►
From YouTube: K9P: Kubernetes as 9P Files - Terin Stock, Cloudflare
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
K9P: Kubernetes as 9P Files - Terin Stock, Cloudflare
K9P, a virtual file system, exposes the state of a Kubernetes cluster as files. Our terminals have been optimized over the last 40 years towards working with files, kubectl not so much. K9P allows us to carry the mantra of "everything is a file" to the distributed computing extreme. K9P allows you to integrate Kubernetes resources into an existing workflow, or create new ones. Scale a Deployment by writing to a file. Locate failing Pods with grep. Update configuration in ConfigMaps with sed.
https://sched.co/Uaa3
A
B
Thank
you
Chris,
so
yeah
I'm
really
talking
today
about
k9p,
which
is
a
project
that
I've
been
working
off
and
on
for
the
past
couple
of
months.
So
kubernetes
is
p9
files,
explain
what
that
means
as
we
get
into
it.
But
first,
that's
me
in
the
background.
Is
my
new
ride?
I
parked
it
just
north
of
here
I'm,
a
kubernetes
engineer
at
file
flare
and
you
can
follow
me
at
Taryn,
jokes,
on
Twitter
and
github.
B
So,
let's
think
of
a
hypothetical
system,
what
if
we
could
orchestrate
compute
across
most
multiple
hosts
or
orchestrate
storage
across
multiple
hosts
and
run
compute
separate
from
users?
They
can
work
on
their
local
machine
and
the
computer
can
run
somewhere
else.
Those
applications
can
run
in
the
sandbox,
they
see
their
own
processes,
they
see
their
own
CPUs,
they
see
their
own
files,
they
don't
see
other
things
that
are
running
there
and
there's
a
system
for
service
discovery,
probably
never
heard
of
such
a
system.
Before.
B
Well,
the
next
1.0
is
still
two
years
away
process
containers.
What
we
now
call
C
groups
is
14
years
away.
Let
me
container
that,
for
you
and
docker
is
21
years
in
kubernetes,
wouldn't
be
open
source
for
another
23
years
well
way
back,
then
we
had
a
thing
called
plan
9,
which
was
a
system
designed
around
multi,
multiple
machines,
computing
single
tasks,
or
rather
multiple
machines,
doing
multiple
tasks
where
you
might
have
a
workstation,
that's
separate
from
a
compute
form,
that's
separate
from
a
storage
and
they
could
all
work
together.
B
And
as
much
as
I
want
to
turn
this
into
a
papers,
we
love
I've
been
told
that
I
only
have
35
minutes.
So
instead,
I
do
recommend
going
reading
the
play
nine
papers.
You'll
find
that
a
lot
of
it
that
they
were
thinking
about
in
the
early
90s
is
the
same
types
of
things
that
we're
talking
about
here.
At
this
conference.
B
But
let's
talk
about
k9,
p,
k9
p
is
kubernetes
exposed
as
a
9p
file
system
and
my
goal
when
I
set
the
start,
this
was
to
make
it
easy
to
interact
with
kubernetes.
You
can
use
familiar
tools.
A
few.
Your
know,
your
on
the
command
line.
You
can
use
LS
cat
them.
If
you
prefer
graphical,
you
can
use
Finder
windows
explorer
Nautilus
or
you
can
integrate
it
into
other
work
village.
B
So
what
is
1990
was
a
network
file
system
originally
from
plan
nine
that
it
was
used
all
over
the
system.
So
the
windows
that
were
running
on
your
screen
were
connected
over
P
nine.
You
can
interact
with
them
with
P
nine
network
connections
existed
in
a
file
system
called
net
FS,
which
you
could
interact
with
to
make
new
connections
to
other
systems
and
the
cool
thing
about
these
sub
things.
Since
they're
a
file
system,
you
can
mount
them
over
the
network.
B
So
if
you
want
to
build
a
proxy,
all
you
do
is
connect
to
another
machine
and
mount
its
net
FS.
And
then
you
can
create
new
network
connections
as
that
machine
and
look
at
processes
either
your
own
processes
or,
if
you've
mounted
in
other
processes,
process
file
system.
You
could
see
those
as
well.
You
can
see
open
files.
B
B
B
B
Mac,
OS
and
Windows.
The
story
is
a
little
different.
Both
operating
systems
do
have
support
for
9p
internally
to
the
operating
system.
