►
From YouTube: SIG - Storage 2023-09-11
Description
Meeting Notes:
https://docs.google.com/document/d/1mqJMjzT1biCpImEvi76DCMZxv-DwxGYLiPRLcR6CWpE/edit#
A
All
right,
I
think
now
is
a
good
time
to
get
started
so
welcome
everyone
to
the
September
11th
2023,
Sig
storage,
meeting
I've
got
the
agenda
shared,
so
you
guys
can
see
that
and
our
first
topic
is
from
Arnon.
So
if
you're
ready,
why
don't
you
go
ahead
and
introduce
that.
B
B
According
to
our
storage
profiles,
to
those
who
don't
know
what
storage
buffers
is,
it's
basically
kind
of
a
table
which
is
most
of
it
is
the
is
build
by
us
a
long
long
time,
at
least,
which
is
the
key,
is
the
the
storage
classes,
storage,
provisioner
and
it
recommends
the
settings,
the
volume
mode
and
access
mode
combinations
and
also
the
recommended
the
Clone
strategy.
If
it's
snapshot,
snapshot
of
CSI
clones.
B
Will
give
us,
but
currently
it
was
supported
only
when
you
use
the
data
volume
now.
What
I
added
in
this
compact
PR
is
the
PVC
mutating
web
hook,
a
which
is
simply
calling
the
rendering
function,
which
was
used
by
our
data
volume
controller
for
doing
the
fulfilling
in
the
missing
Fields,
based
on
the
based
on
the
storage
profiles.
B
B
B
The
kubernetes
default
volume
mode,
which
is
always
for
system
so
in
this
PR,
in
order
to
test
that
it
works
all
the
PVCs
are
created
using
this
label,
so
we
cover
all
PVCs,
only
all
PVC
creation
flows
using
this
mutating
weird
hook.
B
B
Currently,
the
web
intercepts
only
explicitly
see
their
label
policies,
so
in
case
CDI
API
server
is
down.
There
is
no
effect
on
the
cluster
stability
and
no
other
PVCs
will
be
touched.
B
Okay,
now,
Michael,
English
Michael's
initial
command
was
that
at
minimum
we
need
to
have
a
feature
game
which
is
okay
on
the
deployment
and
the
labeling
of
the
PVC
I'd
like
to
hear
your
feedback
on
this
idea.
A
Okay,
thanks
for
the
intro
any
comments
from
anyone.
Have
you
seen
the
pr
anyone
want
to
bring
anything
up
here?
I
guess
the
one
I'll
start
with
a
question
so
for
the
idea
of
intercepting
only
the
explicitly
labeled
PVCs.
Is
it
required
for
those
PVCs
so
that
you
don't
have
to
check
if
it
was
actually
applied?
A
B
D
But
at
the
web
hook
it'd
be
on
the
kubernetes
level.
It's
required.
B
We
need
to
I
was
discussing
it
with
David
Russell.
You
could
see
the
discussion
there,
a
decoration
currently
at
least
with
the
current
setting
of
the
Web
book.
A
will
fail.
B
A
Sorry
I'm
fighting
the
the
zoom
bar.
That's
in
exactly
the
location
where
I
need
to
click
the
tab
this
one
or
was
it
a
different
link.
A
B
You
could
look
for
the
string
which,
which
appeared
in
PR
itself,
it's
citations
from
the.
B
It's
not
so.
A
Yeah,
so
yeah
so
like
I
would
expect
that
those
PVCs
would
fail
to
create.
But
if
you
had
a
PVC
that
did
not
carry
the
label,
it
would
create
successfully.
B
And
no,
if
it
will
be,
if
it
will,
if
it
does
not
have
the
label
and
missing
some
Fields
it
will,
they
will
fail,
of
course,
really
I'm
not
if
there
are
missing
fields,
of
course,
for.
C
A
D
I
think
my
point
is
I.
Think
I
want
this
I
think
it
would
be
best
if
this
is
like
opt
in
at
every
kind
of
level.
You
know,
as
in
you
know,
the
the
the
the
web
Hook
is
future
gated
and
each
PVC
has
to
be
explicitly
labeled
and
then
the
behavior
has
to
be
documented
as
well,
because
it
it
yeah
I
mean
I,
think
it
was
unsure
to
me
and
probably
others
that
would
be
using
this
feature.
D
How
exactly
it's
going
to
work
so
I'm
wondering
if
it
perhaps
makes
sense
to
to
have
a
design
proposal
around
this.
Okay.
E
D
Then,
just
also
a
plan
to
figure
like
how
this
fits
in
with
them,
like
yeah
I
mean
if
this
is
something
we're
going
to
be
publicizing
and
yeah
I
think
it
just
makes
more
sense
to
to
have
a
lot
of
communication
around
this
yeah.
E
E
A
So
I
mean
yeah,
so
I
think
the
one
like
just
to
kind
of
agree
with
Michael
too
we've
had
a
couple
of
false,
starts
on
trying
to
correct
some
of
the
unfortunate
behaviors
of
data
volumes
with
respect
to
backup
and
restore
and
Disaster
Recovery.
