►
From YouTube: CSI Volume Snapshots: On the Way to Faster and Better Backups - Adnan Abdulhussein & Nolan Brubaker
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
CSI Volume Snapshots: On the Way to Faster and Better Backups - Adnan Abdulhussein & Nolan Brubaker, VMware
Users need persistent data to support their business goals, and regular backups are critical to business continuity. The CSI Snapshot API provides users a consistent way of creating volume snapshots within Kubernetes, regardless of their storage provider. This is a big improvement over the prior, provider-specific methods for creating snapshots. This session will first cover the basic CSI Snapshot CRDs, as well as demo snapshotting and restoring a stateful application. Additionally, we'll share how disaster recovery tools, such as the open-source Velero, can build on CSI Snapshots, as well as future enhancements coming to CSI, and what migrations from current in-tree cloud provider implementations will look like. Agenda: - Overview of the CSI Snapshot CRDs - Demo of snapshot and restore workflow - Future direction for application snapshotting and quiescing in CSI
https://sched.co/UaXR
A
A
A
We
have
dynamic
provisioning
using
persistent
volume
claim
and
persistent
volume
api's,
and
we
can
do
so
across
multiple
different
storage
providers
using
the
storage
class
abstraction,
so
qnc
supports
multiple
different
storage
providers
and
you
can
see
in
this
table
from
the
documentation
some
of
some
of
these,
some
of
the
ones
that
are
supported.
You
see
the
usual
suspects,
so
abus
EBS,
Azure,
disc,
GC,
persistent
disk,
but
you
also
see
things
like
cinder,
SEF,
NFS
cluster
and
it
pull
works
and
a
few
other
storage
providers
as
well.
A
So
this
is
great.
We
have
a
way
to
support
multiple
different
storage
providers
and
dynamically
provision
volumes,
but
storage
doesn't
just
stop
at
provisioning
when
we
care
about
complex
stapler
applications,
we
also
care
about
the
ability
to
snapshot
volumes,
clone
volumes,
restore
resize
and
a
few
other
operations,
and
specifically
snapshot
and
toning,
and
restore
give
us
the
framework
to
perform
backup
and
disaster
recovery,
which
is
really
vital
for
running
production
systems
in
kubernetes.
A
Luckily,
we
have
a
great
ecosystem,
and
this
is
just
a
few
of
the
tools
that
are
used
for
backup
and
disaster
recovery
inside
of
Cuba
Nettie's.
Each
of
each
of
these
tools
have
different
strategies
for
actually
performing
backup
and
and
restore
in
capabilities.
Some
of
them
may
call
out
to
the
storage
provider
API
directly.
So,
for
example,
a
tool
could
call
out
to
the
AWS
API
to
take
a
snapshot
of
an
EBS
volume
and
then
on
restore
call
out
to
the
AWS
API
again
to
restore
that
snapshot.
A
Others
may
use
a
file
system
level
snapshot
tool
such
as
something
like
rustic,
but
this
has
performance
implications
for
large
volumes
and
in
order
to
interface
with
different
storage
providers,
some
of
these
plugins
have
plug-in
archetype.
Sorry,
some
of
these
tools
have
plug-in
architectures
that
allow
storage
providers
to
implement
a
connection
to
the
tool.
A
So
that
means
we
have
orchestrators
that
need
to
connect
storage
and
tools
that
need
to
connect
storage
and
for
each
of
these,
a
storage
provider
needs
to
implement
a
separate
plug-in
and
maintain
that
plug-in
and
keep
it
up
to
date
and
in
feature
parity.
That's
a
lot
of
work
for
storage
providers.
Wouldn't
it
be
nice
if
a
storage
provider
could
just
implement
one
interface
and
have
that
used
by
each
orchestration
system
and
each
tool.
A
This
is
where
the
container
storage
interface
comes
in.
So
it's
a
community
led
effort.
The
communities
of
kubernetes
mesosphere,
also
amazes
doc,
swarm
and
cloud
foundry
came
together
to
support
the
container
storage
interface
and
come
up
with
a
standard
specification
and
G
RPC
API
for
interfacing
with
storage.
A
A
Let
us
take
a
look
at
what
CSI
means,
specifically
for
kubernetes,
so
in
six
storage,
they're
trying
to
move
out
all
the
storage
providers
that
currently
entry
to
outside
the
core
code
base
bye-bye
queue
for
next
year.
