►
From YouTube: Kubernetes SIG Storage 20180719
Description
Kubernetes Storage Special-Interest-Group (SIG) Meeting - 19 July 2018
Meeting Notes/Agenda: https://docs.google.com/document/d/1-8KEG8AjAgKznS9NFm3qWqkGyCHmvU6HVl0sk5hwoAE/edit#heading=h.wej10zne6qo2
Find out more about the Storage SIG here: https://github.com/kubernetes/community/tree/master/sig-storage
Moderator: Saad Ali (Google)
Chat Log:
19:36:46 From Jan Šafránek : I made the hidden lines appear, there was filter on Target Status column to list only alpha features
19:39:12 From Saad Ali : Thanks!
A
A
So
if
you
have
anything
that
you
want
to
discuss,
please
feel
free
to
add
it
to
the
agenda.
High-Level
we're
gonna
go
over
status
for
the
different
features
that
we're
tracking
for
the
current
release.
Then
we're
gonna
discuss
any
PRS
that
need
attention
bugs
that
we
want
to
discuss
and
finally
design
reviews.
So
let's
kick
it
off
by
taking
a
look
at
the
planning
spreadsheet
and
doing
a
review
of
the
features
that
we're
tracking
for
this
release.
A
We
have
a
lot
of
them
I'm
going
to
skip
over
the
CSI
driver
statuses,
because
there's
a
lot
of
CSI
drivers
and
they're
not
going
to
be
entry,
so
we
can
get
an
end
of
quarter
status
at
the
end
of
the
quarter,
instead
of
doing
a
weekly
or
bi-weekly
update
on
them.
So
first
up
is
implementing
out
of
tree
volume,
snapshot
and
restore
controller.
B
Okay
yeah,
so
we
had
a
few
design
meetings
had
some
good
discussions
and
design
spec
was
being
reviewed
and
updated.
We
also
submitted
the
snapshot
API
code
PR,
and
that
was
being
reviewed
as
well.
B
C
A
A
Looks
like
he's
not
on
the
call
today.
I
can
try
and
give
an
update
on
his
behalf,
so
the
meetings
were
discussing
whether
a
registration
mechanism
is
necessary
or
not,
and
if
so,
what
would
it
look
like?
Would
it
be
a
CR
D?
Would
it
be
an
entry
object?
Could
we
even
reuse
the
existing
node
objects
at
the
moment,
we're
leaning
towards
doing
a
first
class
entry
object.
Yan
is
planning
on
putting
together
a
proposal
for
cig
architecture
to
see
what
their
thoughts
on
it
are
and
based
on
that
we'll
proceed
from
there.
A
Next
feature
is
skipping
external
attach
detach
for
non
attachable
volumes
for
CSI.
This
has
a
dependency
on
the
plug-in
registration
mechanism.
If
the
plug-in
registration
mechanism
goes
in,
this
feature
can
use
that
mechanism.
Otherwise
we'll
have
to
come
up
with
something
else
here.
So
it's
waiting
on
number
5
to
get
unblocked
next
up
is
the
mount
library
Travis.
Do
you
want
to
give
an
update
here.
D
D
A
A
A
A
A
So
this
is
a
feature
of
for
CSI,
where
CSI
currently
is
only
accessible
through
PV
PVC,
but
for
ephemeral
volumes.
Ideally,
we
want
it
to
be
accessible
through
inline,
but
at
the
same
time
we
don't
really
want
attachable
volumes
to
be
accessible
in
line,
so
it
puts
us
in
a
weird
spot
and
the
design
here
is
to
come
up
with
solution.
Yon
has
a
proposal
out.
Please
take
a
look
at
it
and
provide
feedback,
currently
its
outstanding
waiting
feedback.
A
A
Next
up
is
the
CSI
spec
having
work
load,
information
passed
to
the
node
published
call.
This
was
one
of
the
items
we
believe
we
need
for
a
specific
type
of
volumes
that
require
identity,
pod
identity
and
we're
trying
to
work
through
those
use
cases.
