►
From YouTube: Kubernetes UG VMware 20220407
Description
April 7, 2022 meeting of the Kubernetes VMware User group with a discussion of reporting metrics associated with the vSphere CSI storage driver to Prometheus.
A
So
welcome
to
the
april
7
meeting
of
the
kubernetes
vmware
user
group
on
today's
agenda.
We've
got
a
number
of
items
and
if
we
get
through
all
those
items
as
usual,
we'll
keep
going
and
attempt
to
run
at
this
as
generic
q,
a
and
birds
of
a
feather
discussion
covering
things
related
to
best
practices
for
running
kubernetes
on
vmware
infrastructure.
A
A
A
Okay,
I
assume
everyone
can
see
that
agenda
document
and
hopefully
the
font
is
appropriate
so
that
it
might
actually
be
readable.
So
on
the
topic
of
this
vsphere
csi
storage
driver
actually,
two
months
ago
in
february
this
came
up,
so
there
was
hints
that
yeah
and
what
we're
talking
about
here
there
were
features
added.
The
primary
one,
I
think
is
exporting
metrics
to
prometheus
that
relate
to
success,
fail
and
performance
of
storage
operators
operations
when
you're
consuming
vsphere
storage
through
kubernetes
a
couple
months
ago.
A
The
stuff
might
have
been
in
there,
but
it
wasn't
documented
and
recently
meaning.
As
recently
as
the
last
month
or
even
the
last
two
weeks,
the
docs
got
published
and
a
great
blog
post
came
out
from
cormac
hogan,
a
member
of
the
group-
who
I
don't
think
is
here
today,
but
he
has
specific
instructions
to
walk
through
setting
this
up
and
by
setting
it
up,
I
mean
even
hosting
prometheus
and
grafana
to
put
together
a
complete
system
with
a
gui
dashboard.
A
I
tried
to
walk
through
this
myself,
but
I
just
started
it
yesterday
and
was
doing
it
part-time,
so
I
didn't
have
didn't
get
far
enough
along
that.
I
can
actually
demo
it,
but
cormac's
blog
post
has
screenshots
of
what
this
looks
like.
I
will
warn
you
like.
I
was
trying
to
follow,
along
with
cormax
and
depending
on
your
kubernetes
distribution,
whether
you
use
a
commercial
one,
you
know
an
open
source,
one
or
a
community
one
or
just
use
pure
upstream
code
roll
it
yourself.
A
Some
of
the
instructions
appear
to
be
related
to
a
particular
distribution.
So
when
it
came
to
his
instructions
on
identifying
where
some
of
these
csi
components
live,
he
called
up
using
vmware
hyphen
system,
hyphen
csi
as
the
name
space
and
on
mine
it
turned
out
it
was.
I
was
using
the
tonzo
community
edition
that
installs
the
csi
components
and
cube
system,
but
other
than
that
the
stuff
appeared
to
work.
As
I
went
along
miles
and
I
had
a
little
sidebar
chat
before
the
meeting
officially
opened
so
miles.
Here's
my
question
for
you.
A
You
know
I
was
curious,
since
I
had
run
an
installer
to
install
kubernetes
as
to
what
version
of
csi
actually
was
there.
You
know,
I
think
that
would
be
a
common
scenario
where
you
use
a
commercial
distribution
or
something
with
an
installer
and
you're,
not
really
sure
what
version
of
csi
you're
dealing
with,
and
I
was
curious
as
to
what
best
method
is
for
determining
that.
You
know
I
poked
around
with
a
google
search
and
said
to
say
maybe
there's
room
for
improvement
here,
because
it
didn't
seem
straightforward.
B
No,
you
know
you're
absolutely
right.
There
is
no
easy
way
to
tell
particularly
in
different
tanza
distros
of
kubernetes,
so,
for
example,
tkgs,
which
is
what
cormec
was
using.
I
believe
for
that
or
it
could
have
been
just
vanilla
with
a
straight
up
install
on
it,
because
it
was
not
a
pre-packaged
csi
driver.
So
when
it
comes
to
vsphere
with
tensor
tkg
service,
that
runs
a
private
proprietary
fork
of
the
csi
driver
right.
That's
why
they've
got
the
quota
mechanism
and
all
that
kind
of
stuff
in
there.
B
So
that
is
different
and
there's
no
sort
of
you
know.
Release
number
for
that.
So
whatever
you
get
in
the
10
to
kubernetes,
release
for
1.21
is
just
what
you
get.
You
can't
upgrade.
It's
not
supported
to
upgrade
it's
a
little
different
when
you
look
at
like
tkg,
multi-cloud
or
tkgm,
which
you
install
a
management
cluster
and
then
provision
your
other
clusters
from
it.
B
You
know
because
it
will
have
the
image
tag
in
the
in
the
the
tag
for
the
csi
driver
itself,
so
that
would
probably
be
the
other
way
when
it
comes
to
running
the
pure
upstream
stuff
like
if
you
have
a
vanilla
cluster
that
you
built
yourself
and
you
installed
csi
driver
yourself.
