
on
the
conditions
that
are
set
by
external
component.
So
if
cubelet
sets
those
fields
based
on
that,
if
Q
ability
is
everything
I
think
that's
the
way
it
is
currently
designed.
F
F
Reacts
to
the
the
funk,
the
main
job
of
Q,
that
is,
to
identify
the
problems
on
the
node
and
based
on
that
it
is
going
to
take
an
action.
It's
not
that
the
external
influences
or
the
triggering
point
should
be
from
from
an
external
entity.
Rather
Cubitt
itself
should
do
it.
Is
this
what
I
think
okay.
A
B
Ravi
take
a
look
at
the
DPR
and
there
are
clouds
raised
the
thing
1.12
he'll
be
giving
in
last
minute
of
the
one
control
which
is
fixed
performance
of
tenth
note
condition.
So
in
the
beginning,
Kraus
engine
ensures
a
solution
based
out
of
code
of
tenth
manager,
I
guess,
which
was
used
as
a
slice
of
China
ethics,
kind
of
seems
didn't
perform
well,
so
finally,
he
changed
really
limited
queue
that
seems
to
work.
So
if
you
want
I
can
point
you
out
back
here,
yeah.
F
That'll
be
great
that
pier
I
think
it
was.
There
was
a
scalable
shoe
which
otech
has
created
for
the
five
thousand
nodes
limit
and
cross
has
created
that
pier
to
solve
that
problem,
but
I
am
Not
sure
the
scalability
test,
if
it
has
started
working
fine
because
of
the
change
the
because
after
the
change
has
gone
in,
there
were
only
a
couple
of
times
where
the
test
was
failing
and
then
after
something
the
test
has
started.
Passing
so
I
do
not
know
if
the
change
that
cross
has
create
has
introduced
has
solve
the
problem.
A
Is
that
well,
it's
actually
not
related
to
eviction,
but
there
was
this
issue
with
112
that
if
P,
if
a
cluster
runs,
110
or
or
older
parts
110
is
actually
supported,
we
support
cubelets,
which
are
two
versions
behind
the
control
plane.
So
if
you
run
a
cluster
112,
you
can
run
one
time.
There
was
an
issue
with
those
classes
sent
a
PR.
It's
not
measure
as
far
as
I
can
tell
yet.
A
At
least
it
was
not
measure
until
last
night.
So
the
issue
was
that
110
didn't
one
thing:
he
didn't:
have
the
capacity
or
ability
to
check
node
affinity.
So
this
was
conflicting
with
the
new
demons
of
the
scheduling
by
the
default
scheduler
and,
as
a
result,
some
of
these
demons,
scheduled
by
the
invite
the
default
scheduler
we're
getting
rejected
by
those
older
cubelets
in
the
cluster
class,
has
a
PR
out
that
is
going
to
fix
this
for
112,
newer
versions
of
the
cubelet
like
111
and
on
will
have
that
change.
A
F
Yeah
suicide
I,
initially
looked
at
the
issue.
I
had
a
discussion
with
the
the
person
who
has
filed
that
issue.
I
was
initially
thinking
that
we
could
backward
the
the
change
that
Kraus
has
made
for
1.11
I
believe
but
wondered
well.
We
could
back
4210
but
Jordan.
He
told
that
we
will
not
back
for
the
EPA
changes
to
110.
F
A
Yeah
I
guess
we
do
need
to
really
I,
don't
know.
We
don't
really
need
to
carry
all
these
changes
forward,
because
110
is
the
last
version.
Basically,
that
is
supported.
Do
it
with
112
and
from
the
next
release.
113,
we
are
not
going
to
support
cublas
older
than
111.
111
will
be
fine,
so
it's
it
makes
sense.
To
just
add
this
change
to
the
two
one
thing
cubelets
for
now:
I
guess
for
for
112
reasons.
Yes,.
A
True,
okay,
so
one
Ravi,
you
have
one
more
item
on
your
plate,
which
is
deprecated
critical,
pod,
annotation
I,
don't
know
if
you
were
gonna
have
time
for
doing
this,
but
given
the
number
of
contributors
that
we
have
on
a
lot
of
people
who
are
asking
for
more
work,
if
you
don't
have
the
time,
please
feel
free
to
tell
me
so
that
we
can
reassigned
to
someone
else.
Yeah.
F
A
F
One
question
related
to
this
P
alpha,
pod
affinity:
I
think
there
is
an
incubator
project
called
boot
queue
that
redhead
I
started
using
extensively.
We
found
a
bug
in
in
the
latest
version
once
that
got
merged,
so
I
pointed
to
a
like
offline,
but
I'm
curious.
If
is
the
thing
that
we
are
interested
in
getting
going
forward,
the
patch
that.
C
One
sort
of
a
little
naive
question,
so
please
feel
free
to
ask
me
to
look
at
into
awkward.