At
the
moment
we
had
a
CSI
community
call
yesterday,
where
we
discussed
this
at
at
length.
A
E
F
Yes,
sorry
so
we
got
the
CSS
spec
mouthful.
You're
gonna
have
a
call
on
Monday.
Like
we
act,
we
had
to
have
a
general
consensus
about
it.
You
had
a
call
about
CSI
driver
is
history
on
Monday
or
last
Monday,
and
the
idea
is
like
we
have
a
general
consensus
how
to
proceed,
but
we
have
some
remaining
items
that
made
you
need
to
figure
out
like
how
to
take
into
account.
F
F
So
we
had
a
call
CSI
and
the
CSI
community
meeting
yesterday
and
we
agreed
on
the
broader
approach
of
not
calling
it
matrix,
but
calling
it
like
capacity
mission
that
will
expose
the
both
eye
notes
available,
bytes
available
that
we
have
currently
and
allied
open
a
proposal
in
CSI
to
expose
that
capacity.
Information.
A
new
RPC
call
at
the
node
code
level
and
that
can
be
used
by
by
communities
and
other
SEOs
are.
E
F
I
notes
could
be
faced
from
the
yeah,
the
using
Stata
fiscal.
So
that's
a
that's.
That's
a
that's
an
old
property
where
the
load
is
hosted.
The
capacity,
though,
is
not
depends
on
how
the
volume,
what
type
one
column
is,
so
the
plug-in
will
report
both
and
it
will.
It
will
report
from
the
node,
so
I.
A
E
E
F
E
G
F
G
H
F
F
All
right,
so
this
is
I
think
this
is
the
one
that
we
are
talking
about
earlier,
like
the
the
things
that
are
left
to
be
discussed,
because
certain
volume
types
not
only
there
is
a
limit
account
of
how
many
volumes
you
can
attach,
but
there's
also
how
much
you
can
attach
a
capacity
related
constraint.
So
I
think
this.
This
is
kind
of
like
I'm,
sorry.
F
I
A
A
F
There's
a
spec
already
out,
but
there's
like
at
least
like
I,
think
there's
a
couple
of
items
that
are
one
item
that
is
pending
is
whether
the
online
next,
the
capability
of
what
kind
of
expansion
the
volume
the
column
plug-in
supports
online
versus
offline,
whether
it
should
be
exposed
as
a
capability
or
not
is
that
one
is
not
quite
nailed
down
yet
and
I.
Think
that's
where
the
mean
this
thing
is.
We
had
a
discussion
last
week
in
CSI
this
week.
We
didn't
talk
about
this
particular
thing.
A
Sounds
good,
thank
you
come
on
moving
on,
we
have
a
few
CSI
drivers
which
I
am
going
to
skip
over.
If
you
are
a
feature
owner
for
the
CSI
driver,
please
continue
to
work
on
it
and
we'll
get
a
status
update
near
the
end
of
the
quarter,
CSI
library,
the
Sif's
common
library.
Deep,
do
you
want
to
talk
about
this?
Yes,.
H
H
H
This
insig
windows
I
think
that
the
decision
was
pretty
much
that
there
was
no
no
confidence
case
where
this
would
be
used
in
Windows
and
since
the
support
doesn't
exist,
it
will
take
a
bit
of
work
apparently
to
surface
the
the
block
devices
through
the
windows
containers.
So
as
a
result,
unless
there's
like
explicit
customer
requirement
or
a
use
case
around
it,
okay.
H
A
See:
okay,
fair
enough,
so
we'll
go
ahead
and
scratch
it
for
this
quarter,
then,
and
if
anybody
believes
that
block
support
for
Windows
is
more
critical,
please
speak
up.
A
A
A
G
I've
only
started
looking
at
this
in
detail
this
week,
so
I
don't
know
for
sure,
but
my
suspicion
is
that
it
probably
there's
no
reason
to
believe
yet
that
it's
going
to
be
needed
at
all.
