►
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
Linkerd 2.0 brings dramatic improvements to performance, resource consumption, and ease of use to Linkerd. With its minimal footprint and incremental approach, it’s designed to give service owners and platform owners the critical tools they need to be successful in a cloud native environment.
In this webinar, join William Morgan, core maintainer of Linkerd, as he shares a demo and introduces new Linkerd 2.0 features. William will be joined by Ben Lambert, CTO of Just Football, who will describe how Just Football uses Linkerd 2.0 in production to scale their wildly popular service in the face of massive amounts of traffic.
You can expect to learn how to:
- Diagnose and identify instability and latency in your Kubernetes services.
- Access granular level metrics on how each service is behaving.
- Watch real-time requests and responses for all our container service debugging.
- And more!
A
With
some
asterisks
and
I'll
show
off
some
of
that
stuff
in
the
demo
right
now,
Dex
is
kubernetes
only
we
are
going
to
expand
that
in
the
future,
but
we
had
to
start
somewhere,
and
so
it's
a
so
starting
point
for
us,
okay
and
a
service
mesh.
Hopefully
you
all
know
this.
By
now,
service
mesh
is
an
infrastructure
layer
that
tries
to
manage
the
runtime
communication
between
your
different
services.
So
there's
a
reliability
aspect,
there's
a
security
aspect
as
a
visibility
and
observability
prospect
and
I'll
show
off
some
of
that
stuff
in
the
demo.
A
So
far,
so
good
send
all
complaints
to
Tamaki
via
chat,
I'm
gonna
jump
straight
into
the
demo
and
locking
believe
me
earlier,
and
he
said
that
I
should
show
off
some
of
them
really
experimental,
really
fancy
stuff
and
so
I'm
gonna
try
and
do
that,
but
it's
extremely
risky.
So
if
it
doesn't
work,
we
all
know
they're
to
blame
all
right.
So
let's
get
my
terminal
and
everyone.
Okay,
so
I'm
in
a
demo
not
just
link
your
D
2.0,
but
the
edge
release.
A
So
we
have
weekly
edge
releases
and
if
you
run
this
command
curl
die
chefs
run
down
like
a
good
eye
on
slash,
install
edge.
You
get
the
CLI
that
will
bring
you
the
latest
edge,
so
really
living
on
the
edge
okay.
So,
let's
see
I
ran
a
strand
already
because
it
takes
like
you
know,
and
we
can
add
to
see
a
live.
My
path
so
now
I
have
the
ability
to
say,
link
redeem
and
the
very
first
thing
that
we're
gonna
do
we're.
Gonna
run
this
command
called
maker
D
check.
A
Let
me
drag
this
to
zoom
window
over
to
the
side
here
out
of
my
way
and
just
to
set
the
context
here.
Actually,
even
before
I
run,
this
command
I
have
an
empty
kubernetes
cluster,
nothing
up
my
sleeve
it's
running
on
docker
format,
kubernetes
and
I.
Don't
spend
up
for
62
days.
Possibly
you
know
the
longest
little
communities
cluster
in
the
universe
and
that's
just
been
like
adding
and
removing
services
to
it.
So
that's
been
fun,
okay
and
then
we're
gonna
do
liquidy
check.
A
Sorry
you
check,
Korean
and
and
what
this
command
does
is
first
checks
to
see
whether
the
kubernetes
cluster
is
like
ready
for
later
deep
right,
because
there's
all
sorts
of
weird
are
back
things
that
could
happen.
There's
all
sorts
of
you
know
different
versions
or
whatever.
So
this
is
just
the
same
any
check:
okay,
everything's
fine.
We
can
install
winter
date
on
this
on
this
cluster
and
actually
I.
Don't
know
if
you
are
a
kubernetes
expert,
but
you
can
see
create
some
cluster
moles
and
cluster
bindings.
A
We
have
an
experimental
feature
that
I'm
not
going
to
demonstrate
called
a
single
namespace
where,
rather
than
cluster
wide
permissions,
we
actually
just
look
at
rolls
and
roll
bindings.
If
you,
if
you
are
a
user
of
kubernetes,
you
have
contain,
you
have
access
to
a
namespace,
but
you
don't
have
access
to
the
cluster
as
a
whole.