Those
images
are
tagged,
v,
2.1.
A
B
And
I'll
say
from
like
google's
perspective
like
if
you
install
google
anthos
gke
on-prem,
it
uses
a
csi
driver,
but
they
too
forked
the
vanilla,
csi
driver
and
they
do
their
own
builds,
which
have
their
own
different
tweaks
here
and
there
they
change
some
of
the
failover
properties
of
certain
pods
and
that
kind
of
thing.
So,
even
if
you
go
and
say,
look
at
red
hat
or
if
you
look
at
google
or
whatever
red
hat
are
still
used
in
upstream.
B
A
It's
I
perfectly
understand
that,
but
you
know
when
it
comes
to
the
csi
driver,
publishing,
release
notes
with
new
features.
I
think
it's
natural
for
a
user
to
expect
that
hey,
I
want
to
know
with
you
know.
My
vendor
chose
this
for
me,
but
I'm
curious
as
to
what
release
I'm
actually
on.
In
the
event
I
want
to
enable
you
know
prometheus
metrics,
and
I
wish
it
were
better.
One
thing
I
thought
I
might
work
was
to
just
go.
Look
at
the
logs.
A
I
had
this
theory
that
gee,
if
I
can
find
the
pods,
where
the
csi
components
live
and
go,
get
the
logs
I'll
bet
that
they
dumped
the
version
when
it
started
up,
but
it
didn't
appear.
No
it
doesn't
it
did.
That
might
be
a
nice
feature
to
add.
So
maybe
I'll
take
an
action
item
to
to
open
an
issue
to
suggest
that
because
then,
even
if
somebody
was
to
fork
it,
I
doubt
that
they'd
remove
the
you
know
the.
A
That
said,
you
know
this
is
a
recent
version
so,
depending
on
when
you
installed
your
kubernetes,
where
you
got
it
from,
you
may
or
may
not
have
these
metrics
available,
but
I
would
guess
that
they'd
be
coming
soon.
If
they're,
not
there
already.
The
metrics
that,
I
think
are
you
know
boiled
to
the
top
is
useful.
There
are
these
success
fail
things
I
think
it's
a
three-way
switch
that
comes
out
on
these
various
operations,
like
volume,
create
delete,
attach,
etc,
so
that
that
would
seem
something
useful
to
monitor
and
then
latency
of
these
operations.
A
You
know,
ideally
these
should
be
low,
but
if
you
go
look
at
cormac's
blog
he's
got
a
grafana
dashboard
that
shows
these,
so
that
strikes
me
as
something
really
useful
to
have,
and
I
think
there
are
a
few
other
things
in
there
as
well.
There
is
documentation
linked
here
on
collecting
these
metrics.
A
You
know
if
you
were
to
look
at
the
source
that
aren't
disclosed
in
the
documentation,
or
that
was
my
impression
as
I
was
poking
around
so
but
hey
some
documentation
is
better
than
none,
which
is
the
state
it
was
in
a
couple
months
ago,
so
right
for
whatever
it's
worth
this
might
be.
I
suspect
this
would
be
useful
to
very
many
people
running
this
in
a
production
scenario.
It's
almost
you
know
if
the
metrics
are
there
and
useful,
it's
almost
crazy,
not
to
use
them.
In
my
mind,
could
save
you
yeah.
B
They're,
some
of
them
are
gauges.
Some
of
them
are
like
histogram
type
things,
so
you'll
get
like
op
types,
so
they'll
be
like
what
part
of
the
crud
operation
was
it
or
you
know?
Is
it
a
volume?
Expansion
or
is
this
a
query?
You
know
snapshot
operation,
so
it'll
give
you
the
operation
whether
it
passed
or
failed.
That
kind
of
stuff,
the
latency
stuff,
for
example,
that
steve
was
talking
about,
is
operation
latency,
not
storage,
latency
critically,
because
they
might
look
the
same
sometimes,
but
some
operations,
you
know,
take
50
milliseconds
of
process.
B
A
Another
thing
I
want
to
point
out
this,
this
picture
that
I
scrolled
up
to
shows
kind
of
the
basic
architecture,
because
some
might
not
even
be
familiar
with
the
terms
used
by
prometheus,
but
service
monitors
are
the
things
that
are
put
in
place
to
gather
this
information
and
they
feed
into
a
prometheus
server.
Then
you
might
put
something
on
top
of
that,
like
rafana
to
actually
render
these
metrics
in
a
a
gui
so
that
you
can
get
the
dashboard
experience.
A
One
thing
that
was
pointed
out
by
members
of
this
group
back
when
we
covered
this
two
months
ago.
I
think
it
was
miles
robert
and
scott,
who
I'm
not
sure
has
joined
us
today
pointed
out
explicitly
that
in
some
distros
there
might
be
a
package
of
prometheus
but
there's
a
particular
version
with
an
operator
that
is
much
easier
to
deal
with
than
using
some
of
the
things
like
the
packaging
built
into
tonzo
distributions.
