►
From YouTube: Kubernetes SIG Storage - Bi-Weekly Meeting 2023-03-23
Description
Kubernetes Storage Special-Interest-Group (SIG) Bi-Weekly Meeting - 23 March 2023
Meeting Notes/Agenda: -
Find out more about the Storage SIG here: https://github.com/kubernetes/community/tree/master/sig-storage
Moderator: Xing Yang (VMware)
A
B
A
B
Yeah
there's
a
peer
out
for
that
removes
the
I,
don't
know
if
it
there
was
a
PR
out
to
remove
the
likely
Mount
Point
check,
During
the
Reconstruction
right.
B
D
B
Yeah
I'm
sorry
I
was
saying
that
the
likely
is
that
issue
is
I,
think
fixed.
Now.
The
thing
is
that
now
the
next
thing
here
like
similar
to
that
one
is
that
we
assume
that
the
like
the
node
published
path
is
one
which
is
not
correct,
so
this
yeah,
so
this
score
now
we
are
trying
to.
What
we
are
trying
to
address
is
we're
trying
to
say
that.
Okay,
we
cannot
assume
that
the
Reconstruction,
the
the
what
the
paths
will
be
publish
path,
will
be
at
one
point.
So
that's
the
that's.
A
And
the
next
one
is
the
Warren
group,
so
one
group
snapshot,
so
this
one
yeah
we
are
making
progress
implantation
for
the
controller
are
ongoing,
so
we're
trying
to
wrap
it
up
and
also
there
is
a
placeholder
blog
PR,
so
so
the
date
for
blog
PR
to
be
ready
for
review
I
think
it's.
It's
April
4th
so
still
have
a
few
days
for
that.
A
Thanks
the
next
one
is
since
I
wouldn't
have
traditional
metrics
yeah,
so
this
one's
the
owners,
no
update,
I
need
to
ping
run
C.
If
he's
still
working
on
this.
A
Next
one
is
change
block
tracking,
let's
see
if
leave
on
or
okay,
so
yeah.
We
discussed
this
again
in
the
data
protection
going
group
meeting.
So
basically
it's
still
going
through
design
discussions.
A
Next
one,
the
new
rewrite,
only
red
ones,
access
mode.
Do
we
have
everything
merged,
Michelle,
Chris,.
A
A
A
C
I
think
this
one
Mauricio
is
not
really
going
to
have
time
to
continue
on
this.
Oh
okay,
I
think
it's
gonna
be
on
pause
unless
another
person
wants
to
take
over.
B
Okay,
there
was
issues
up
some
of
like
in
in
our
company
somebody.
It
is
that
with
the
privileged
containers,
it
doesn't
need,
like
some
kind
of
Windows
kernel,
binary
bits
like
it
cannot
like
privilege
container.
It
cannot
be
like.
Can
it
be
made
like
a
can
a
drivership
as
a
fully
open
source
module
right?
B
B
C
Know
Mauricio
did
some
prototyping
in
this
area.
I
was
not
I
did
not
think
the
container
image
was
unusual,
but
I.
B
A
Okay,
so
next
one
is
a
no
suspension,
Secret
I'm,
not
sure.
If
humble
is
here
I,
so
he
has
a
dark
PR
that
is
being
reviewed.
D
A
A
And
what
is
this?
Oh,
this
is
the
control
body
mode
version
in
source
and
Target.
So
this
this
one,
it's
just
the
the
talk
is
already
done.
That's
that
was
a
merged
earlier.
We
don't
need
the
new
docs,
so
the
only
remaining
is
blocked.
So
Run
Access
is
not
sure
it
has
not
much
update.
He
may
not
write
the
blog,
but
we'll
see
we
still
have
some
time
to
decide.
A
C
F
Yep,
so
all
of
the
code
is
merged
and
I
have
a
blog
post
to
do
and
then
update
the
docs.
F
Yeah
I
think
the
only
dock
update
is
switching
the
you
know.
It
says
it's
Alpha
and
if
we
seem
to
switch
that
to
beta
I,
don't
think
there's
any
changes
required
to
the
the
a
substance
of
the
docs.
