►
From YouTube: Kubernetes SIG Storage 20170216
Description
Kubernetes Storage Special-Interest-Group (SIG) Meeting - 16 February 2017
Meeting Notes/Agenda: https://docs.google.com/document/d/1-8KEG8AjAgKznS9NFm3qWqkGyCHmvU6HVl0sk5hwoAE/edit#heading=h.vfa89xk0z3sk
Find out more about the Storage SIG here: https://github.com/kubernetes/community/tree/master/sig-storage
Moderator: Saad Ali (Google)
Chat Log:
09:42:45 From serge : will do
09:56:46 From Chris Dragga : I think to really make this work for replicated file systems, you may need storage plugins specific to the file system.
09:59:04 From stephenwatt : Got to drop. Good discussion!
A
A
Items
that
folks
are
working
on
for
this
quarter
as
a
reminder:
release
code
freeze
for
1.6
in
one
week
away.
We
have
next
week
and
then
the
following
monday
is
code
freeze.
So
if
you
have
features
in
flight,
please
make
sure
that
they
get
time
to
make
sure
we
check
in
on
what
we're
planning
for
this
release.
Let's
apply
just
go
over
the
spreadsheet,
meaning
I'll
see.
If
I
can
share
my
screen.
B
A
Cool,
so
out
of
tree
volume,
plugins
was
supposed
to
be
in
design
this
quarter,
so
the
announcement
here
is
that
it
is
currently
in
the
works.
A
lot
of
folks
me,
Michael
tim
has
started
talking
to
folks
from
mezzos
cloud
foundry
and
dr.
about
having
some
sort
of
standard
interface,
a
container
storage
interface
that
we
can
all
agree
on
as
a
standard
model
and
then
kind
of
a
dessert
we're
hoping
the
implementation
falls
out
of
that.
So
those
discussions
are
on
going
we're
going
to
start
sharing
the
notes
from
that
more
widely
right.
A
B
F
A
G
A
H
So
I
put
out
a
design
proposal
is
paying
it's
like
the
coaching
yourself,
I've
started
barely,
but
there's
a
lot
of.
We
are
getting
close
to
a
final
design.
That's
like
that
should
be
done
by
today.
I
hope
and
I
should
be
able
to
present
the
work
sometime,
like
you
know,
like
maybe
wednesday
next
week.
Actually
the
first
cut
of
the
code
for
that.
Okay.
A
And
so
I
think
the
last
open
question
there
was
whether
we
were
okay
bumping,
that
field
from
volume
specific
to
a
first
class
field
globally
yeah
in
the
pod
definition-
and
I
think
the
answer
to
that
is
it
should
be
fine.
It
is
different,
for
example,
FS
tight,
but
it's
not
completely
unheard
of,
for
example.
Mount
path
currently
is
a
first-class
feel
that
gets
interpreted
by
volume
plugin,
so
having
mount
options
be
a
first-class
feel
that
gets
interpreted
by
volumes
is,
should
be
ok
as
well.
Ok,.
I
Michelle's
got
this
amazing
proposal.
Just
go
one
item
back
on
this:
persistent
storage
for
local
disk
will
block
devices
there's
loads
of
feedback
on
it,
but
I'm
noticing
the
usual
characters
in
the
community
that
are
trying
to
run
data
management,
platforms
and
stuff
like
I'm,
not
without
commenting
on
it,
but.
E
I
Take
any
case-
yes,
maybe
they
just
don't
sort
of
know
about
us
or
so
I'm,
trying
to
take
a
second
if
you're
working
on
something
like
1,
I'm
cassandra
in
cube
using
local
disks
or
or
glycerol
safe,
or
you
know
any
kind
of
storage
platform
or
something
that
uses
loaf
of
this
fit
erase.
That
provides
an
API
for
consumption
of
that
storage.
This
is
a
freaking
massive
proposal.
It's
hugely
important
as
big
as
dynamic
provisioning.
In
my
opinion,
you
didn't
my
vacation
too
yeah.
G
A
J
I
think
we've
made
some
really
good
progress
with
the
testing
getting
rid
of
flakes
getting
quite
a
few
PRS
merged.
Since
the
last
meeting,
I've
also
had
some
more
people
ask
to
be
invited
to
the
meeting
through
the
community,
so
I
think
we're
getting
more
involvement
in
more
buy-in,
we're
continuing
to
work
on
the
destructive
testing
and
some
of
the
the
high
priority
tests,
including
the
the
wedging.