A
big
a
big
push
for
this
is
going
to
be
embracing
CSI
as
the
primary
storage
integration
in
order
to
connect
to
CSI
the
the
Cuba.
Next
API
is
going
to
have
some
storage
api's
to
turn
to
act
with
D
CSI
drivers
and
we'll
take
a
look
at
those
in
a
second
I'll
possible
noted,
not
sure.
B
So
I'm
going
to
look
at
how
we
specifically
use
the
volume
snapshot,
portions
of
the
CSI
API
for
provisioning,
attaching
and
detaching-
that's
all
GAA
in
kubernetes,
but
for
volume
snapshotting,
that's
only
just
now
getting
to
beta
in
117,
and
it
was
alpha
from
112
to
116.
So
we're
gonna,
look
at
some
of
the
alpha
data
structures
and
workflows
for
some
of
my
diagrams
I
just
wanted
to
show
some
of
my
conventions.
B
Okay,
so
first
we're
gonna,
just
look
at
a
cluster
and
we're
gonna
zoom
in
really
closely
and
look
at
a
PVC,
a
persistent
volume
claim
pointing
to
a
persistent
volume
to
a
volume.
This
is
kind
of
your
basic
storage
scenario.
You
would
then
have
a
pod
pointing
to
that
PVC
so
that
you
can
actually
get
to
your
file,
glory
block
or
whatever
you're,
using
that
for
to
actually
take
a
snapshot
of
that
you
would
submit
this
volume
snapshot
custom
resource
definition.
This
is
namespaced.
B
B
B
B
It
calls
that
driver
and
that
provider
that
provider
driver
will
create
the
snapshot
42
and
while
that
snapshot
is
being
created,
the
controller
is
going
to
update
your
volume
snapshot
to
be
from
having
ready
to
use
nil,
to
ready,
to
use
false
and
then
is
going
to
get
the
snapshot
handle,
which
is
some
snapshot.
Id.
Whatever
scheme
your
storage
provider
uses
that
will
be
returned
back
to
the
CSI
snapshot
or
controller
and
that
will
be
put
into
a
volume
snapshot,
content
object.
B
Then
you
have
the
providers
driver
monitoring
that
snapshot
looking
to
see
for
whatever
state
the
provider
needs
to
judge
that
snapshot
is
ready
to
use
restoration.
Once
it's
ready,
it's
going
to
tell
the
CSI
snapshot
er
that
hey
you
can
use
this
and
then
that
will
get
passed
into
the
volume
snapshot
and
the
user
can
then
use
that
volume
snapshot
for
restoration
or
cloning.
B
B
You
also
have
a
reference
to
the
persistent
volume
that
you
created,
the
snapshot
from,
and
you
have
a
pointer
to
the
volume
snapshot
that
kicked
it
off
and
also
note
that
again,
this
is
the
alpha
version
of
the
data
structure.
The
definition
has
changed
a
little
bit
with
the
beta
versions
and
there's
also
a
status
for
this.
Now.
B
Okay,
so
once
your
volume
snapshot
is
ready
to
use
you'll
notice
down
here
at
the
bottom,
you've
got
a
creation
time
and
that
you've
got
ready
to
use
set
to
true
and
also
up
in
the
spec.
You
have
a
snapshot
content
name
pointing
to
the
ID
of
a
snapshot.
Content,
object.
Well,
I'm,
sorry,
a
volume
snapshot,
content
object,
and
this
is
diagram
of
what
your
cluster
is
going
to
look
like.
So
up
at
the
top.
B
You
have
the
normal
storage
you're,
probably
used
to,
and
at
the
bottom
in
the
bottom
left,
you
have
the
volume
snapshot,
which
is
referencing
the
original
PVC,
which
has
a
bi-directional
relationship
with
volume
snapshot
content,
which
is
then
tying
together,
everything
to
the
snapshot
and
persistent
volume
so
to
restore
or
make
a
copy.
Even
you
would
create
a
new
persistent
volume
claim
with
this
new
data
source
field,
which
is
now
beta
as
of
kubernetes
117
and
was
Alfre
from
112
to
116.
B
B
So
if
we
look
at
the
restore
process,
the
diagram
is
very
similar,
but
some
of
the
arrows
are
pointing
in
a
different
direction.
So
in
this
case
we
still
have
the
volume
snapshot
with
its
bi-directional
relationship
with
the
volume
snapshot
content,
but
we
now
have
a
PVC
pointing
to
the
volume
snapshot
instead
of
the
other
way
around.