Okay.
A
Thank
you
and
okay,
then
this
is
for
tracking
okay.
So
that's
all
we
have
here.
We
have
a
topic
here
for
design
review.
Sunny
view
address.
G
Yes,
can
you
hear
me?
Yes,
okay,
so
this
is
revision,
two
of
the
volume
provision
IO
cap-
and
there
are
some
changes
that
I
adopted
from
the
previous
feedback.
G
The
first
one
is
that
we
are
making
the
API
modified
volume
to
make
it
more
generic,
and
for
this
revision
it
will
only
provide
the
functionality
to
update
apps
and
throughput,
but
also
keep
in
mind
in
the
future
that,
if
there
are
other,
more
flexible
parameters
we
want
to
provide
depends
on
the
cloud
providers.
G
C
Cool
I
haven't
taken
a
look
at
your
updates.
G
C
It
would
be
good
to
go
a
little
bit
more
into
detail
about
like
what
Fields
you
would
use
and
sort
of
the
logic
to
to
count
the
quota,
especially
for
operations
that
are
in
progress,
because
I
think
that
when
we
did
the
volume
resize
function,
that
was
a
big
part
of
the
design.
For
that,
and
there
was
a
lot
of
trickiness
around
how
to
handle
like
someone
quickly
increasing
and
then
decreasing
their
requests,
and
you
know
how
to
make
sure
we
capture
any
in
progress
operations.
C
So
I
think
if
you
have,
if
you
can
provide
some
more
details
on
like
exactly
which
fields
are
going
to
track
and
progress
things
the
in
progress
allocations,
and
you
know
that
whole
sequence
of
how
we
serialize
that
I
think
that'll
be
useful.
A
And
this
will
be
depending
on
changes
right
on
the
the
status,
not
that
okay.
C
I
guess
the
other
thing
is
I
would
probably
I
would
I.
Guess,
let's
see
in
terms
of
next
steps,
because
I
want
I
think
we
should
open
up
the
CSI
spec
PR2
to
start
discussion
there,
because
I
have
a
feeling
we're
going
to
spend
a
fair
amount
of
time
debating
the
meaning
of
biops
and
throughput,
and
so
yeah
I
think
we
should
get
that
discussion
going
soon.
A
C
B
It's
it's
yeah,
that's
it,
but
it's
simpler,
also,
because
there's
no
no
introductions
required
with
the
the
node
side
so
that
helps
things
along,
but
like
the
quota
code
in
in
particularly
the
quota,
changes
has
to
be
kind
of
carefully
review
than
somebody
who
is
familiar
with
quota
stuff
has
to
also
review
like
I,
don't
know
direct
car
or
someone
has
time
to
do
it.
B
C
Okay,
yeah
so
I
guess
in
terms
of
next
steps,
I
think
when
I
say
it's
one:
let's
open
up
the
CSI
change,
so
we
can
start
discussing
and
then
second
thing
is,
if
you
can
add
more
details
about
how
the
quota
will
handle
shrink
and
things
like
that,
and
then
we
can
also
pull
in
someone
from
signal
to
to
help
look
at
that.
But.
C
G
Mode
I
that
should
be
a
valid
arguments.
C
G
Fail
I
was
thinking,
you
can
still
pass
down
the
iops
and
then
the
driver,
the
specific
driver,
should
decide.
C
Or
I
guess
it
I
mean
generally
in
the
past.
The
way
we've
done
we've
done
this
is
we've
made
a
capability
right
and
we
check
if
the
driver
supports
the
capability.
E
C
E
C
C
Tricky
part
would
be
like
modifications
like
say
if
somebody
modified
their
request
to
to
something
that
the
driver
doesn't
support.
I
guess
we
can
always
just
put
events
on
it.
F
But
isn't
this
what
we
would
use
the
extended
status
for
so
like?
If
you,
if
you
make
a
you,
know,
iops
request
of
18
gazillion,
you
would
get
a
status
saying
like
in
yeah
and
and
then
depending
on
the
driver
either
you
would
be
back
to
the
same
or
maybe
would
be
in
an
invalid
state,
but
that
would
be
reflected
in
the
status
yeah.