So
we're
on
track
and
I
look
forward
to
seeing
some
new
members
tomorrow
at
the
meeting
an.
A
H
So
Prasad
payment,
oh
yeah,
like
we'd
CFS
of
working
on
this,
like
we
have,
we
have
calls
Thursday
1pm
jealous.
If
we
have,
we
have
consulted
some
items
like
like
access
to
track
on
and
they're
working
to
so
the
things
working
on
the
caching
proposal.
I
opened
a
proposal
for
bulk
volume
that
will
the
bulk
volume
pollen
pulling
and
I'm
going
to
have
a
peregrine
have
a
call
today
to
discuss
that
proposed.
Also
that
we
eliminate
some
extra
cause
that
we're
making
in
a
SS.
K
L
L
H
A
So
some
of
the
items
is
there
purely
bug
fixes
could
theoretically
be
done
after
code
freeze,
but
it
sounds
like
a
lot
of
fitted
feature
work,
so
I
definitely
bring
that
up
at
the
meeting
today.
If
you
can
okay
sure
we
reminder
to
everyone.
Alright
cool
next
up
is
the
NFS
wedging
issue.
This
I
think
we
may
have
dropped
the
ball
on
a
little
bit.
Does
anyone
know
if
anyone
took
ownership
of
this.
J
A
A
A
N
F
A
Not
expand
some
of
your
reviews
to
John
yeah
this.
That
would
be
great
yeah
cool,
okay,
Thank,
You,
Brad,
alright,
so
next
up
is
default.
Storage
classes
for
cloud
providers.
Yawn
already
had
the
pr
out
for
this
I
think
the
only
open
items
here
were
coming
up
with
a
better
mechanism
for
being
able
to
disable
the
default
storage
classes
because
they
request
for
add-ons
and
they
require
access
to
the
master,
to
delete
I.
Think
Michelle
had
a
work
that
she
had
a
bug
that
she
was
tracking
Michelle.
G
G
A
H
A
O
A
P
Q
A
A
That
would
be
awesome
and
we've
got
a
week
left,
so
we
may
need
to
scramble
some
resources
to
get
this
done,
and
then
we
need
to
decide
if
it's
a
blocker
or
not
for
going
ga
for
storage
glasses.
The
other
open
issues
for
going
ga
for
storage
classes
is
whether
we
are
allowed
to
promote
the
storage
class
field
to
a
first
class
field
in
pvcs.
I
think
brian
grams
had
some
input
there,
but
he's
out
in
tahoe
at
the
open
source
can
leadership
convention.
He
should
be
back
next
week
and
will
have
his
input.
A
P
Was
another
discussion
around
the
attribute
stuff?
Is
that
right
now,
with
TVC
anybody
being
able
to
set
the
attribute
for
TVs
to
be
belonging
to
a
storage
classes
that
basically
reduce
the
storage
class
down
to
a
label
and
I,
don't
know
if
the
seeker
6
feels
that
they
should
be
a
stronger
relationship,
the
similar
to
what
exists
between
claims
and
PV,
where
the
relationship
between
the
PV
and
a
storage
class
is
more
hardened
so
that
it
actually
needs
to
be
something
that
really
belongs
so
the
stuff
not
just
0e?
P
A
P
A
A
Storage
classes
to
GA,
we
already
went
over
metrics
no
capabilities.
I
have
not
started
working
on,
but
it's
supposed
to
be
designed
for
this
quarter,
so
I
plan
on
starting
on
it
after
code,
complete
our
code,
freeze,
snap,
shots
I
think
we
do.
You
know
what
the
status
of
this
is
I
believe
we're
punting
on
this.
N
N
I
Also
just
to
address
where
you
know
for
openshift
we're
pretty
big,
I'm
trying
to
get
snapshots
in
sooner
rather
than
later,
so,
we've
also
had
on
the
Red
Hat
side
internal
design
discussions.
So
maybe
it
might
be
an
idea
just
to
get
together
with
the
folks
that
are
designed
just
designed
together
just
to
move
the
ball
forward
on
the
set
yeah.
A
L
B
H
L
A
Thank
you
do
that
next
up
is
bringing
storage
classes
to
GA,
we've
already
discussed
it.