A
B
So
if
you
install
tkg
extensions
or
kinds
of
extensions,
or
whatever
they're
called
today
on
top
of
your
tensor
cluster,
so
that's
like
a
bunch
of
out
of
the
box
bits
and
pieces
to
help
you.
You
know,
get
a
baseline
for
running
your
case
cluster.
It
uses
k
app
to
install
it
all
and
that's
fine.
It
gets
you
what
you
need.
However,
it
doesn't
include
the
prometheus
operator
and
the
prometheus
operator
is
what
gives
you
access
to
the
service
monitor
crd
type,
which
you
absolutely
want
to
be
using
to
discover
all
your
metrics.
B
So
if
you
are
using
the
tanzo
extensions,
I
and
you
want
to
use
service
monitors,
which
you
really
should
be
using
in
prod
for
for
prometheus.
I
would
advise
not
to
use
the
prometheus
and
grafana
that
comes
with
tkg
extensions
and
instead
use
the
helm
chart,
because
it
rolls
all
that
stuff
out.
There's
one
called
q.
Prometheus
stack,
I
think,
is
the
name
of
the
helm.
Chart
I'll
drop
a
link
into
the
chat
anyway,
and
that's
the
one
that
I
would
advise.
You
guys
take
a
look
at.
A
A
Take
the
easy
path,
this
this
operator
method
is
maybe
going
to
save
you
some
time
and
headaches.
B
I
dropped
the
chart
into
the
chat
there,
so
that
is
the
one
I
encourage
you
to
have
a
look
at
the
critical
piece
that
it
contains
is
the
prometheus
operator.
You
could
roll
out
the
prometheus
operator
yourself
to
your
tkg
cluster,
along
with
tkg
extensions
and
probably
make
it
work
to
me.
It
feels
like
a
bit
too
much
work
to
do
something
like
that,
and
this
is
always
known,
tested
released
as
a
package,
so
I
would
say
seriously
look
into
cube
prometheus
stack.
A
Okay,
let
me
cut
and
paste
that
right
into
the
note
stock-
well
I'll,
do
it
after
the
meaning,
since
I'm
sharing
the
screen:
okay,
here's
an
advisory
related
to
csi
and
the
upcoming
1.24
release.
A
Now,
just
this
morning
I
saw
an
announcement
that
maybe
this
kubernetes
1.24
release
is
going
to
be
delayed,
so
you
know
that
they
had
plans
for
a
release
candidate
coming
out
very
soon,
like
you
know,
you
know
this
month
in
a
week
or
so,
but
they
announced
that
release
candidate
might
be
pushed
back
a
little
bit
and
then
one
two
four
obviously
is
likely
to
get
delayed
too.
If
the
release
candidate
is
late,
but
in
any
event,
if
you
do
upgrade,
there
is
an
advisory
on
csi.
A
This
isn't
just
vsphere
csi,
but
I
believe
it's
related
to
csi
in
general
of
needing
to
drain
out
nodes
in
an
upgrade
scenario,
and
I
put
the
link
to
the
pr
here
that
has
more
details
on
that.
So
you
know
if
you
live
on
the
bleeding
edge
and
jump
on
kubernetes
releases
as
soon
as
they
come
out.
This
is
something
you
might
want
to
keep
in
mind
to
not
get
ugly
surprises.
A
Then
I'll
turn
this
stage
over
to
you
miles
that
the
csi
now
supports
csi
snapshot,
natively.
B
Yeah,
so
this
is
something
that
people
have
asked
for
for
forever.
Since
we
launched
the
csi
driver
was
csi
snapshot
support.
I
guess
just
some
points
of
clarity.
I
mean
the
high
level
thing.
Is
we
now
support
the
csi
snapshot
feature
right?
It's
a
boilerplate
feature.
That's
part
of
csi
spec,
nothing
special
there.
It
does
not
back
up
your
applications,
it
does
not
guarantee
application
consistency.
B
It
does
not
do
any
of
that
magic
stuff.
You
still
need
to
quiesce
your
app.
You
still
need
to
make
sure
that
it's
consistent
before
you
take
your
snapshots.
These
are
not
backups,
but
that
functionality
is
there.
If
you
need
it,
so
it
does
a
snapshot
at
the
vsphere
layer.
Then
it
takes
a
consistent
copy
of
the
data
and
then
it
releases
the
snapshot
of
vsphere
layer.
So
you
have
a
consistent
copy
of
your
data
doesn't
mean
that
the
application
was
shut
down
and
quiet
properly,
but
you
will
have
a
copy.
A
And
it
could
well
be
that
whatever
backup
technology
you
use
is
prepared
to
integrate
with
this
functionality
in
a
csi
driver,
but
yeah
that's
outside
the
scope
of
what
this
group
gets
into.
But
you
know
there
are
solutions
like
valero,
along
with
rustic,
that
might
do
some
of
these
steps,
and
then
there
are
packaged
versions
of
commercial
backup
products
too.