C
If
you
had
maybe
in
here,
if
you
can
put
examples
of
like
what
the
what
the
object
looks
like
in
those
various
error
scenarios,
I
think
that
will
help
perfect.
F
Yeah
and
I
I
I
guessed
this
I
wonder.
If
then,
we
want
in
the
API
between
the
sidecar
and
the
driver
to
have
something
like
after
you
do
one
of
these
operations.
The
driver
reports
back
with
what
the
current
iops
range
is,
so
that
if
there
is
some
error
case
at
least
we
can
report
what
exactly
the
volume
is
doing.
So
you
know
what
I
mean
like
there
has
to
be
like
a
two-way
communication.
It
can't
just
be
succeeded
or
or
failed
because
it
might
be
ambiguous.
What
state.
E
But
to
be
clear,
the
this
is
just
a
limit
right
like
if,
if
I
say,
if
I
say
50
000
iops
is
my
limit,
but
the
device
can
only
handle
50,
000
apps
total
and
it's
serving
a
bunch
of
other
users
like
it's.
It
will
still
say
sure
sure
your
limit
is
50
000
and
then
you'll
just
get
less
than
that
right,
because
the
device
can't
actually
go
that
fast
yeah.
E
F
E
F
Right,
but
there
is
so
like
I
mean
we're
thinking
of
of
the
of
some
specific
upcoming
features
where
users
will
pay
for
particular
provisioned
IO,
and
so
you
know
there
are
a
bunch
of
asterisks
and
stars
on
to
like
when
gcp
will
actually
be
able
to
fulfill
that,
but,
like
ultimately
like
the
user
is
specifying
a
number
that
is
implications
for
billing.
So
we
have
to
make
sure
that's
accurately
communicated
and
reflected.
E
F
F
I'm,
actually,
thinking
specifically
of
the
problem,
the
you
know
like
currently,
the
the
the
the
the
user
has
provision,
two,
a
thousand
you
know
things
and
so
like
they're,
actually
being
billed
for
two
thousand
and
then,
if
they
make
a
request
which
is
invalid,
like
what
exactly
are
there
be?
Are
they
being
billed
for
if
they
they
change?
This
thing
from
two
two
thousand
to
you
know
in
infinity
they're
going
to
be
built
for
Infinity
they're
gonna
be
built
for
something.
Is
it
two
thousand?
Is
it
the
maximum?
F
You
know
which
could
be
fifty
thousand,
like
all
that's
going
to
be
driver
dependent,
but
I
think
there
has
to
be
a
way
to
communicate
that
back,
so
that
the
user
knows
what
exactly
their
paying
paying
for,
or
you
know
how
much
their
bill
is
going
going
to
be.
F
B
Did
you
did
you
still
wanted
the
vendor?
This
thing,
like
feels
in
the
resources
like,
like
whatever
vendor.io,
slash,
something.
G
So
we
are
thinking
that
for
this
cap
we
just
want
to
provide
the
functionality
of
updating,
iops
and
throughput,
but
keep
that
vendor
specific
parameter
in
mind
and
not
blocking
implementation,
because
I'm
thinking
there
is
another
way
to
pass
in
the
vendor
parameters
like
even
in
the
same
modified
volume
request.
You
can
have
a
flexible,
flexible
map
where
the
parameters
just
like
what
we
do
in
storage
class.
B
G
C
Yeah
I
would
I
would
probably
say,
like
I,
don't
think
like
in
the
CSI
spec.
We
should
add.
You
know
like
a
opaque
set
of
values,
because
I
think
that
is
going
to
open
up
a
big
can
of
worms.
C
B
C
That's
an
interesting
question:
I
feel
like
that's
kind
of
like
a
resize
V2
kind
of
thing
like
if,
if
CSI
drivers
support
it,
we'll
pass
the
resizing
request
through
to
modify
volume
request,
but
I
think
that's
I
think
that'll
be
pretty
tricky
right
because
we'll
basically
have
to
have
like
two
different
resize
controllers.