Mac
OS
supports
it,
starting
in
1014,
for
when
they've
launched
their
new
Mac
Pros
and
you
might
actually
run
Mac
OS
virtualized
and,
as
I
said
before,
Windows
supports
it
four
windows
subsystem
for
Linux.
You
can
use
fuse
implementations
on
both
of
those,
so
you
can
have
a
virtual
file
system
for
your
virtual
bio
system.
B
So
what
are
the
types
of
things
that
I
thought
out
to
do
or
thought
I
thought
out
to
deal
with
with
k9p
when
I
started
out,
so
everything
is
file?
All
you
have
to
know
is
your
usual
file
tools.
So
if
you
want
to
see
information
about
the
deployment
you
can
do
LS
and
you
can
see
the
actual
underlying
yamo
file
and
then
there's
some
other
control
files
that
are
broken
out,
that
you
might
want
to
interact
with
so
the
like
scale
or
you
look
at
the
status
directly
see.
B
B
B
B
And
then
you
might
decide
that
you
want
to
do
even
crazier
things
like
building
a
controller
on
top
of
this,
so
you
can
build
it
in
your
favorite
language.
You
don't
have
to
worry
about
there
being
a
finding
the
kubernetes,
because
there
everything
is
a
file
and
most
tools
have
great
or
most
programming
languages
have
great
file
system
support.
B
B
Don't
think
it
logs
anything
first,
because
why
would
I
do
that
so
here
I'm
using
the
fuse
version,
because
I'm
on
crostini
on
Chrome
OS
and
it
doesn't?
Let
me
mount
9p
file
systems,
so
I'll
choose
the
fuse
version
and
I
mounted
it
to
the
mount
directory.
Should
do
that
and
see
we
have
cluster
scoped
resources
and
named
chase
resources.
So
we
can
look
at
namespaces,
see
all
the
name
spaces
on
the
system
and
like
ways
we
can
look
into
specific
deployments.
I.
C
B
So
how
did
I
implement
k9p
so
the
server's
implemented
in
go?
That's
my
primary
language.
It
exposes
the
90
server
over
TCP
and
it's
using
client
go
under
the
hood
so
that
it
has
great
interactions
with
the
kubernetes
api
right
now
each
mount
is
associated
with
a
new
connection.
The
kubernetes.
This
is
so
that
in
the
future,
you
can
have
multiple
users
connect
with
their
own
credentials,
and
you
won't
have
that
confused
deputy
problem.
All
the
information
will
be
contained
to
one
mount.
B
B
B
B
Some
challenges:
9-p
has
no
polling,
there's
no
way
to
do
subscriptions,
so
this
makes
it
difficult
to
tell
the
client
that
there's
been
changes
to
deployments
or
pods
or
anything
after
they've
kind
of
opened
the
file,
and
so
they
can't
do
like
a
poll
select
or
anything
that
they
might
want
to
do
so.
Hopefully
we
can
work
around
this
in
some
ways,
so
you're
right
now,
you're
required
to
open
the
directory
or
file
to
observe
changes.
B
B
So
it's
a
normal
operating
system
where
you
give
the
process
a
file,
descriptor,
ID
and
19.
The
client
gives
you
the
ID
that
it
wants
the
file
to
be
associated
with
there's
a
map,
the
resources,
a
lot
of
state
is
generated
on
the
man.
The
performance
isn't
that
great
I
was
actually
surprised
at
how
well
it
worked
in
this
demo,
but
normally
I've
sat
there
for
like
20
seconds
after
mounting
it
long
doing
like
ok,
where's
all
the
files,
the
kubernetes
api
uses,
resource
versions
to
prevent
modifications
based
on
old
versions
of
resources.
B
B
B
So
whether
or
not
your
clients
will
actually
see
the
error
messages
you
get
back
is
up
up
up
and
their
authentication
I
haven't
implemented
right
now,
9
fee
defers,
authentication
to
the
implementations,
so
there's
a
file
when
you
first
connect
that
you
can
write
to
in
exchange
with
the
server
any
negotiated
authentication
protocol
right
now.
I
currently
don't
use
this
so
just
automatically
accepts
anything
any
connections
but
in
the
future,
I
want
to
set
up
authentication
protocol,
so
you
can
exchange
it
either
your
user
or
yeah
or
service
tokens.
B
So
some
discussion,
this
kind
of
only
scratches
the
surface
of
what
I'm
P
can
do.