That's
been
discussed
in
previous
calls.
A
This
is
an
attempt
to
sort
of
work
around
that
by
not
needing
data
volumes
by
using
PVCs,
but
still
getting
the
like
the
Lion's
Share
of
the
benefit
of
the
data
volume,
but
we
just
need
to
make
sure
like,
as
we
try
to
solve
this
problem,
that
we
don't
just
add,
like
additional
complexity
or
additional
flows
without
actually
really
fully
solving
it,
because
that
just
diverts
our
energy
and
that
sort
of
stuff.
It's
a
really
cool
idea
and
a
potential
great
solution
to
the
problem.
B
B
D
B
B
D
I
think
it's
so
nice
that
we
should,
you
know,
really
have
communicate,
communicate
it
and
you
know,
be.
D
D
To
to
you
know
an
admin
or
or
someone
that
doesn't
know
too
much
I
just
be
like
what
you're
going
to
put
a
web
hook
on
PVCs
I,
don't
think
so.
But
if
it's
you
know
if
we
can
clearly
document
how
it
works
and
it
won't
screw
up
other
applications.
That
would
definitely
help
okay.
B
B
E
C
I,
just
added
some
some
of
my
thoughts
on
the
first
scenario.
We
should
make
sure,
but
as
I
see
it,
it
should
100
fail
like
it
wouldn't
make
sense
if
it
didn't
fail.
A
And
that
would
be
the
desired
Behavior.
So
that's
good,
great!
Okay!
Why
don't
we
move
on
to
the
next
topic.
B
A
We
I
just
wanted
to
take
a
step
back
from
this.
Do
we
understand
what
is
the
valid
use
case
of
having
multiple.
B
C
D
D
C
Concept
so
please
go
ahead,
so
there's
a
concept
with
volume
snapshot,
classes
of
a
default
volume
snapshot
class
just
like
there
is
with
default
storage
classes.
So
if
we
want
to
start
slow
with
implementing
this
feature,
then
maybe
we
could
just
respect
the
default
storage
class
like.
A
Honestly,
yeah,
that's
what
I
was
gonna
ask
like
there
should
be
a
default
and
I
think
we
should
treat
this
exactly
like
you
do
with
a
storage
class
and
so
in
your
data
volume
or
your
PVC
you're
expressing
a
storage
class
name.
If
you
want,
and
if
you
leave
it
out,
then
a
default
one
is
selected
based
on
the
semantics
of
the
cluster.
So
why.
A
A
A
A
I,
don't
sorry
I
think
we're
like
the
the
one
thing
I,
don't
think
that
the
storage
profile
is
the
right
place
because
who's
to
say
that
for
an
entire
storage
class
you
always
want
the
I
mean
if
the
use
case
for
multiple
volume
snapshot
classes
is
to
select
certain
options
like
retain
or
delete,
or
if
there's
like
a
grade
of
snapshot
or
something
like
that.
A
This
is
not
on
a
per
storage
class,
but
rather
on
a
per
snapshot
base
that
you
would
want
to
make
the
selection,
so
it
needs
to
be
in
the
actual
request,
not
in
the
storage
class.
If
you
ask
me.
B
B
B
A
It's
I
think
it's
actually
so
I
think
the
first
step
as
Alex
suggests
would
be
to
implement
the
def
using
the
default
if
it's
defined
and
then
the
second
would
be
to
I
mean
I,
think
it's
a
property
of
the
Clone
operation.
So
if
you
are,
if
you
look
at
the
spec
in
the
data
volume
for
the
like
the
source,
I
would
say
that
you
could
put
it
as
a
field
in
there
and
then
that
translates
into
the
populator
CRS
as
well.
A
Like
you
add
the
it's
a
weird,
it's
weird
to
put
volume
snapshot
class
in
one
section
and
then
storage
class
in
a
different
section,
but
I
think
that's
when
we
could
discuss
that
in
the
PR,
but
I
think
that's
where
you
would,
where
you
would
put
it,
because
it's
related
to
the
Clone
operation,
and
we
can
say
that
if
this
the
Clone
strategy,
that's
selected
is
not
a
snapshot,
then
that
field
is
ignored.
But
if
it
does
choose
a
snapshot,
then
it
would
use
that
class.
C
So
I
think
I've
been
thinking
about
the
first
bullet
point.
Is
it
a
real
use
case
for
a
few
minutes
and
I
I
guess
it
is
like
you
can
always
carve
like?
You
could
always
see
like
a
real
setup
where
it
has
tons
of
parameters
and
stuff,
but
but
maybe
we
would
like.
Maybe
we
want
to
take
it
slower
like
go
with
it,
see
a
cyclone,
maybe
in
those
cases
like
delegate
the
responsibility
somehow.
C
A
Also
I'm
confused
whether
the
snapshot,
the
default
volume
snapshot
classes
per
provisioner.