That
may
or
may
not
support
this.
Yet
since
it's
new,
I
think
that
it's
possible
that
some
solutions
aren't
prepared
to
hook
up
to
this
yet.
B
So,
as
far
as
I
understand,
because
this
is
just
the
standard-
csi
snapshot-
part
of
the
csi
spec,
any
backup
app
that
you
have
today
take
it
valero
or
castin
or
whatever
it
is
that
you
use
to
do
your
cloud
native
backups
if
they
support
csi
snapshot
and
that
part
of
the
spec,
it
should
just
work
with
this.
You
know
it's.
There
should
be
no
proprietary
integration
needed
to
to
support
this
feature.
E
B
E
E
To
say
we
put
a
ui
around
valero,
but
yeah.
B
Yeah
so
we've
talked
about
this
at
length
in
in
other
meetings
and
other
forums
about,
should
you
do
stretch
cluster
with
kubernetes,
and
my
personal
opinion
is
absolutely
not,
though
we
do
have
a
lot
of
customers
that
keep
asking
if
asking
us
for
it
anyway.
So
these.
B
That's
that's
a
good
starting
point.
Let's
do
that
so
a
stretch
cluster
is
whenever
you
have
two
compute
clusters
on
different
sites
or
different
buildings
or
whatever,
and
you
have
a
single
storage
array
or
what
appears
as
a
single
storage
array.
It's
you
know
a
single
logical
storage
array
but
backed
by
two
physical
devices
and
they
synchronously
replicate
to
each
other.
That
means
every
I
o
operation
is
completely
blocking.
B
So
if
there's
a
vm
on
site
a
and
it
writes
to
the
data
store
on
site
a
it-
does
not
get
an
acknowledgement
back
that
okay,
that
was
written
until
it
goes
to
the
other
side,
gets
an
ack
that
comes
back
and
then
they
both
act
to
say
that
was
a
successful
transaction.
So
that's
synchronous,
replication
and
a
stretch.
Cluster
essentially
does
that,
but
across
multiple
sites.
B
So
it
was
a
a
way
for
years
and
years
and
years
to
protect
applications
that
had
no
ability
to
shard
or
spread
or
replicate
data
themselves,
so
think
about
your
microsoft,
sql
server,
that's
a
bad
example!
It
does
have
that,
let's
say
mysql
as
an
example,
or
you
know
any
of
those
more
traditional
type
databases
they
don't
have
built-in
replication
or
sharding,
or
the
ability
to
scale
out
or
anything
like
that.
B
So
you
would
protect
them
with
a
stretch
cluster,
so
if
your
entire
site
dies
no
problem,
all
your
data
lives
on
the
second
site
and
your
workloads
come
up.
This
poses
a
lot
of
problems
for
kubernetes,
because
kubernetes
was
designed
to
not
come
back
from
those
kinds
of
failures.
It
was
essentially
designed
that
if
something
dies
it
stays
dead.
You
know
it
isn't
recovered
somewhere
else,
and
that
does
cause
some
challenges
and
particularly
from
like
a
failure
domain
architecture,
point
of
view.
It
is
very,
very
challenging
to
try
and
align
kubernetes
failure.
B
B
You
know
sort
of
separation,
logic
inside
kubernetes,
be
it
anti-affinity
policies
or
whatever
now
no
longer
applies,
because
the
underlying
topology
has
changed.
Based
on
what
kubernetes
thought
it
was.
So
there
are
reasons
why
some
companies
want
to
use
stretch
cluster
that
is
essentially,
this
is
the
way
we've
always
done
things
or
the
only
cluster
we
have
is
a
stretch
cluster.
Please
tell
us,
we
can
just
run
it
there
and
we'll
pin
it
to
one
site
or
something
like
that,
but
it
is
not
an
architecture
that
you
should
go
into
thinking.
B
This
is
the
best
way
to
run
kubernetes.
This
is
the
best
way
to
make
everything
redundant
because
it
will
just
cause
you
headaches
and
a
lot
of
pain
going
down
the
road
that
said,
the
csi
driver
now
supports
it,
and
this
is
essentially
just
a
sign
off
that
says.
Yes,
we
will
support
you
if
you
run
it
on
a
stretch
cluster,
but
we're
not
recommending
that
you
run
it
on
a
stretch
cluster.
B
Yeah,
it's
it's.
It
works
the
same
way.
It's
always
worked,
they
did
add
some
extra
test
cases
and
I
believe,
there's
like
one
or
two
edge
cases
that
they
caught.
So
there
might
have
been
a
few
really
niche
things
that
they
fixed,
but
the
larger
work
was
validating
it
figuring
out
how
we
support
it
from
a
gss
perspective,
figuring
out
how
engineering
supports
it,
it
was
mainly
validation,
yeah.