G
And
I
would
also
say
that,
from
my
past
experience
of
like
using
databases,
the
word
you
call
like
resizing
a
database.
Sorry
I'm
just
talking
about
my
use
case.
That
doesn't
happen
so
often
to
some
point,
especially
if
you're
using
some
relief
non-relational
database
and
then
you
will
start
to
click.
Some
other
parameters
like
this
news,
for
example
in
the,
if
you
know
dynamodb
they
actually
provide
the
one
thing
for
you
to
configure,
is
RCU
and
wcu,
and
that's
just
changing
how
much
you
can
do
in
your
iops
and
throughput.
E
Well,
from
the
application
side,
I
think
that's
true,
but
from
the
storage
side
the
opposite
is
true
right,
most
storage
devices
have
a
fixed
amount
of
capacity
and
a
fixed
amount
of
iops
and
they
prefer
to
carve
it
up
and
fix
chunks.
So,
like
every
gigabyte
gets
three
iops.
So
if
you
have
a
terabyte,
you
get
3000
iops
and
like,
and
that's
not
fungible
right.
That's
how
storage
prefers
to
operate
because
the
realities
of
physics
say
you
know.
E
C
B
There's
another
like
for
an
AWS
case:
for
example,
user
increase
the
size
from
10
GB
to
100
GB,
and
then
they
are
done.
Then
they
realize,
though,
now
that
I
haven't
read
the
disk
size.
My
Ops
is
too
low,
but
the
quota
of
the
API
call
will
prevent
them
from
increasing
the
iops
for
another
6
to
10
minutes.
That's
a
bit
I!
Think
six
minutes
at
minimum.
C
B
G
So
I
think
right
now,
gcp
is
going
to
provide
the
update
API,
but
that
one
only
update,
iops
and
throughput
and
we
will
add
the
size
later
as
well.
So
you
can
do
that
all
in
one
API,
but
the
res.
We
also
have
a
resize
API,
that's
Standalone,
just
for
resizing
and
that
one
we
used
to
function.
B
E
E
What
you
may
actually
do
is
you
may
give
them
a
whole
terabyte,
even
though
they
don't
want
it,
because
that's
how
much
space
you
need
to
give
them
the
iops
they
want
and
and
and
so
like,
but
what
your
billing
could
do
is,
it
could
just
say,
look
I'm
going
to
bill
you
for
either
the
higher
of
the
capacity
you
want
or
the
iops
you
want.
I'll
do
the
math
and
figure
out.
E
You
know
what
the
higher
one
is
and
that's
what
you
could
build
for
and
then,
if
you
leave
space
on
the
table
or
you
leave
iops
on
the
table,
that's
your
problem
right
that
I'm
still
getting
my
money.
So
the
billing
system
could
hide
this
and
you
could
still
efficiently
carve
up
your
raw
storage
and
its
capacity
and
performance
gonna.
B
B
What
you're
asking
like
I
I
need
to
do
like
someone
either
me
or
somebody
has
to
do
some.
Some
research
on
this,
like
I
I,
would
have
initially
I
initially
thought
that
okay,
you
specify
iops
per
GB
during
creation,
and
then
you
then
automatically
you
increase
the
size.
Then
IOP
should
increase,
but
there's
iops
you
can
set
in
the
modify
volume
call
also-
and
there
has
been
bugs
reported
in
it,
so
I
wanted
to
ensure
that
we
are
not
proposing
something
that
will
get
the
user
stuck
like
they
resize
the
volume
for
size.
C
Yeah
I
guess
one
of
maybe
one
of
the
questions
that
we
should
answer
in
this
cap
is
for.
Do
we
want
to
should
the
iops
field
be
used
when
the
underlying
storage
system
kind
of?
Has
this
like
iops
per
GB
thing,
or
should
it
really
be
only
used
if
it's
truly
independent
parameters
that
you
can
modify.
F
C
No,
no
no
I
think
The
Proposal
is
to
allow
it
to
be
independent.
Okay,
because
I
think
for
for
a
lot
of
existing
storage
systems
like
the
iopsis
kind
of
implicit
and
people
control
it
by
the
gigabytes,
but
then
I
think
some
storage
systems
do
allow
you
to
actually
control
them
independently
and
I.