There
are
some
open
items
that
may
put
this
at
risk.
We'll
carry
this
discussion
offline
and
then
hopefully
have
a
better
answer.
Next
week
you
doing
metrics
on
volume
controllers,
I
thought
we
already
reviewed
that
wasn't
that
tracked
above
somewhere.
M
A
Let
me
mark
it
as
orange
as
a
reminder
to
myself
to
go
in
and
review
the
proposal,
but
yet
it's
non
critical
for
this
release.
Yeah
we
great
if
we
can
get
it
in
if
not
no
big
deal
next
up
is
the
deli
MC
scale,
io
volume
plugin.
I
think
this
has
been
blocked
on
me
for
code
review,
but
I've
assigned
it
to,
I
believe,
jing
or
michelle
to
take
a
look
at
it.
A
R
A
A
A
A
I
L
I
I
A
A
G
A
G
A
A
I
The
late
you
can
exchange
slips
up
the
color
interesting,
like
I,
didn't
I
looked
at
James
proposal.
Initially
I
had
a
misconception
that
I
discovered
later,
which
was
that
things
like
glossary.
When
you
take
a
snapshot,
it's
immediately
available
as
an
additional
money
that
you
can
mount
that
snapshot
right
away,
which
is
a
true
for
a
lot.
You
know,
apparently,
all
the
cloud
providers
a
lot
of
matter
of
Technology,
so
the
axe
phoenix.
I
What
I
wasn't
thinking
one
was
before
I
discovered
that
was
sort
of
being
able
to
support
of
some
sort
of
annotation
on
a
claim
or
something
that
would
kick
off
the
semantic
profit
profit
snapshots
of
work
possess
the
back
ends
of
life
I
associated
with
the
distance
and
volume
and
create
a
mirror
for
system
volume
and
claim
for
that
cream.
So
that's
the
developer
could
initiate
the
snapshot,
see
them
all
themselves
when
they
do.
You
know
to
control
gear.
Pvc
then
see
the
fair
shot
in
the
list.
I
You
know,
label
claims
act,
snapshots
or
something
like
that,
but
that
all
got
a
bit.
I'm
aroma.
When
you
know
it's
not
as
simple
as
just
creating
it's
like
a
two-phase
process
create
snapshots,
and
then
you
know
turn
that
that
snapshot
into
an
EBS
s-
or
you
know
something
to
that
effect.
So
anyways,
like
I,
guess
I'm
just
interested
in
hearing
what
jingle
thinking
about
her
approach
around
that
problem,
and
you
know
just
moving
the
ball
forward
over
the
line.
I
wonder.
A
I
L
I
T
I
L
M
Yup
I
mean
no
I
did
want
to
our
final
touches
on
that:
okay,
massage
just
photos
of
who
is
actually
working
at
the
young.
The
story
slide
on
and
go
back
inside.
So
when
we
are
forced
to
PR
three
know
who
should
contact
you
for
the
youth?
Sorry,
three
working,
my
god
is
actively
owner
story
side,
not
here.
So
what
we
are
now
communicates
with
you.
You
know
who
have
the
bandwidth
to
review
and
well.
A
Who
should
be
reviewing
us
so
it
should
be
pretty
well
distributed
across
the
storage.
Big
I
think
pretty
much
everyone
who's
been
contributing.
The
code
can
be
a
reviewer
and
I
think
they're
automatically
reviewers.
There
are
a
handful
of
approvers,
but
as
far
as
I
understand,
approvers
are
more
or
less
rubber
stamps,
so
they
are
to
the
last
case
with
the
reviewers.
It
looks
good
the
approver
just
come
in,
take
a
quick
look
at
it
and
improve
it.
So
you
should
be
able
to
spread
the
bandwidth
across
pretty
much.
The
whole
team
are.
A
A
I
think
so
they
we
have
these
owner
owners
files
and
inside
the
owners
files
there
are
approvers
and
reviewers
defined
for
each
each
directory
and
the
the
system
are
automatically
will
distribute
code
reviews
across
the
reviewers
and
that
reviewer
list
is
fairly
large
and
if
you'd
like
to
be
added
to
that
reviewers
list,
please
just
create
a
PR
and
add
yourself
to
it
and
sign
it
to
me
or
Tanner.
All
these
folks,
okay,.