A
And
I
suppose
you
know
the
the
one
thing
to
consider
then,
since
it's
not
recommended,
but
if
you
have
kind
of
these
legacy
style
databases
running
your
option
would
probably
be
to
just
run
them
outside
of
kubernetes
entirely,
and
they
should
work
perfectly
fine.
The
way
they
have
for
a
decade
or
more
and
you'd
be
able
to
take
advantage
of
this
stretch,
cluster
that
you
already
invested
in
and
get
the
high
availability
that
you
invested
in.
B
Precisely
yeah,
don't
don't
add
a
layer
of
complexity
that
you
don't
need.
If
you
don't
need
to
put
your
existing
database
on
kubernetes,
don't
do
it
if
it
runs
fine
in
vms
and
you've
got
it
protected
by
a
stretch
cluster
and
it
works.
Leave
it
alone,
build
your
new
stuff
in
case
and
leave
the
old
stuff
where
it
is
until
you
decide
we're
going
to
re-platform
into
a
completely
different
database
and
it's
going
to
be
a
real
cloud
native
database.
Sorry,
my
blog
just
saw
someone
outside.
A
Okay,
thanks
miles
for
covering
that
another
update,
the
a
number
of
cloud
provider,
new
versions
came
out
in
the
last
couple
of
weeks.
You
know
the
latest
version,
the
1.2.6,
but
they
backfilled
some
bug
fixes,
and
I
don't
know-
maybe
even
some
feature
enhancements
on
some
of
the
older
versions
as
well.
A
You
know
we
we
support
those,
because
sometimes
distributions
or
users
are
running
on
something
less
than
the
very
latest
release
so
that
you
know
it.
It
can
be
important
to
support
users
who
are
on
things
going
back
a
few
versions.
A
I
don't
know
I'm
not
going
to
read
you
the
list
here.
One
thing
is
the
helm:
chart
moved
so
a
helm
chart
is
the
recommended
way
for
installing
the
vsphere
cloud
provider
if
you're
running
the
route
of
pure
open
source
or,
if
you're,
in
the
position
where
you
actually
publish
a
distribution
or
installer
of
kubernetes
yourself.
D
I
got
a
question
on
that
so
like
we're
in
the
case
that
you
mentioned
where
we
still
have
older
versions
of
other
things
that
have
blocked
us
from
migrating
to
the
csi.
One
of
the
things
we
found
was
that
the
vsphere
api
can
cause
basically
the
cubelet
to
not
fully
start
up.
D
A
A
My
tongue,
I'm
not
all
that
familiar
that
they
linked,
I
think
two
or
three
issues
or
bug
reports
that
have
been
fixed,
so
the
details
should
be
in
there
if
you've
been
having
this
issue
yourself.
Hopefully
it
aligns
with
one
of
these
open
issues.
Otherwise
you
should
open
a
new
one.
D
Yeah
we
have
an
issue
open
internally
with
vmware
our
vmware
support
on
the
vsphere
api,
basically
needing
to
get
restarted
anyways.
But
when
that
was
happening,
we
were
experiencing
impact
on
the
clusters.
If
cubelet
restarted
during
that
time,
because
it
would
try
to
connect
the
vsphere
api,
it
wasn't
available
and
it
would
get
stuck
in
a
state
where
it
can
finish
starting
up.
D
B
As
far
as
I'm
aware,
if
you're
on
cpi
and
david's
on
the
call-
so
maybe
he
can
confirm
here
as
well,
but
if
you're
using
cpi,
it's
not
integrally
part
of
cubelet
anymore,
so
cube
should
be
able
to
start
up
without
cpi
establishing
a
connection
to
the
vsphere
api.
Because
it's
an
external
component.
E
That's
true,
I
have
a
cluster
that
was
up
when
my
visa,
like
that
nodes,
crashed
and
came
back
up
and
my
vcenter
was
not
accessible.
It's
you
know
not
100
necessarily
functional
and
cubelet
is
up
and
everything
is
running.
You've
got
other
issues
around
provider,
node
refs
and
things
like
that.
You
know,
but
the
cluster
will
come
back
up.
The
node
cubelet
is
not
it's
not
in
the
data
path
anymore.
The
access
to
the
cloud.
B
D
Was
a
little
mysterious
because
cubelet
would
be,
it
would
look
like
it
was
up
and
it
would
start
reporting
to
the
api
server
and
then
it
would
just
stop
it
was.
It
was
like
it
partially
got
up
and
then
it
hit
that
that
spot
where
it
was
trying
to
make
that
connection
and
then
just
stopped,
but
it
still
kind
of
looked
like
it
was
up.
Just
wasn't
ever
communicating
back
to
the
the
kubernetes
api
server.
After
that
point,.
B
Yeah
I
mean
the
the
the
entry
pi
provider.
We
know
we
know
has
a
lot
of
issues
right.
It
was
a
prototype
thing
that
we
put
out
years
and
years
ago,
and
that's
that's
why
we
built
cpi
and
csi
was
to
get
to
get
away
from
a
lot
of
those
inherited
problems
that
that
the
vsphere
cloud
provider
had,
but
I
would
say
a
lot
of
those
kind
of
like
weird
niche,
especially
there's
stuff,
to
do
like
disc
failures
and
failures.