A
You
can
use
this
and
even
the
salt
lingered
in
the
same
namespace
as
your
app
I'm
not
going
to
demo
that,
but
we
want
to
really
make
it
as
light
and
was
easy
to
install
liquidy
into
a
kubernetes
cluster
as
possible.
We
don't
want
you
to
have
to
ask
permission
all
right,
for
anyone
should
be
able
just
do
so,
but
but
for
this
you
know
for
this
demo.
A
I'm
gonna
actually
install
I'm,
gonna
kind
of
demo,
the
opposite
side,
which
is
we're
gonna,
do
a
cluster
wide
installation,
in
fact
we're
gonna
use
and
mission
webhook
mutating
controlling
things
to
automatically
add
new
services
to
the
linker,
the
control
plan.
So
let
me
actually
boot
up
my
demo
app
first
okay.
So
before
we
do
anything
with
with
with
linker
D
right,
we
run
the
CL
ICU
license
everything's
fine.
A
If
we
get
all
name
space,
we
will
see
whoopsie
we're
going
to
do
name
space
right,
we'll
see
that
it's,
you
know,
kind
of
a
couple,
pods
some
services
and
then,
as
this
web
service,
which
has
a
cost
for
an
external
IP,
because
this
is
Dockers
for
a
Mac
plus
kubernetes
I
can
actually
look
at
localhost
right
on
my
laptop.
So
let
me,
let's
close
this
thing:
let's
check
out.
A
A
If
we
look
at
the
kubernetes
dashboard,
let's
actually
go
to
the
so
I've
got
a
refresh
it
when
they
get
to
the
emoji
photo
namespace
all
right,
so
we're
looking
at
the
kubernetes
dashboard.
Everything
is
green,
all
right
as
far
as
kubernetes
is
concerned,
everything
is
fine.
The
pods
are
running,
you
know,
health
checks
are
responding
right
and
and
it's
not
a
failing
of
kubernetes.
That's
just
what
kubernetes
understands
about
this
world
is:
are
the
health
checks
responding
and
are
the
pods
running
and
everything's
fine
right?
A
So
now
we're
in
this
very
common
situation,
which
is
we
have
an
app
that
we
know
it's
broken.
You
know,
but
kubernetes
doesn't
know
anything
about
them
because
kubernetes
doesn't
know
about
success
rates,
it
doesn't
know
about
services
being
slow.
Do
you
know
about
any
of
that
stuff?
Ok,
so
what
I'm
gonna
do
now
is
I'm.
Actually
I'm
gonna
delete
this
I'm
just
gonna
delete
the
name
space
delete.
A
And
we're
gonna
we're
gonna
reinstall
this
app
in
a
second.
The
reason
why
I'm
going
to
do
that
is
because
it
basically
makes
the
demo
a
little
bit
easier.
Ok,
so
now
that's,
let's
install
the
linker
decontrol
plane
all
right
so
enough
about
emoji
voto,
let's
install
into
be
control,
plane
and
I'm.
Passing
in
these
two
highly
experimental,
highly
dangerous
flags.
A
Okay,
one
is
called
TLS
equals
optional,
which
means
that
we
are
gonna
actually
put
TLS
in
between
every
single
service
automatically,
and
then
this
proxy
auto
injects
means
that
we
are
going
to
use
kubernetes
fancy
new
dating
web
book
admission
controllers
or
to
automatically
inject
linker
key
data
plane
pods
next
to
every
application.
This
is
like
the
super
fancy
way
to
do
it.
There
are
less
fancy
ways
of
doing
it
but,
like
I,
said,
locking
kind
of
bullied
me
into
showing
off
the
good
stuff,
and
this
was
a
contribution
from
someone
named
Ivan
sim.
A
A
It's
often
a
namespace
called
:
linker
D,
like
various
things,
are
like
booting
up,
because
this
is
running
on
docker
for
Mac
and
meanwhile
using
zoom
like
you
know,
which
consumes
500
course,
but
this
liquidy
check
command
I,
don't
know
if
you
notice
that
we'll
just
wait
there,
it'll
retry
until
everything's,
ready,
okay,
so
now
everything's
firing.
You
have
your
D
check.
Alright,
again
the
instantaneous
everything's
okay.