A
T
S
S
A
L
A
C
A
B
A
D
A
J
Have
I
made
sure
on
our
side
that
it
would
not
be
a
huge
impact
openshift,
probably
direct
management
agreed
that
they
could
just
play
in
the
current
release.
That
would
be
deprecated
in
the
next.
So
the
business
impact
on
our
side,
which
I
think
was
we
were
worried,
was
going
to
be
the
greatest
minimal,
so
I'm
happy
to
just
I
guess
I
need
to
what
do
I
need
to
do
then.
Do
I
need
to
go
to
Brian,
grant
Mac,
we
can
have
an
exception
or
what
what's
the
are.
A
So
so,
just
to
put
this
into
context
for
everyone
else.
The
proposal
here
is
the
acceleration
deprecation
process
for
the
recycler
policy,
because
we
technically
been
trying
to
deprecate
it
for
a
longer
period.
The
social
I
think
brad
had
additional
deprecation
proposal
that
was
as
up
in
merge
chef
Brad
thank.
F
A
So
regardless
I
think
is
that
we
may
not
have
followed
the
official
process
yet
to
the
official
countdown
clock,
which
is
it
here
or
major
release
is
basically
screwing
us
over
and
we're
going
to
see
if
I
could
get
a
a
get
around
that,
but
yeah
Brian
grant
will
be
the
right
person
to
talk
to
that
about
Aaron.
Take
the
lead
on
that
that'd
be
great
yep.
A
O
L
A
A
U
A
G
U
U
A
G
I
But
then
you
have
to
carefully
curated
look
like
that
right
now:
sort
of
adaptive,
self-healing
capability
with
includes
an
ad
that
you
can
leverage
with
that
pattern.
So
if
we
can
take
local
file
systems,
you
know
in
direct
attached
storage
and
expose
them
as
persistent
volumes.
Then
we
can
hooker
things
like
respectful
set.
Then,
basically,
I
believe
idea
is
to
update
their
scheduler
with
some
additional
scheduling.
Intelligence
that
says.
Look
at
this
this
part
or
replica
in
the
stateful
phase,
is
using
persistent
longing.
That
happens
to
be
a
local
disk.
I
And
then
for
a
faithful
set
like
so
for
self-healing,
capability
or
Kinesis
to
say
I
have
a
sec,
no
Cassandra
cluster.
So
since
my
fateful
stateful,
Cassandra
and
I've
got
you
know
six
persistent
bonding
queen,
but
are
all
local
disks
associated
with
each
replica
and
staple
set
that
that
that
works?
That's
also,
but
if
one
dies
that
we
obviously
don't
have
the
capability
to
dynamically
provision.
So
you
just.
We
just
have
like
service
life,
additional
persistent
volumes
that
can
then
be
used
for
that
Rick.
Okay,
in
the
faithful
thing,
yeah.
G
V
G
O
G
G
G
I
And
then
varying
levels
of
forgiveness
right
so,
like
you
know,
you
want
to
avoid
creating
a
data
replication
storm
every
time.
I
know
it
goes
down,
so
you
could
play
like
look
if
you
know
parts
being
unavailable,
41
minutes
in
the
stables.
Let's
do
nothing,
you
know
if
it's
been
unavailable
for
one
hour
do
nothing,
but
if
it's
been
available
unavailable
for
one
day
replace
it,
it
had
yeah.
G
I
I
But
your
stakes
will
set
my
sony,
be
5,
5,
5,
45
minutes,
so
it'll
go
just
pick
5
and
go
run
them
according
to
whichever
maybe
quality
of
service
each
specified
for
your
local.
Just
that
you
want
and
we'll
just
go
around
them
and
it
will
find
of
those
die.
So
let's
go
find
five
other
nodes
with
the
similar
local
this
and
is
automatically
going
to
make
sure
it
starts
up
and
runs
over
there.
So
there's
a
huge
step
forward
and
self
healing
capabilities.
K
I
have
a
question
so
looks
like.
If
the
note
goes
down,
I
mean
chances
are
for
many
of
these
local
storage
systems.
Its
data
is
replicated
and
if
a
note
goes
down,
you
just
don't
want
to
grab
a
new
TV
cool.
You
want
to
be
able
to
grab
a
TV
at
points
they're
replicas,
so
your
application,
you
know,
can
resume
already
left
off.
Is
that
something.