A
Yeah
and
the
kubernetes
project
itself,
they
keep
kicking.
This
can
down
the
road,
but
they
keep.
You
know
for
two
years
now
it's
been
one
year
from
now
we're
going
to
kick
the
entries
out
of
the
tree
and
that
you,
you
will
be
forced
to
migrate.
The
reality
is
they,
I,
I
think,
are
maybe
exaggerating
the
speed
at
which
that's
going
to
happen
to
encourage
people
to
actually
you
know.
A
D
Yeah
yeah,
so
you've
got
some
hardware.
That's
can't
get
moved
to
that.
So
we've
we're
looking
at
I'm
like
okay.
D
Are
we
going
to
separate
those
off
and
do
something
different
with
those
because
of
what
the
hardware
is,
but
anyway,
so
we're
getting
close
to
be
able
to
start
testing
that
we
could
move
some
of
our
environment,
but
then
it
gets
confusing
when
you're
like
okay,
half
our
environments
like
this
half,
it's
like
that,
so
it's
almost
easier
to
just
wait
till
you
can
get
everything,
but
when
it
these
versions,
that's
called
out
in
that
note-
is
that
the
kubernetes
version
version.
A
These
align
with
the
version
of
kubernetes
that
that
cloud
provider
supports,
but
I
believe
they
will
support
older
ones
than
that
and
not
newer
ones.
So
there
is
some
correlation
where
these
semantic
versions
do
track,
but
it
isn't
necessarily
a
one
for
one.
So.
D
I
guess
my
my
question
is:
if
I'm
talking
about
just
the
entry
provider,
I'm
getting
those
updates
like
that.
It
would
say
here
if
I'm
getting
to
that
version
of
kubernetes.
E
Okay,
yeah
the
cpi
in
general,
is
the
major
and
minor
version
has
to
match
with
the
kubernetes
version.
So
any
120
needs
to
be
with
a
120
cpi,
but
it
could
be
any
120
will
work
with
any
120
of
the
cpi
is
kind
of
the
general
contract
that
is
made
beyond
that.
121
same
thing,
122
the
same
thing,
but
you
can't
use
a
121
with
120..
E
There
is
a
contract
about
that.
Just
in
terms
of
breaking
changes
doesn't
mean
it
won't
work,
but
with
the
changes
of
go
versions
and
things
being
vendored
in
and
different
kubernetes
things
being
vendored
in
it's.
The
supported
way
is
to
be
in
the
same
release,
same
minor.
D
So
that
next
bolt
or
the
last
bullet
point
you
see
there,
so
it
says
vsphere,
csi
and
now
has
no
dependency
on
cpi.
A
I'm
not
sure
what
the
scenario
would
be
that
that
would
take
place,
but
apparently
there
is
one
because
they've
called
it
out.
B
Yeah,
it's
that's
an
interesting
one.
It's
I'll
be
honest
in
internal
engineering
decision,
the
engineering
team
behind
csi
wanted
to
essentially
allow
the
csi
to
work
without
any
dependency
on
the
cpi,
and
the
only
thing
that
it's
ever
depended
on
cpi
for
was
to
populate
the
node
uuid,
which
csi
has
now
moved
away
from.
They
use
a
different
method
instead
of
node
uuid,
to
actually
do
storage
alignment
with
the
node
that
it's
in.
So
there
is
no
tight
coupling
between
the
two
anymore.
B
B
You
don't
need
the
vsphere
cpi
now,
as
I
understand
it
from
a
kate's
perspective,
you
either
need
to
choose
an
in
cloud
or
entry
provider
as
part
of
cubelet
bootstrap
or
choose
external.
B
E
Although
it
does
have
load
balancing
capabilities
now
with
nsx,
it
can
do
nsx
load
balancing
for
you.
If
you
have
nsxt
yeah.
B
A
A
I
think
dan
has
had
some
other,
perhaps
more
recent
talks
on
it
in
other
venues.
I
can't
remember
if
it
was
a
kubecon
or
a
cloud
rejects
or
something
like
that,
but
yeah.
I
I'm
really
a
fan
of
it
myself
and
run
it
in
my
home
lab
it.
I
don't
know
that
it
would
hold
up
in
an
enterprise,
large
scale,
enterprise
scenario
as
a
load
balancer,
but
certainly
for
home
lab
and
maybe
for
edge.
C
B
So
that's
a
good
question
david,
so
it
it
has.
It's
got
its
own
implementations
of
zones
and
regions
as
well.
As
far
as
I
understand,
it's
never
actually
used
the
cpi
for
zones
and
regions
and
they've
always
done
their
own
zones
and
regions.
I
think
it
ignored
the
stuff
that
the
cpi
put
in
there.
I
might
be
wrong,
but
that's
just
what
I
recall
from
the
last
time
I
I
talked
to
engineering
about
it.