So
now
I
have
the
linker
D
control
play.
A
In
fact,
we
can
just
look
at
the
dashboard
okay,
so
we
have
the
control
plane
installed
in
this
namespace
called
linker
D
by
default
and
well
link
your
D
is
already
good.
At
is
understanding
what
is
the
traffic
and
what
is
it?
What
are
the
success
rates
of
all
the
various
components?
In
fact,
later
d
by
default
also
monitors
itself,
so
we
can
look
down
in
here.
We
can
say:
oh,
this
is
everything
that's
in
the
linker
D
namespace.
We
can
see.
There's
a
fauna,
Prometheus,
there's
two
new
things.
A
Luckily,
that
might
be
new
to
you.
The
proximately
of
this
proxy
injector
component,
we
have
a
CA
component,
so
actually
linker
D
is
going
to
be
running
its
own
CA
and
it's
going
to
be
generating
kilohertz.
That's
we
distributing
them
to
each
of
the
services.
So
again,
zero
config
writes
well,
I,
don't
want
to
I.
Don't
wanna
have
to
write
any
animal
at
some
point.
There
might
be
a
little
bit
again,
but
right
now,
there's
no
y
anymore
and
now
I'm
going
to
install
an
ogee
mode
again.
A
So
let's
do
that
same
thing
background
the
dashboard.
Let's
do
emoji
namespace,
but
let's
install
that
thing
here:
okay,
so
I'm,
taking
that
same
manifest
on
the
planet
again
now
I
mentioned
previously
that
there
are
other
ways
of
doing
this
in
the
demo
than
I
normally
give
emoji
voter
remains
running
the
entire
time
we
install
them.
We
control
playing
over
to
the
slide,
and
then
we
balanced
the
pods
and
then
the
app
remains
up
the
entire
time.
That's
pretty
fun,
but
for
this
case,
I
wanted
to
show
off
the
fancy,
auto
injection.
A
So
what
you'll
notice
here
that
I
haven't
done
anything
other
than
apply
that
manifest?
It's
usually,
in
fact,
if
we
go
now
to
the
linker
d
dashboard
will
see.
We
are
automatically
discovering
this
emoji
voter
namespace
and
we're
discovering
that
the
pods
have
been
meshed
and
we're
starting
to
see
metrics
for
each
of
those,
things
will
have
to
drag
zoom
over
to
the
other
side.
Okay,
so
what
linker
D
is
done
with
all
this
fancy
magic
and
once
again,
there
is
a
non
magic
version
of
this,
but
I
want
to
show
the
magic
is.
A
It
has
installed
the
tiny
little
product
see
next
to
every
pod
reason:
old,
sidecar
pod.
In
fact,
if
we
look
at
the
Cuba
needs
to
pin
a
nice
dashboard
okay,
this
was
the
old
version.
If
I
hit
refresh
here,
what
we'll
see
is
a
slightly
different
view,
which
is
yes
not
only
do
we
have
that
emoji
vote
a
web
service,
what
we
have
a
little
proxy
next
to
it,
and
then
this
thing
is
like
an
init
container
that
sets
up
the
docent
iptables
magic
so
that
all
traffic
is
going
through
here.
A
Okay,
so
going
back
to
my
linker
to
dashboard,
what
we'll
see
is
immediately.
We
have
these
deployments
and
they're
starting
to
see
success,
rates
right
and
obviously
emojis
in
great
shape.
We
have
a
big
red
funny
face
next
to
the
voting
service
of
next
web
service.
Folk
bot
is
a
little
bot,
that's
sending
traffic,
so
it
actually
doesn't
have
any
inbound
traffic.
So
there's
no,
no
success
rate
really
to
report
for
it.
A
So
how
do
we
actually
debug
this
right
and
I
want
to
give
you
the
tools
to
debug
this
thing
in
a
way
that
doesn't
involve
any
kind
of
rewriting
the
graph
instrument
in
your
code
or
anything
all
right,
and
so
we're
looking
now
at
the
linker
D
dashboard
for
the
web
service,
and
what
we
can
see
is
that
you
know
here's
our
web
service
in
the
middle.