So.
B
Now,
when
we
reconcile
state
on
restore
or
copying,
we
have
the
CSI
provision
or
controller
again,
a
common
controller
written
by
the
six
storage
community.
This
is
going
to
watch
for
a
persistent
volume
claim
that
has
a
data
source.
It's
going
to
create
the
PV
and
then
it's
going
to
ask
the
providers
driver
to
create
the
volume
the
providers
driver
will
then
give
the
volume
ID
to
the
CSI
provisioner.
B
So
I
can
update
the
PV
with
that
information
and
and
once
that
volume
ID
is
added
to
the
PV,
it
can
point
to
the
volume
and
then
we
end
up
with
a
state
again
that
looks
pretty
similar
to
what
we
looked
at
before.
But
in
this
case
we
have
a
a
persistent
volume
claim
that
actually
has
a
data
source
field
pointing
at
a
volume
snapshot
and
then
finally,
one
piece
of
data
that
I
didn't
go
over.
Is
the
volume
snapshot
class
custom
resource
definition?
This
is
very
similar
to
a
storage
class.
B
This
holds
fields
such
as
the
snapshot
er,
which
is
going
to
be
your
provider
driver
that
does
the
work.
It's
also
going
to
hold
the
parameters.
It's
a
key
value
pairs
to
give
to
your
driver,
and
this
will
vary
based
on
your
driver
or
plugin,
and
it
also
holds
the
deletion
policy
for
your
snapshots,
which
is
to
retain
the
snapshots
beyond
the
life
of
your
kubernetes
objects,
or
to
delete
your
snapshots
when
the
kubernetes
objects
are
deleted
and
both
the
volume
snapshot
and
the
volume
snapshot.
Content
objects
will
point
to
your
volume
snapshot
class.
B
Thank
you,
I
pre-recorded,
this
just
in
case
the
Wi-Fi
was
unreliable,
so
I
set
this
up
with
the
kind
of
standard
WordPress
and
my
sequel,
demo,
I
populated
the
site
with
a
post,
also
customize,
the
name
of
the
site.
We
have
a
comment
and
I
also
leave
a
comment
here.
It's
just
so
I
can
show.
We
have
data
in
here
and
we're
also
going
to
move
over
to
a
console
and
we're
going
to
take
a
look
at
the
namespace.
B
B
So
now
we're
gonna
we're
going
to
delete
the
namespace
now
something
I'm
gonna
mention
here
is
the
volume
snapshots
are
names
based,
so
those
are
going
to
go
away.
The
volume
snapshot,
content
object
is
not
namespaced,
so
we
can
still
get
to
our
data,
but
we're
going
to
need
to
recreate
the
volume
snapshot.
B
And
I
actually
have
the
code
for
this
example
up
on
github.
If
you
want
to
see
how
some
of
this
got
wired
together,
it's
not
this
particular
part
isn't
shown
in
the
demo.
But
if
we
look
inside
of
our
restore
contents,
you
can
see
here.
I
have
the
name
of
the
volume
snapshot.
Content
objects
are
already
populated
into
my
file.
B
And
one
more
thing
we
have
to
do
is
the
relationship
between
a
volume,
snapshot,
content,
object
and
a
volume
snapshot
is
bi-directional,
so
we
have
to
make
sure
that
we're
not
pointing
to
the
old
ones.
So
don't
do
this
in
production,
but
yeah,
but
we're
gonna
delete
the
new
IDs
of
those
very
specific
objects
so
that
we're
binding
to
the
names,
not
the
actual
object.
B
This
has
changed
a
little
bit
in
the
beta
versions
of
the
Sierra
DS,
so
that
you
don't
have
to
mess
with
the
spec
as
much
but
in
the
Alpha
versions.
This
is
one
thing
that
we
found
you
might
have
to
do.
If
you
run
into
an
issue
where
you
deleted
the
namespace
and
you
need
to
recreate
a
volume
snapshot
pointing
to
the
volume
snapshot,
content.
B
B
A
Awesome
thanks
Nathan,
so
great.
We
saw
that
we
can
use
the
volume
snapshot
api's
and
the
volume
snapshot.
Contents
API
is
to
successfully
restore
staple
applications.
So
what
does
the
future
of
CSI
look
like
well?