C
Yeah
so
then,
well,
if
someone
like
kubernetes,
like
you
know,
starts
to
do
like
you
know,
node
like
zone
placement
and
stuff
like
that,
does
it
take
the
you
know,
take
the
side
of
like
I'm
gonna,
look
at
the
csi
stuff
or
the
seat,
or
you
know
if
cpi
is
up
and
running,
you
know
which
one
is
the
one
that
kind
of
wins
out,
because
there's
only
you
know,
there's
only
one
well,
two
labels
right
that
actually
determine
where
that
happens.
So
if
you
have
both
running,
will
they
conflict
or.
B
E
Don't
conflict
with
one:
they
don't
conflict
with
one
another.
One
is
for
pod
scheduling
and
the
other
is
for
persistent
volume,
scheduling,
they're
used
for
different
things.
The
csi's
will
not
deal
with
pod
scheduling
onto
nodes.
The
csi
will
purely
do
zone
topologies
for
where
to
create
the
persistent
volume.
E
That's
it's
purely
zoning
for
persistent
storage,
not
for
the
pods.
So
if
you
want
to
have
full
availability
zone,
topology
aware
you're
going
to
need
the
cpi
in
order
to
be
able
to
get
topology
awareness
in
pod,
scheduling
on
to
nodes.
C
Yeah,
it
seems
a
little
it's
interesting
because,
like
if
you
use
there's
like
a
whole
other
mechanism
to
do
like
volume
scheduling
somewhere,
it
would
make
sense
that
you
would
want
the
pod
to
land
in
a
place
where
it
you
could
actually
access
that
volume
yeah.
That's
why
I
brought
up
the
question.
I
was
just
interested
in.
E
The
only
way
of
really
solving
that
without
the
cpi
is,
I
mean
you
can
set
in
volumes.
If
you
have
the
provision
before
schedule-
or
I
forgot
the
names
but
there's
provision
the
disk
before
creating
the
pod
or
schedule
the
pod
and
then
create
it,
there's
different
settings.
You
can
do
with
that,
so
you
could
probably
play
with
the
ordering
yeah
so
wait
for
first
consumer
for
first
consumer-
and
you
know
you
could
probably
play
with
that
and
maybe
get
away
without
the
cpi,
but
it
seems
pretty.
E
I
think
the
only
case
that
I
see
that
is
that
in
the
docs
you
mentioned
there
there's
something
here
about
for
the
adventurous
who
want
to
try
kubernetes
on
a
vm
in
arm-based
things.
If
you
try
things
on
like
esxi
on
arm
or
on
laptops
running
these
here,
you
know,
whatever
you
may
be
a
bit
more
constrained
in
resources
and
the
cpi
does
take
quite
a
lot
of
resources
relatively,
so
it's
just
getting
rid
of
another
pod
from
there.
So
I
think
for
edge
cases.
A
Yeah,
I
think
this
whole
topic
of
zone
awareness
and
reaction.
You
know
where
potentially
you've
got
infrastructure
failure,
domains
as
a
lower
layer
to
kubernetes
failure,
domains
and
we've
had
chats
about
this
before
it's
by
necessity,
complex
if
you're
not
going
to
be
opinionated
and
force
people
to
particular
choices
which
the
vmware
infrastructure
tends
not
to
do.
A
You
know
you
could
elect
to
have
a
data
center
that
has
redundancy
by
rack
by
aisle
by
region,
and
it's
all
out
there
if
you
want
to
do
it,
but
it
makes
it
hard
to
put
a
can
solution
that
just
works
out
of
the
box
without
configuration
settings.
A
Anyway,
let's
let's
move
on,
unless
somebody
has.
D
I
have
one
more
question
on
the
cs5
stuff.
I
read
somewhere
that
there
was
a
minimum
hardware
version
of
15.
B
It's
for
all
the
nodes
that
are
going
to
use
csi
so
any
case
node
has
a
minimum
requirement.
Minimum
requirement
of
v15
I've
been
asked
this
a
lot
in
the
past.
You
know:
can
it
be
used
with
a
lower
hardware
version
and
people
say
hey?
I
got
it
to
run
with
hardware
vm
hardware
13..
You
can
get
it
to
work
with
13,
but
there
are
some
nasty
bugs
and
edge
cases
that
are
resolved
in
v15,
particularly
around
failure,
failures
and
disk
remounting
that
you
do
not
want
to
get
involved
with.
B
D
D
B
Right
but
there,
and
because
that
was
the
first
release,
we
didn't
realize
that
there
are
these
bugs
at
the
time,
in
the
first
release
that
we
discovered
after
the
fact,
which
is
why
we
retrospectively
said
v15
is
where
you
need
to
be.
But
as
far
as
I
know,
v
6.7
u3
does
support
vm
hardware
15..
D
B
I
just
googled
it
I
find
my
own
blog
on
it,
so
it
says
in
my
own
blog
that
the
requirement
is
b
13
for
6
7
u3.