It
has
this
miserable
success
rate,
it's
making
it's
taking
traffic
from
the
pot
and
extending
traffic
to
emoji
the
emoji
service
into
the
voting
service.
A
And
obviously,
one
of
these
is
fine
and
one
of
them
that's
failing
all
right.
We
can
click
on
the
voting
service
and
you
can
see.
Okay
voting
service
is
taking
traffic
from
the
web
service.
It
has
some
miserable
success
rate
taking
about
one
request,
a
second,
that's
pretty
good
for
my
laptop,
which
is
starting
to
heat
up
and
in
this
table
down
here.
What
we
have
is
a
live
sampled
view
of
all
the
requests
that
are
happening
to
the
service.
A
Ok,
and
we
can
see
that
you
know
a
lot
of
them
are
totally
fine,
but
some
of
them
vote.
Squishy
round-faced,
is,
as
is
zero
percent
success
rate.
Okay,
so
we're
starting
get
a
clue
now
of
where
the
problem
might
be.
In
fact,
we
can
click
on
this
tap
interface
here
and
I'm
actually
going
to
stop
this
here.
This
is
another
tool
I
really
like
from
from
liquor.
D
now
have.
This
is
now
tapping
all
requests
between
web
and
voting.
A
So
I'll
show
me
a
sample
view
of
all
the
live
requests
that
are
happening
between
the
web
service
and
the
buildings
for
this,
and
we
can
see
that
you
know.
There's
a
lot
of
requests,
and
this
app
is
a
gr
PC
app
for
those
of
you
who
are
familiar
with
gerrae
can
see
everything
is
a
HTTP
200,
but
the
G
RPC
status
is
either
ok
or
unknown
and
under
me
there.
Ok,
so
what's
actually
pulse
that's
topless
now
so
we
can
see.
You
know
every
time
that
we
that
we
need
this
call
to
vote.
A
Squishy
smelly
brown
face
is
an
error
and
all
the
other
calls
are
ok.
So
at
this
point
we
have
now
the
tools
that
we
need
to
go
talk
to
the
team
that
owns
a
service
or
to
start
debugging,
the
actual
code
itself,
or
to
look
at
the
logs
or
whatever
it
is.
We
know
that
this
one
call
is
about
failing
yeah,
ok,
so
that's
the
nice
kind
of
debugging
story
again
without
having
to
do
anything
to
our
application.
We
don't
even
know
what
language
this
stuff
is
written
in
right.
A
We
have
an
instrument,
anything
there's,
two
brief
things:
I
want
to
show
off.
First,
before
we
go
back
to
slides,
first
is-
and
you
may
have
noticed
that
in
this
table
we
actually
have
these
little
orange
links
here.
Let's
pick
a
good
one
here,
let's
go
to
you
when,
and
this
will
take
you
to
the
pre-configured,
perform
a
dashboards
just
a
little
bit
slowness,
Llodra
file,
and
so
we
have
what
we
try
and
do
is
build
you,
these
pre-configured,
their
final
dashboards
for
each
service.
A
Let
me
show
you
kind
of
the
top-line
metrics
or
the
golden
matrix
here.
So
we've
got
one
for
success
rate.
Here's
a
success
rate
web.
Here's
a
request
rate
here
is
the
latency.
Then
we
can
see
we
got
the
report
in
p9,
1995
and
P
50,
and
everything
was
looking
good
until
I
started
for
fauna
and
if
it's
a
spike
another
kind
of
back
to
I,
wouldn't
call
that
very
good,
and
the
other
thing
you'll
notice
here
is
that
we
have
this
little
lock
icon
next
to
the
web
service.
A
I,
don't
know
if
you
can
see
that
I
can't
move
the
mouse
over
it,
which
means
that
all
this
traffic
is
being
pls.
If
we
had,
if
this
web
service
we're
taking,
you
know
different
bits
of
traffic
that
we're
taking,
you
know
non,
kill
us
and
kill
us
and
we'd
see
those
split
out
here
and
then
finally,
we,
actually
you
know
if
you
scroll
down
and
in
here
we
actually
group
your
inbound
and
outbound
traffic.
Here
you
can
see
so.
A
Here's
inbound
here's,
the
outbound
traffic
that
the
web
services
making
that's
making
bunch
of
calls
to
it
Mogi
service
and
a
bunch
of
calls
to
the
voting
service.