K
Yes,
yes,
sets,
are,
you
know,
instance,
that's
running
on
a
machine,
but
many
of
the
posts
like
sources
that
you
know
they
have
the
data
replicated
on
different
notes,
and
when
that
machine
crashes,
we
can
launch
that
instance
on
a
different
note.
That
has
a
replica
of
the
data
that
was,
you
know
you
so,
instead
of
erratic
a
new
TV
that
points
to
scratch
space,
you
want
to
be
able
to
point
with
volume
that
has
a
copy.
Okay,.
G
L
O
That's
a
that's
right.
I
mean
I
concur
with
what
the
hour
earlier
person
said.
It
may
not
have
a
pot
on
it
so
like,
for
instance,
with
port
works.
We
may
have
another
replica
set
placed
on
a
different
note.
So
let's
say
there's
10
nodes
in
the
cluster.
Three
of
those
nodes
may
actually
already
contain
the
data,
so
it
would
be
preferred
to
take
any
one
of
those
others.
We
know
that
application
will
just
review
again,
even
that
we
do
accomplish
today
by
way
of
node
labels
and
note
collectors.
O
I
Like
sense,
that
makes
sense
yeah
but
sounds
usually
so
I
get
it,
but
with
data
replication
you
usually
want
to
make
some
sort
of
quorum
rights
of
you
know.
You
want
two
replicas
or
three
red
Booker's.
So
if
you've
lost
the
note
you're
down
a
replica,
so
you
want
to
basically
be
able
to
you
know,
generally
speaking
across
mostly
data
management
platforms
right
with
it,
whatever
it
is
because
spin
up
a
new
pod
on
a
new
host
with
the
blank
disk
and
have
that
sets
be
available
and
then
rebound.
L
So
that's
actually
one
of
the
questions
that
I
have
a
multiplication
document
because
obviously
is
sort
of
similar,
but
not
exactly
so.
If
it's
MongoDB
kind
of
application,
then
you're
right,
it's
exactly
like
the
cell
replicates
down,
which
means
the
manga
video.
Whatever
the
databases
is
one
sharp
fewer.
You
need
to
create
a
new
short
and
repopulated,
but
if
Akitas
volumes
you
might
want
to
start
with
that
replica
as
a
seed
for
the
new
shark
yeah.
I
L
I
More
like
a
block
replication
as
an
application
right.
So
if
I,
if
I'm
understanding
you
correctly
we're
talking
about
like
hey
I,
have
you
know,
Dave
ACB
on
turtle,
one
and
I,
and
maybe
I
have
a
file
system
on
it
and
then
I
want
to
replicate
the
content.
Use
some
mechanism
to
replicate
the
context
of
this
Davis
BB.
To
give
FDB
on
server2
is
that
right,
yeah
for
its
Bella.
L
K
I
L
T
A
G
G
Okay,
yeah
so
yeah.
Thanks
for
the
explanation
we
were
targeting,
we
were
mostly
targeting
we
use
cases
where
applications
do
the
application
level
replication
and
not
so
much
the
volume
level
replication.
We
can
definitely
think
about
how
how
this
can
support
the
use
cases
also.
Maybe,
if
you
can
send
me
some
more
details
about
that
use
case
specifically,
if
we
have
anything
like
that,
hey
excusing
that
doing.
That
would
be
good
to
see.
R
E
R
Q
R
Time
so
you
have
the
biscuits
actually
only
one
pd,
logically,
it
just
change
its
location
transfer.
The
previous
active
note
to
the
new
active
note,
so
you
don't
have
to
express
it.
Soupy
DS.
You
can
express
this
as
one
pd.
That's
currently
have
a
location
on
notte
now
switch
location.
To
note,
be
that
makes
things
simpler
if.
L
I
L
A
D
Thanks
yeah
I
think
that
this
is
a
hybrid
between
local
storage
and
volume
mouth,
because
you've
got
two
sons
growing
in
here
that
there's
it's
not
an
open-ended
external
attached,
because
there
are
constraints
on
what
compute
node
it
might
be.
Attachable
to
and
potentially
with
some
forms
of
replication
is
glossy
so
that
it
you'd
view
it
as
essentially
at
the
image
copy
that
doesn't
have
hundred
percent
fidelity,
but
maybe
the
app
it
could
take
advantage
of
that
and
save
effort
on
the
rebuild
yeah.