Think
that's
what
we're
trying
to
support
yeah.
A
C
Yeah
yeah,
so
I
think
we
need
to
kind
of
make
that
clear
in
the
the
spec
like
when
we
describe
what
it
means.
E
C
C
A
A
Thanks
so
moving
on
to
the
next
topic
and
see
if
I
think
he
yeah
I,
think
the
my
two
he
said
he
couldn't
join
the
meeting.
He
added
this,
so
he
was
in
one
of
the
Cozy
meetings.
He
was
talking
about
if
it's
possible
to
add
a
new
object:
storage
particle
for
dial,
UCS,
open
issue
here
so
right
now
we
support
three.
We
support
I!
Think,
okay,
it
looks
like
he
added
a
few
more
things
right
now.
A
So
I
thought
you
know
we
should.
You
know
bring
this
up
here.
A
D
I,
don't
have
any
objections,
I
think
that's
perfectly
reasonable.
Ask
I
think
the
only
question
really
would
be
when
you
introduce
a
protocol.
You
know
how
widely
used
or
is
it
given
the
effort,
but.
C
E
D
A
That's
a
good
question
yeah.
This
is
for
dell
ECS,
so
I
assume
you
know
they
will
be
maintain
their
own
part,
but
I'm
not
sure
what
how
do
we
Define
that
I
guess
so?
How.
A
C
I
think
in
the
Cozy
stuff
itself
there
isn't
anything
specific
to
the
protocol
if
the
protocol
is
mostly
used
as
like
a
handshake
between
the
user
and
the
driver
to
make
sure
that
they
speak
in
the
same
or
like
the
driver
supports
the
protocol
that
the
user
wants.
A
Right
so
this
is
basically
I
think
this
is
in
which
one
is
that,
like.
C
Cozy
controllers
actually
like
do
anything
specific
to
any
protocols.
It's
really
like
a
past
School
value.
A
A
Okay,
so
let's
say
if
they
want
to
add
one:
what
what
is
the
What
is
the
process
for
them
to
get
there?
If
they
want
one
want
to
do
this,
so
we
could
definitely
ask
them.
You
know
if,
if
they
want
to,
if
they
can
maintain
it,
they
probably
will
say
yes
because
you
know
since
now
they
you
know
they
are
racing
this
request
and
are
there
anything
else
we
want
them
to
do
to?
Or
will
this
be
like
going
through
a
some
review.
D
I
think
asking
for
a
long-term
commitment
is
a
reasonable
one
and
beyond
that
I
think
let
them
Drive
the
implementation,
and
you
know
cozy
folks-
can
help
review
it.
A
A
A
Yeah
I
was
I,
don't
remember
what
you
know
what
we
discussed
at
that
time
and
was
saying
we
have.
We
start
with
three
I,
don't
remember
what
we
said
if
we
want
to
add
more
or
what's
the
requirement
so,
okay,
so
probably
we'll
just
to
let
them
hold
the
pr
for
a
while
until
you
get
some
reviews
from
you
thought,
yep.
A
Okay,
thank
you,
okay,
so
that's
this
one
and
and
then
we
have
this.
The
seek
storage
and
you
report
out
so
yeah
add
a
link
here.
So
please
help
review
it.
B
Yes,
so
I
need
a
favorite
language
volunteered
to
work
for
Staples
at
expansion,
I
sent
an
email
she
took.
Some
notes
like
I
could
basically
reconstruct
this
note.
But
one
of
the
issues
like
with
the
stateful
set
expansion
is
it's
a
very
tricky
feature
and
when
people
open
enhancement,
they
just
you
know
like
they
just.
G
B
A
So
you're
asking
for
notes
from
Shalini
is
that
your
yeah
yeah
I
think
she
I
think
she's
out
for
maternity
leave.
I
have
not
seen
her
for
ages,
I,
don't
know
where
she's
back
I
can
ask,
but
I
does
not
look
like
she's
coming
back
soon,
but
I
can
definitely
ask
I'm,
not
sure.