We
split
out
the
request
rate
by
that,
and
in
fact
we
all
just
put
out
the
success
rate.
We
try
and
give
you
these
dashboards
all
set
up
by
default
and
want
to
make
your
life
easy
and
that's
why?
A
If
you
were
wondering
you
know
in
this
little
view
of
the
linker
d--
namespace
here
we
have
a
little
graph
on
ax
and
a
little
Prometheus
instance,
and
the
way
that
works
is
we
we
cap,
these
we
put
pretty
tight
boundaries
on
the
amount
of
data
and
Prometheus
will
start
expiring
data.
You
know
after
a
couple
hours.
This
is
not
meant
to
be
a
huge
for
me,
Thea's
install
that
you
then
have
to
manage,
because
it's
holding
all
metrics
for
all
time.
A
We
want
you
to
get
your
data
out
of
this
Prometheus
and
put
it
into
a
real
one.
It
was
basically
cured
in
order
to
do
you
know
some
fancy
fancy
stuff
for
us.
That's
part
of
the
dashboard
and
also
as
part
of
the
CLI
I,
haven't
demoed
this
too
much,
but
we
can
do
things
like
I
know,
I
need
to
give
it
to
Tony.
A
All
right
good,
so
this
is
actually
a
noose,
my
form
a
little
bit.
This
is
a
CLI,
you
reduce
it
more.
So
you
can
see
all
this
stuff
see
a
live
view
of
one
of
those
tables.
So
you
don't
just
have
to
look
at
the
web
browser
actually
when
they
give
you
a
CLI
view-
and
this
is
you
know
literally
it's
like
top.
You
know,
looking
at
processes
except
here
we're
looking
at
calls
R
aggregating,
you
know,
so
we
can
see
all
right
since
we
started
to
talk.
A
All
right
clearly
I
am
the
type
of
person
who
clicks
around
at
a
web.
Ui
CLI
locky
probably
knows
how
to
run
this,
and
you
know
refer
to
his
video
for
for
the
answers.
Ok,
so
let's
go
back
here
now.
This
is
basically
the
end
of
the
of
the
demo.
There's
a
couple
things
I
wanted
to
point
out.
The
first
is
that
you
know
what
I
demoed
here
was
like
the
full
fancy
version
where
we
install
CA
we're
doing
TLS.
We
have
cluster
wide
access.
A
We
have
a
you
know:
mutate
web,
put
the
admission
controller,
there's
a
much
simpler
kind
of
easier
world
where
you
don't
have
any
of
that
stuff.
You
just
deploy
into
a
single
namespace.
It's
really
a
kind
of
a
different
focus
for
the
product
or
for
the
for
the
project,
but
we
want
to
make
it
really
flexible
and
most
of
all,
we
want
to
make
this
really
fast
and
lightweight.
A
A
Actually,
when
we
do
this,
let
me
save
my
precious
CPU
for
zoom
delete
that
thing,
because
the
bot
is
actually
running
in
sunny
traffic
okay,
so
that
was
the
demo.
So
let
me
give
you
a
little
context
about
linker,
D
2.0
and
why
it
exists.
What
I
just
done
loading
this
I
know
you
know
linker
D
I
think
he
saw
in
the
previous
life
it's
been
around,
for
you
know
it's
been
around
for
three
years
of
those.
A
You
know
it's
been
in
prod
for
at
least
two
years,
and
it
was
the
original
service
mission
brought
that
term
to
the
world,
and
we've
learned
a
lot
since
going
through
that
process,
because
it
spent
a
lot
of
our
day.
Not
just
you
know,
making
a
linker
be
better,
but
actually
trying
to
help
companies
and
and
engineers
adopted
into
their
architecture,
and
what
we
learned
is
that,
basically,
the
you
know
the
value
prop
to
the
service
mesh,
really
great
right.
There's
a
lot
of
value
there.
A
You
get
reliability,
you
get
visibility,
you
get
security,
you
know,
and
you
get
it
all
at
the
infrastructure
layer.
You
don't
have
to
get
the
developers
to
like
write
their
own
major
object,
but
there
are
some
big
downsides
to
link
31
dot
X
and
to
kind
of
the
other
service
meshes
that
follow
that
pattern,
which
is
that
they're
big
pieces
of
equipment
right.