It's
just
amount.
So
if
that
changes,
all
oh
I
should
find
out
soon,
because
I'm
really
spending
a
lot
of
time
with
the
CSI
NFS
driver.
This
week,
perfect.
J
F
A
F
I
I
A
A
A
I
A
I
E
A
I
A
A
A
G
Well,
yeah
I
only
bring
it
up
because
I'm
I
don't
understand
like
what.
What
didn't
go
right
here,
because
the
code
was
more
or
less
ready
for
111
and
it
was
sort
of
late
in
the
release.
So
I
mean
I'm,
not
super
mad
I.
Didn't
it
didn't
merge,
but
I
want
to
know
like
what
should
I
have
done.
Should
I
have
been
poking
people
personally,
yeah.
A
G
A
G
A
E
E
We
already
have
some
tests
here
for
ephemeral
volumes.
As
part
of
this
we
can
evaluate
if
there
are
more
test
cases,
we
should
add
to
the
ephemeral
volumes
such
as
the
sub
path
feature,
and
maybe
some
I'm
not
sure-
maybe
some
node
failure
handling
cases,
some
new
ideas
that
I
had
of
test
cases
that
could
fit
in
core
performance.
Our
PDC
features
that
don't
require
a
specific
volume
plug-in
to
actually
test.
So
these
are
things
like
pre,
binding
or
a
static
binding.
E
The
PV
protection
feature
testing
that
we
claimed
retain
reclaimed
policies,
some
things
like
that.
So
that's
the
basic
idea
behind
the
first
suite
the
then
there's
the
second
suite
which
I
am
calling
the
persistent
volume
plug-in
profile.
This
is
the
profile
suite
I
think
is,
is
basically
a
concept
where
it's
not
strictly
required
for
kubernetes
conformance,
but
it
can
be
some
optional,
additional
formed
in
sweet
that
kubernetes
distributions
can
optionally
conform
to
so
here.
E
So
those
kind
of
things
that,
like
you,
know,
basic
functionality
that
every
plugin
should
support
can
go
here,
and
the
idea
with
this
profile
suite
is
that
distributions
can
bundle
one
or
more
persistent
volley
plugins
by
default
in
order
to
conform
to
this
profile
suite
and
then
the
third
category
of
tests
is
basically
the
this
is
sort
of
like
that.
Everything
else
suite
where
we
test
a
lot
of
features
that
not
all
volume
plugins
may
support.
E
E
E
E
E
E
That's
correct
so
I
want
to
formulate
some
of
these
test
cases
from
the
pod
perspective
and
also
like
the
node
perspective,
so
like
say
like
the
pod
can
recover
if
it
gets
rescheduled
on
a
different
node
or
it
can
recover
if
the
node
fails
or
it
can
recover.
If
the
node
reboots
you
test
cases
like
that,
I
love.
H
E
I
think
there's
right
now:
I
just
have
like
a
really
rough
idea
right
now,
I'm,
like
possible
scenario,
so
test
I,
think
when
we
actually
go
and
investigate
writing
these
test
cases.
We
can
actually
work
out
those
details,
so
I
have
a
spreadsheet
that
I
started
drafting
some
more
specific
test
cases,
although
right
now
the
spreadsheets
also
pretty
generic
and
not
very
detailed,
but
I-
think
the
idea
is
each
each
kind
of
common
class
of
test
cases
we
want
to
add.
E
We
can
open
up
an
issue
for
that,
and
then
we
can
work
out
more
of
the
details
of
exactly
what
the
test
needs
to
do
and
how
to
propagate
or
how
to
you
know,
trigger
some
of
those
failures.
So
if
you
guys
have
any
ideas
of
more
test
cases
that
we
can
add,
feel
free
to
add
them
to
this
spreadsheet
I
have
I've
given
six
storage,
edit
capabilities.
E
I
E
I
definitely
think
like
workload.