I
do
recommend
trying
out
for
fun
there's
a
lot
of
great
project
ideas
on
github,
where
people
are
playing
around
with
9p,
be
cool
to
have
coupon
talks,
that's
a
file
system,
so
you
don't
have
to
navigate
youtube
if
you're
interested
in
this,
but
don't
really
care
about
files.
You
care
about
blocks,
there's
n
BD,
which
is
network
block
device,
which
is
a
similar
idea
where
you
have
an
RPC
message
for
block
devices
which
I
think.
B
Actually,
if
it
has
a
lesson,
are
PCs,
then
the
9p,
so
it's
pretty
easy
to
implement.
Then
you
can
make
block
devices
appear
out
of
nowhere
and
QS
allows
you
to
build
character
devices.
So
you
can
build
support
for
old
old
devices
in
user
space
rather
than
having
to
upstream
Linux
kernel
drivers,
so
do
recommend
checking
those
out
as
well.
B
If
you
like
this
project,
but
want
to
do
even
more
things
with
it,
such
as
mounting
gke
or
google
cloud
or
AWS
s,
files
Papa
has
wash,
which
is
a
shell
for
interacting
with
remote
and
called
native
infrastructure.
With
unix-like
tools
and
some
of
those
things
are
exposed
in
a
fuse
file
system
that
it
also
has
so
do
recommend
checking
that
out.
B
B
There's
locking
support
for
things
that
you
might
expect
from
a
modern
Linux
system.
Similarly,
sometimes
work,
sometimes
not,
and
whether
or
not
you
can
do
relative
symlinks
or
they
have
to
be
absolute
depends
on
the
client
and
things
like
X,
adders,
sockets
name.
Fifo
is
everything
that
you
might
expect
to
exist
on
a
file
system.
Some
of
those
just
can't
be
represented.
B
So
the
question
is
who's,
coordinating
that
if
you
look
at
the
G,
Weiser
implementation,
I
think
there's
eight
versions
that
they've
put
on
top
of
things
that
only
the
G
Weiser
implementation
or
like
the
internal
Google,
9
P
clients
know
about
so
maybe
those
things
could
be
can
be
used
in
other
places.
But
who
knows
you
can
negotiate
versions,
but
there's
no
standard
and
what
those
versions
are
so
use
at
your
own
risk.
It
might
delete
something
important.
D
E
B
B
B
F
B
Right
now,
if
you
look
at
like
the
Linux
implementation,
it
like
does
a
string
search
based
on
the
strings
that
you've
sent
back.
So
maybe
it's
peaking
at
that
codified
to
the
various
implementations.
You
use
those
same
strings
or
maybe
have
a
different
negotiated
from
version
of
9ps
for
returned
numbers
I'm.
So,
like
g-virus
implementation,
if
you
negotiate
version
instead
of
giving
back
strings,
it
will
give
back
IDs
for
files
for
the
user,
IDs
and
group
IDs.
G
G
B
H
B
B
J
B
There's
a
major
advantage:
I
mean
breast
is
a
very
similar
idea,
just
your
ERP,
C's
or
structured
differently.
You
still
have
crate,
leave
update
so
very
similar
idea.
You
could
turn
a
REST
API
into
a
file
system,
probably
fairly
easily
I.
Think
I've
seen
fused
file
systems
that
do
that.
The
main
advantage
here
was
that
I
wanted
to
put
things
in
a
isolated
namespace
who
didn't
have
network
access
but
could
still
interact
with
kubernetes
which
to
use
a
REST
API
directly
in
that
project
could
require
having
network
access.
K
B
C
B
Scale
end
points
on
replica
sets,
so
if
I
had
implemented
that
you
could,
that
would
already
work,
otherwise
might
just
be
like
a
standard,
a
standard
implementation
just
for
unimplemented
resource
types,
so
you'd
still
get
maybe
status
labels
and
the
EML
format
would
not
have
anything.
Fancier
Thanks.
B
Now
I
don't
do
any
of
the
are
back
stuff.
It
just
relies
on
when
you
do
that
when
you
do
lookup,
so
you
make
changes.
Does
the
kubernetes
api
server
reject
that?
I
could
implement
that
things
so
that
when
you
do
an
LS,
you
can
see
the
correct
permissions,
but
I
haven't
and
what
did
that
right
now
just
says
you
have
read/write
two
things.