It's
you've
just
adopted
koukin
IVs,
which
is
you
know,
huge
amount
of
machinery,
and
now
we're
asking
you
to
adopt
this
other
big
layer
of
infrastructure.
A
And
if
you
do
well,
you
know
this
good
value
out
of
it.
But
you
know
it:
shouldn't
have
to
be
that
complicated
shouldn't
feel
like
this
big
infrastructure
layer.
So
a
lot
of
what
we've
done
with
the
$2
own
release
of
link
Athene's
it
actually
strip
stuff
out
and
try
and
make
it
as
minimalist
as
that's
possible.
You
know
again
so
that
you
as
a
developer
even
could
install
it
in
namespace
and
not
have
to
go
through.
A
So
it
shouldn't
have
to
be
like
you
install
a
service
across
everything
or
you
get
nothing,
you
should
be
able
to
do
it
incremental
and
then
the
final
big,
big
kind
of
learning
I
mean
I'm
not
going
to
dwell
on
this
too
much.
Is
that
everything
that
we've
built
in
in
the
one
9x
release-
and
you
know
emphasize-
that
our
continuing
I
mean
game
isn't
like
we've
got
lots
of
adoption
and
we
love
our
one,
not
accusers,
but
you
know
we're.
A
The
focus
was
really
on
this
on
the
kind
of
platform
operator
roll
right
over
the
platform
owner,
and
we
were
felt
we
had
this
feeling
that
we
were
under
serving
the
service
owner
or
the
actually
developer.
We
had
to
use
kubernetes
and
for
whom
the
platform
owner
is
actually
building,
stuff,
right
and
and
I
think
we're
all
kind
of.
On
the
same,
we
all
have
the
same
goal,
which
is
we're
not
doing
infrastructure
for
infrastructure
sake.
I
mean
it's
fun,
but
like
it's
not
it's
not.
A
You
know
it's
a
means
to
an
end,
and
that
end
is
we
want
to
have
our
developers
and
we
want
to
have
our
service
owners
be
able
to
do
their
jobs
better.
So
a
lot
of
what
we've
tried
to
do
with
with
to
dotto
is
give
tools
to
service
owners
developers
to
the
people
who
has
to
use
this
stuff
make
their
lives
better
in
in
a
very
direct
way.
Okay,
so
I
said
a
lot
of
this
in
the
previous
slide,
you
know
we've.
We
saw
a
lot
of
these
views
during
the
demo.
A
We
saw
kind
of
the
service,
centric,
dashboards
and-
and
you
know
some
of
the
debugging
and
diagnostics
tools
we
can.
You
can
actually
install
the
link
again
just
one
service
which
it's
really
cool.
So
if
you
own
one
service
and
one
namespace,
you
just
get
a
lot
of
value
out
of
linker
D,
even
without
having
to
ask
for
permission,
zero,
config
sixty
seconds
to
install
at
least
no
Plus
Gravano.
A
You
know
health
check
time
or
whatever
and
then
a
big
focus
on
visibility
and
debugging,
and
then
we
saw
some
of
the
CLI
tools,
no
permissions
required
inside
learning,
tagline
and
then
for
platform
engineers.
The
consequence
of
this,
which
I
really
like
that
you
can
actually
be
incremental
about
linking
right.
You
can
run
it
on
one
service
and
then
you
can
add
another
one
in
I
know
and-
and
so
like
you
know,
in
one
circuit-
is
constantly
calling
service.
A
So
much
I
think
we
call
it
service
sidecar,
you
know,
because
just
one
servicing,
what
are
you
missing?
But,
as
you
add
more
more
service
of
this,
you
know
you're
kind
of
in
back
into
mesh
world.
This
should
be
good
for
for
platform
owners
as
well
for
service
owners,
because
you
can
be
incremental,
you
can
do
things
slowly
and
then
safely
alrighty
and
since
we
have
engineers
on
the
call
we
spend
three
seconds
talking
about
the
architecture.
This
is
actually
pretty
similar
to
every
other.
A
Almost
every
other
service
mesh.