I
was
I
just
had
this
question
from
the
new
scheduling
framework,
so
to
quote
from
there
we
mentioned
that
that
keeping
scheduler
Vivaan
backward
compatibility
is
an
on
goal
specifically
with
the
v1.
Extenders
won't
work
in
the
new
framework,
so
this
statement
did
we
mean
that
the
but
the
new
framework
also
says
that
it
will
still
have
the
out
of
process
plug
and
supported.
So
does
that
mean
the
out
of
process?
C
A
Initial
plan
actually
later
on
I
actually
later
on.
Maybe
this
this
is
not
updated,
correct.
You
know
that
was
our
initial
plan.
They
see
our
initial
plan
was
to
build
a
framework
from
scratch.
Sort
of
that
was
that
got
changed
later
on.
We
felt
like
it
would
be
too
much
to
build,
as
well
as
to
roll
out
and
particularly
rolling
out
and
use.
Scheduler,
which
was
not
backward-compatible,
would
take
almost
infinite
in
fronts
of
time
in
clusters
that
drawn
production
workers.
A
C
Okay,
thanks
and
just
one
more
thing
related
to
that
I
mean
the
in
process
will
definitely
be
the
most
driving
forces
like
I,
guess,
performance
to
prevent
the
marshalling
and
and
marshalling,
but
other
than
that,
so
in
process
would
definitely
mean
that
we
would
need
to
recompile
or
like
I
from
the
documented
did
not
seem
like
a
one
statement.
I
can't
seem
to
look
at
exactly
where,
but
one
statement
referred
to
that
we
might
not
need
to
recompile
I
was
kind
of
confused
at
how
would
we
have
in
process
plugins,
which
would
not
recompile
ation?
C
A
Actually,
given
that,
given
that
we
don't
have
a
clean
concept
similar
to
like
dll's
in
go
yeah,
we
compilation
is
gonna,
be
actually
yeah
all
right.
You
need
to
copy
files
into
our
country
scheduler
and
maybe
just
register
your
plugins
in
our
registration
process,
but
otherwise
other
than
this
and
the
final
you
need
to
recompile.
Hopefully
you
don't
need
to
make
any
other
changes
or
you
don't
need
to
do
any
complicated
marriages
of
the
code
with
my
quick
compilation
is
ended
and
that's
a
bummer
but
yeah.
B
About
the
10
best
eviction,
so
what
update
is
that
I
have
some
conversation
with
surviving
in
the
car,
so
Robbie
and
I
are
working
on
that
so
right
now,
Revere
most
focused
on
the
increase
that
that's
coverage
of
a
urine
test
and
booking
on
the
e
to
e
test.
So
once
we
are
down,
we
kind
first
working
on
improve
the
coverage
on
any
questions
estimated.
So
and
another
thing
is
actually
it's
about
me.
So
actually,
I
didn't
I,
didn't
understand
a
tank
based
addiction
crackery.
B
So
my
my
understand,
who
was
that
the
picture
can
chose
that
if,
if
there
is
a
no
executing
and
it
will
give
Vic
to
the
part
right
but
and
it'd
be
his
behavior
is
not
controlled
by
10
best
eviction,
it's
just
controlled
by
internal,
no
execute
and
manager.
No
matter
can
basically
be
changing
in
be
enabled
or
not.
So
what
exactly
can
base
the
eviction?
Controls
is
actually,
if
there
is
some
already
known
already
or
ritual
condition
comes
out,
and
this
feature
is
enabled
in
well.
E
B
A
E
E
B
B
Already
exists
and
when
I
checked
each
we
test
found
funny
is
that
the
test
that
was
named
by
a
Pacific's
and
a
Scott
has
a
dog
that
goes
so
it's
not
enabled
by
default
in
to
you,
because
it's
2d,
the
kind
of
like
union
task
so
I
raised
a
few,
are
two
kind
of
also
be
opposed,
rename
it
and
also
increased
some
coverage.
True
tested,
multiple
paths
and
also
to
test
on
multiple
toleration
which
have
different
origin
seconds
on
the
switcher.
B
F
A
A
F
B
One
last
thing
about
me:
at
the
Interpol
affinity,
which
we
are
going
to
support
its
mentioned,
multiple
parts
so
about
their
features.
We
need
to
kind
of
redesign
right,
so
I
have
worked
out
a
draft
of
ocean
but
haven't
checking
yet
so
I'm
gonna
I'm
gonna
create
some
benchmark
tests.
I
have
some
conversation
with
Swiss
with
Harry
and
I
will
create
some
benchmark
test
and
verify
that
it
works
well,
so
I
hope
I
can
I
can
check
him
before
Jonathan's
face
refactoring
I
don't
want
to
because
three
stories
kind
of
all
predicates
priorities.