There
can
be
a
default
or,
if
there's
one
for
the
cluster,
which
would
be
would
break
us
in
some
we'd
have
to
understand
what
the
actual
low-level
semantics
are
for
that,
because
we
need
to
pick
one,
that's
appropriate
for
the
provisioner.
C
Right
yeah,
it's
per
cluster
I,
think
yeah.
A
So
that
doesn't
that
doesn't
match
if
there's
only
one
per
cluster
and
there's
multiple
provisioners,
you
could
have.
You
know
two
Snapshot
classes
for
what
provisioner
a
and
two
Snapshot
classes
for
provision
or
B,
but
you're
only
allowed
to
designate
one
of
those
four
volume
snapshot
classes
as
the
default.
Then,
if
you're
using
the
other
provisioner,
there's
no
default,
so
you're
still
broken.
B
A
Yeah-
and
it
should
be
noted
that
this
problem,
I
guess,
would
affect
like
keyword,
VM
snapshots
as
well.
A
And
that's
not
nice
because
you
could
have
you
could
have
multiple
VM
disks
on
different
provisioners,
so
you'd
have
to
be
able
to
individually
request
those.
A
So
I
mean
I,
guess
that
does
make
sense
why
this
that
does
go
as
an
argument
back
to
using
the
storage
profile,
because
it
would
be
simpler
that,
like,
if
you
have
a
weird
cluster
with
this
weird
configuration,
you
could
set
it
once,
but
I
also
hate
exposing
this
idea
to
admins,
who
would
have
another
place
to
set
things.
B
E
A
B
Link
between
the
two
zeros,
so
we
provided
the
simplest
way
they.
A
Yeah
I
mean
I
would
say
that
it's
potential
yeah
I
mean
it's
potentially
valid
to
have
that
updated.
But
then
are
we
just
okay
to
say
that
if
you
don't
specify
it,
then
a
random
one
will
be
selected
or
or
that
will
use
alphabetical
order
or
I
mean
it's
a
we
have
to
I
guess
we
have
to
define
the
behavior.
A
A
Any
other
thoughts
or
comments
on
how
ideas,
suggestions
on
a
path
forward.
A
I
guess
maybe
we
should
take
a
look
at
what
the
storage
profile
option
looks
like,
but
I
think
another
thing
that
would
be
I
mean.
This
is
my
suggestion.
Of
course,
just
one
of
many
potentials,
but
I
think
that
the
volume
snapshot
class
should
appear
in
the
status
of
the
storage
profile.
So
you
can
see
which
one
it
is
selecting
deterministically
by
default,
and
if
you
don't
like
it,
you
can
override
it.
Just
like
the
same
as
the
rest
of
it.
C
A
Really
and
then
yeah
if
I
think
that
could
maybe
be
the
simplest.
And
if
people
don't
like
that,
then
then
we
can
try
to
address
it
in
a
different
way
with
like
a
date,
outdated
volume,
level,
API
or
something.
A
Cool
all
right,
any
other
comments
on
that
topic
sounds
like
we've.
Maybe
got
a
potential
path
forward
or
okay
I'm,
going
to
open
up
the
CDI
issues
in
a
second
here,
I'm
just
cleaning
up
my
tabs.
E
If
you
look
at
that
one,
it
is
using
kubernetes,
121
I
think
with
CDI
157.
Okay,
so
I
gave
them
versions
to
use
for
this
particular
version
of
so
can
we
close
it
I
think
so?
Okay,.
A
A
Yeah
but
I
mean
like
in
the
list
here:
okay
I,
don't
know,
you're.
C
So
it's
a
shame.
We
don't
have
them
on
the
call
I
think
it's
the
get
them
they
all
forklift
I
see
the
link
below
okay.
A
C
E
C
Like
a
set
of
four
or
five
annotations
that
we
are
allowed
to
pass
between
the
DB
and
the
pods,
so
some
Network
ones
like
the
istio
one
and
I
think
they
can
just
capitalize
on
it.
I
linked
to
Alexander's
PR,
with
priority
classes.
B
E
E
E
C
And
we
might
want
to
open
the
gates
to
everything
potentially
like
make
it
user
Define,
which
annotations.
A
Yeah
I
like
the
idea
of
again
like
if
we
don't
need
an
API
or
if
we
don't
need
something
like
I,
think
I
think
it
makes
sense
to
have
it
restricted
down.
I
think
the
original
worry
was
that
you
could,
potentially,
since
since
the
the
labels
are
passed
to
objects
that
the
user
didn't
actually
create,
that
is
there
a
possible
escalation
of
any
kind,
so
we
wanted
to
or
unpredictable
Behavior.
So
we
were
trying
to
limit
the
surface
of
doing
this
to
things
that
we
feel
or
understand
the
behavior
for.
E
C
A
Do
that,
okay,
so
I,
guess
that's
all
the
issues
for
now.
Are
there
any
other
topics
from
anyone
else
who's
joined
today
any
like
open
floor
type
of
topics.
A
Going
once
going
twice
all
right:
well,
I
guess:
we've
got
a
wrap
here.
So
thanks
everyone
for
joining
and
we'll
see
you
at
the
next
one.