Integration
should
be
part
of
this
too
I
forgot,
where
I
put
it
I
might
have
put
it
under
optional
features,
or
did
they
put
it
under,
maybe
I
put
on
their
second
one.
I
had
like
one
bullet
point
about
stateful
sets
yeah
workloads.
We
can
potentially
add
some
more
things
here
and
you
know
flush
it
out
a
little
more
there's.
I
K
E
I
think
we'll
have
to
think
about
it
like,
like
from
a
user's
perspective.
It's
definitely
important
to
from
a
user
that,
like
their
volume,
is
going
to
behave
properly
in
disruptive
scenarios
we'll
have
to
see
we
might
not
be
able
to
run
it.
We
might
need
to
make
changes
to
the
way
that
conformance
tests
are
run
to
be
able
to
allow
distrub
disruptive
tests.
E
E
E
I
C
E
Assuming
well,
this
theoretically
is
just
e
to
e
test
changes
like
we're
not
making
changes
to
the
kubernetes
core
code
in
order
to
enable
this.
So,
theoretically,
if
we,
you
know,
write
this,
we
can
potentially
run
the
test
against
older
clusters
too.
I
think
that
that's
possible
I
think
from
a
conformance
point
of
view,
I
don't
think
they
allow,
like
backporting
tests,
conformance
tests
to
previous
releases
right.
E
I
think
I've
liked
that
the
test
framework
could
be
made
to
like
work
against
other.
You
know
different
kubernetes
releases
I,
think
that's
fine,
so
that
we
can
act
for
conformance
and
might
only
run
against
the
latest
kubernetes
releases.
But
for
just
a
testing
point
of
view,
we
can
certainly
run
these
tests.
E
L
I
L
E
Think
it's
more
so
I
don't
think
we
have
like
plugin,
there's
no
plugins
certification.
So
all
of
this
has
to
be
tested
against
a
kubernetes,
distro.
So
I
think
it's
up
to
the
kubernetes
distributions
to
choose
which
volume
plugins
they
want
to
test
for
conformance
and
say
that
you
know
certify
that
this
volume
plug-in
works
well
on
their
environment.
A
K
Where
I
went
through
the
existing,
it
would
just
be
happy
to
burn
at
least
for
storage.
I
went
through
the
cloud
from
photo
volume.
Plugins.
There
are
cloud
based.
We
have
ng
of
test
for
Amazon,
as
your
UCPD
opens.
The
can
side,
and
this
fear
we
don't
have
any
foreigner
of
fire,
but
it
seems
nobody
cares
and
then.
K
The
test
job
that
actually
runs
the
test
and
the
end
it
gives
the
different
picture
because
the
tests
AWS
and
the
test
GCP
you
none
any
tests
on
a
hurry.
We
don't
have
OpenStack,
we
don't
dismiss
fear
I,
know
that
we
remember
run
bunch
of
their
own
into
invest
and
they
are
fixing
the
bugs
and
pushing
PRS.
But
I
don't
know
if
Microsoft
or
somebody
who
cares
about
OpenStack
runs
the
test,
but
I
don't
think
I
can
do
anything
here.
I
hope
Microsoft
does
around
yeah.
H
K
K
Know,
rich
OS:
there
is
no
test
there
and
I
looked
at
for
free
test
behalf
and
looked
which
do
you
run
and
we
run
quite
a
lot
of
them,
but
we
don't
run
safe,
X,
cozy
and
safe
RDD
and
I
would
like
to
run
them
because
we
have
to
technology
it's
open
source.
We
can
run
it.
Nothing
is
preventing
us,
except
for
the
images
that
the
run
end
to
end
tests
for
are
all
the
tests
run
on
GC
I,
which
is
immutable.
K
So
I
fought
about
it.
How
to
write
and
I
found
couple
of
issues
like
that
surf
server.
Image
is
for
definitive
I
cut,
the
image
is
gone,
but
I
I
am
fixing
them,
but
if
we
add
a
new
job
that
would
run
all
the
tests
we
have,
including
self
and
I
see.