Clearly,
one
area
that
we
kind
of
see
growing,
especially
as
kubernetes
moves
to
moving
moving
current
storage
providers
out
of
tree
and
switching
migrating
to
CSI
is
more
CSI
drivers
with
natural
capabilities.
A
Additionally,
we
need
to
take
the
snapshot.
Csi
API.
Is
that
currently
there
and
bring
them
to
GA
so
they've,
just
we
just
released
the
beta,
but
we
need
to
keep
iterating
on
those
and
and
bringing
Tejay,
and
then
there's
also
Odyssey
great
ecosystem,
tooling.
That
again
can
build
on
top
of
those
api's,
for
example
like
the
Valero
CSI
plug-in
and
improve
the
workflow
in
certain
areas.
A
So
things
like
what
Nolan
pointed
out
there
about
having
to
copy
around
that
snapshot
handle
that's
something
that
we
can
improve
on
in
ecosystem
by
just
improving
that
workflow
other
things
that
are
going
on
in
the
in
the
community.
Around
CSI
and
volume
snap
shopping,
one
is
the
execution
hook
kept
an
API,
and
this
was
this.
Is
all
about
enabling
application,
quiescing
and
I'm
creating?
A
So
in
order
to
take
consistent
application,
consistent
snapshots,
we
need
a
way
to
run
a
command
in
containers
in
pods
in
order
to
do
quiesce
and
an
anchor
es
state
before
taking
a
snapshot.
Another
cap
that's
worth
mentioning
here
is
the
state
full
application
data
management
cap,
and
this
is
all
about
taking
application
level
snapshot.
So
we
kind
of
looked
at
the
vol
snapshot
and
growing
international
content
api's
and
these
kind
of
deal
with
volumes
specifically.
A
But
if
we're
interested
in
taking
a
full
application
backup,
we
want
to
have
a
higher
level
API
to
look
at
that
from
an
application
point
of
view
so
where
to
contribute.
If
you're
interested
in
this
SIG
storage
is
obviously
a
great
place
to
get
involved
in
since
we're
all
like
keep
going.
There's
a
six
storage,
intro
and
deep
dive-
that's
happening
tomorrow
at
2:25
p.m.
so.
Please
do
attend
that
if
you
can
otherwise
6
storage
meets
every
first
day
at
9
a.m.
A
Pacific
time
bi-weekly
and
there's
also
the
six
storage
slack
channel,
as
well
as
the
mailing
list
and
and
github
that
you
can
also
get
involved
in
and
then
CSI
also
has
its
own
slack
channel,
which
is
CSI
and
slack,
and
then
there's
also
the
github
repos
for
for
CSI,
specifically
as
well
as
nolan
mentioned.
The
demo
materials
that
we
use
today
are
all
on
github.
In
this
repository
so
and
again
the
slides
you
can
download
them
from
from
sked
con.
So
you
can
find
these
things.
C
A
D
B
So
one
thing
that
didn't
quite
make
the
cut
for
the
demo
is
there's
what
is
kind
of
called
in
the
in
the
CSI
world
and
import.
So
you
can
create
a
volume
snapshot
content
with
the
snapshot
handle
already
populated
already
pointing
to
an
existing
snapshot
in
your
claw
in
your
provider.
So
let's
say:
I
made
a
snapshot
in
my
web
UI
and
it's
already
there
and
I
want
to
use
it
or
say
I
made
one
with
Alero
and
I
want
to
and
want
to
move
to.
Csi
you
can
do
that.
B
E
A
F
So
one
question
regarding
the
application
who
KPI
for
acquiescing
the
application.
I
don't
know
if
it's
been
considered
all
the
ways.
I
would
say:
yes,
that
there
is
possibility
for
this
application
guessing
command
to
return.
Obviously
the
status
is
access
to
are
successful
whatever,
but
also
maybe
some
extra
metadata,
which
I
don't
know
it
could
be
or
output
or
something
could
be
very
interesting
to
label
the
some
data
about
the
disruption
itself.
The
clearing
process,
yeah.
A
H
So
for
consistency
group
we
are
actually
working
on
that
in
six
storage,
there
is
a
universal
Google
Talk,
where
we
are
designed
in
this
feature,
so
we
have
not
really
reached
the
consensus
yet,
but
you
are
her
welcome
to
join
the
discussion
and
there
was
also
another
thing
that
I
never
mentioned
earlier.
Is
the
application
application
snapshot
so
that
one
will
be
doing
this
application
level?
Consistency.
Is
that
answer
your
question.
Okay,.