So
let
me
just
check
internally,
it
might
have
been
6.7
u3
and
the
first
version
of
the
csi.
We
said
13,
because
that
was
you
know
what
we
originally
tested
on
and
then
for
the
newer
stuff.
B
B
As
far
as
I
remember,
the
actual
issue
was
to
do
with
persistent
ids
for
the
disks
in
v13.
That
doesn't
exist
as
a
feature
as
part
of
vm
hardware
and
we
don't
retrospectively
port
features
into
vm
hardware.
So
we
rev
the
version
and
we
added
the
ability
to
have
persistent
ids
for
disks
in
15..
So
to
I,
as
I've
always
remembered
it
as
a
bug,
fix
and
not
a
feature
thing,
but
I
I
might
be
misre
misremembering.
B
Oh
there
you
go
vm
hardware,
315
is
from
6.7
u2
perfect.
So
if
you're,
if
you're
on
youtube,
you
should
be
good.
A
A
I'll
move
on
again
somebody
a
few
minutes
back
mentioned
running
arm.
This
is
a
different
form
of
arm,
so
there
is
a
new
update
to
or
a
tech
preview
tech
preview,
meaning
that
it's
you
know,
potentially
it's
not
for
production
and
it's
experimental
we're
looking
for
feedback
from
users
of
the
desktop
hypervisor
for
mac
fusion,
and
it
does
support
the
recent
apple
m1
hardware.
A
So
relevance
to
this
group.
You
know
we
declare
that
we
cover
all
the
vmware
hypervisors
and
some
people
might
be
trying
to
run
minicube
or
some
other
form
of
kubernetes
on
there
on
a
laptop.
So
I've
put
a
link
in
the
agenda,
notes,
document
and
here
I'll,
cut
and
paste
it
to
the
chat
too,
since
I'm
not
sharing
my
screen
anymore,
but
you
can
go
enroll
in
that
tech
preview
and
get
a
download
of
that
and
play
around
with
it.
A
I
don't
run
a
mac
myself,
but
I
took
a
look
at
it
this
morning
and
does
look
like
there
might
be
some
bug
reports
and
things
there.
So
you
know,
like
all
tech
previews.
This
is
for
the
more
adventuresome.
A
Finally,
another
note-
and
this
might
relate
to
david's
comment
about
zoning
because
I
think
we'll
get
into
some
coverage
here,
but
the
kubecon
europe
conference
is
coming
up
in
may
I
think
third
week
and
this
group
has
a
maintainer
track
session
there,
so
michael
gash
will
be
presenting
there
on
the
eventing
integration
with
kubernetes
and
if
you'll
recall,
about
a
year
ago,
michael,
along
with
william
lamb,
gave
a
presentation
to
this
group
about
their
fling
that
uses
event-driven
programming
to
monitor
things
going
on
in
the
vmware
infrastructure.
A
A
That
could
be
combined
with
events
being
generated
from
your
kubernetes
level
and
provide
a
unified
way
to
keep
those
things
together
for
troubleshooting.
So
anyway,
that
session
is
occurring
in
about
a
month.
I
think
it
will
be
good.
That
said,
I'm
co-presenting
with
michael
and
we
don't
have
the
deck
done
yet
so,
potentially,
if
you've
got
any
requests
for
things
to
be
covered,
I
can't
promise
we'll
get
them
in
there,
but
maybe-
and
I
would
anticipate
that-
would
be
an
update
on
this
presentation
that
william
and
michael
did
about
a
year
ago.
A
Also
that
kubecon
conference
is
anticipated
to
be
physical.
Who
knows,
what's
going
to
happen?
You
know
the
covet
news
change
seems
to
change
week
by
week,
month
by
month,
but
back
in
the
pre-covered
days,
I
would
I
know
we
were
hoping
along
with
robert,
when
that
kubecon
was
destined
to
be
in
amsterdam
to
actually
have
a
physical
meeting
of
group
members
and
for
now
anyway,
I'm
planning
on
attending
physically.
C
A
Yeah
same
same
here:
okay,
well
great.
Let's
make
that
a
tentative
plan,
then,
because
it's
been
a
while
since
we've
had
that
face-to-face
experience
and
I'd
look
forward
to
it
anyway.
A
If
the
group
is
four,
I
think
I
can
get
away
with
buying
buying
beverages
and
maybe
a
little
food
so
we'll
see
what
happens.
A
So
we've
covered
our
agenda
that
that
was
the
last
item
on
the
agenda
in
the
notes
document,
but
we've
still
got
five
more
minutes.
If
anybody
has
anything,
they
want
to
bring
up
either
to
talk
about
now
or
to
nominate
as
a
topic
for
next
month.
A
A
Okay,
thanks
everybody
for
attending.
If
anything
comes
up
late,
that
you
wish,
you
had
brought
up,
go
for
it
in
the
slack
channel
and
otherwise
we'll
see
you
in
a
month.
Maybe
some
of
you
will
see
also
at
kubecon
europe
bye.