We
could
run
this
test
if
the
job
runs.
Ubuntu
instead
of
you
see,
I
I
know
take
that
couple
of
other
that
just
ordered
a
run.
You
boom-boom,
so
it
shouldn't
be
a
problem.
K
So
we
get
the
necessary
kernel
modules
and
we
also
have
the
possibility
to
run
the
mount
utilities
like
I,
scuzzy,
ATM
or
staff,
these
inside
containers,
and
not
to
have
them
on
the
host.
So
this
way
we
could
run
the
tests
in
some
end-to-end
job
with
every
commit,
or
at
least
once
a
day.
It
requires
couple
of
changes,
mostly
in
the
test
framework.
We
use,
for
example,
I
need
to
add
new
parameter
to
and
to
end
test
binary
that
will
run
the
container
of
the
computer.
K
It
is
on
every
host
before
the
test
start,
so
the
test
itself
expects
that
the
Monte
relatives
are
available
somewhere.
If
somebody
wants
to
install
to
run
the
tests
on
its
own
machine
with
the
utility
on
the
host
and
run
it
and
in
the
same
way
somebody
can
run
the
test
on
its
own
house,
the
wave
the
boundary
boundary
in
the
container
so
I
want
to
test
not
to
care
where
it
runs.
That's
why
I
need
to
end-to-end
test
suite
to
curl.
K
Well,
so
I
need
the
job,
I
need
the
new
parameter
and,
basically
that's
it
so
I.
What
I
need
here
is
somebody
from
6:00
to
show
you
this
combo
stick
which
perspective,
because
I
will
run
all
the
tests
that
have
6
storage
tag
and
I
need
to
skip
all
the
disruptive,
because
on
Ubuntu
it's
not
so
simple
to
restart
you
but
I
think
I
will
skip
the
flaky
I
work
in
the
stereo,
because
the
boots
everything
down
but
I,
will
include
or
slow
tests.
I
did
some
measurement.
It
will
take
16
minutes
fifteen.
K
Sixteen
minutes
run
all
the
tests,
including
one
other
thing
want
in
/.
Oh
yeah
and
everything
works.
So
again.
What
I
need
here
is
somebody
from
sixth
original
review
dates.
If
it
is
good
thing
to
do,
this
will
run
only
the
tests
which
we
have
right
now
in
the
future.
I
can
imagine
that
we
can
extend
to
the
tests
and
coverage
and
I,
don't
know,
run
run
full
example.
K
Soft
pass
test
for
iced
coffee
and
for
safar
baby
and
run
I,
don't
know
what
FS
group
tests
and
all
the
features
we
run.
We
test
if
other
volumes
which
we
could
test,
also
with
the
ice,
Kazi
ancef
and
maybe
in
the
future.
We
could
also
add
some
tests
for
what
works
or
for
which
OS
and
the
storage
vendors
that
they,
which
care
about
kubernetes
I
know
that
crocodile
I
haven't
heard
from
storage
OS
for
a
long
long
time.
K
E
A
E
A
K
A
We
do
something
ugly,
like
have
a
common
method,
that
all
the
storage
methods
will
run
and
every
test
must
pass
true,
true
or
false
to
it,
and
if
it's
true,
then
we
install
it
and
then
the
test
itself
has
to
determine
if
it
should
install
the
utility
and
it'll
have
to
implement
logic
like
if
I
am
you
know,
running
this
test.
With
this
specific
driver,
then
I
will
utility.
A
L
G
A
K
Another
thing
I
wanted
to
get
help.
This
is
I
want
to
write
through
some
room,
the
proof
606
testing.
So
if
anybody
has
any
contact
there,
I
would
like
them
to
approve
this.
Come
on.
Induction
me
in
this
job
would
make
sense
or
they
have
better
options.
I
take
whole
six
testing
on
the
PR,
but
yeah
I.