That's
out
there
there's
a
control
plane,
which
here
we
have
like
kind
of
expanded
on
the
top
there's
a
data
plane.
The
data
plane
is
pretty
easy
right.
You
can
have
a
bunch
of
proxies.
You
stick
it
next
to
the
application
as
sidecars
laker,
D
1
dot
X,
you
can
use
it
as
a
side
bar
you
can
use
it.
A
As
a
daemon
said,
we
had
a
lot
of
demons
that
deploys,
but
you
know
it's
just
as
easy
to
use
it
as
a
sidecar,
and
there
are
some
users
to
do
that
for
two
decks
really
focus
on
the
side
run,
use
case
and
the
reason
why
we're
able
to
do
that.
It's
because
these
proxies
now
are
super
super
lightweight
and
super
fast.
A
So
we've
built
this
networking
stack
in
rust,
which
allows
us
you
know
to
basically
be
native,
have
native
code
performance
with
all
sorts
of
nice
security
guarantees
and
like
you're,
not
gonna,
have
a
buffer
flow
at
XP,
buffer,
overflow
exploit
or
anything.
Because
rust
has
this
in
a
very
strict
memory
model,
but
there's
also
no
garbage
collector
right.
So
you
can
amortize
all
the
memory
usage
across
the
request.
Traffic,
which
means
if
you're
late
and
seeds
are
super
spiky.
Your
latency
distribution
is
super
sharp,
which
is
what
you
want.
A
A
You
can
have
been
a
1
per
application,
part
of
a
service
pod,
and
then
we
have
the
control
plane,
which
is
sitting
off
to
the
side
and
the
demo
that
I
showed
that
those
were
the
components
that
we
were
looking
at,
where
there's
a
web
component
and
there's
a
controller
component
of
Prometheus.
It's
Agrafena
in
this
diagram.
We
don't
have
the
CA,
but
there
was
a
CEA
component
if
you
use
that
TLS
equals
optimal
flag
and
there's
like
it,
the
the
webhook
component.
A
If
you
use
the
such
proxy
auto-injector
component,
you
can
add
you
know
you
can
make
this
like
kind
of
fancier
and
fancier,
but
the
whole
control
plane
itself
is
also
designed
to
be
small
enough,
that
you
can
run
it
in
a
single
namespace
for
single
service
and
it's
not
like
500
Meg's.
It's
like
15
minutes
or
something
ever
noise.
I
think
it
was
under
50
last
time.
A
Okay,
so
think,
small,
lightweight
and
and
and
fast
okay.
So
that's
the
end
of
my
linker
D
demo,
$32
demo
and
monologue.
We
g8
linker
D
2.0
in
just
just
last
month,
we've
had
a
excellent
community
of
the
doctors
and
contributors
and
I
wanted
to
have.
You
know
to
try
and
make
this
more
real
and
less
of
like
a
cerebral
thing.
A
A
B
Of
course,
so
the
football
is
a
start-up
where
our
based
in
Stockholm
so
everything's
kind
of
pitch
black
outside
and
very
cold
a
minute.
We
starting
2016.
We
start
with
four
people
and
now
about
15
people,
so
yeah
we're
going
going
slowly,
but
surely
we
build
an
app
basically
for
kids
to
kind
of
get
them
out
of
housetop
the
planar
consoles
and
start
doing
some
exercise
lately.
B
So
we
basically
try
and
take
the
concept
of
like
FIFA
or
like
NHL
mud
and
all
these
console
games
that
kids
play
all
the
time
and
then
try
and
combine
that
with
actually
playing
the
game
in
real
life.
And
then
you
progress
in
a
virtual
career.
So,
basically,
like
we
give
you
in
terms
of
football,
that
you
go
to
a
pitch
you
check
in
and
then
we
give
you
challenges
to
play.
We
can
do
like
all.
They
asked
of
things.
A
A
B
We
deployed
to
address,
we
have
to
a
device
account,
so
everyone
per
dev,
which
means
that
we
then
out
to
keep
net.
It
goes
to
manage
it
out.
I
know
you
did
some
way.
We've
thought
about
basically
bring
just
one
cluster
and
having
some
cool
stuff
for
services
to
just
do.
Devin
live,
but
for
now
we've
got
to
we're
not
using
UK
or
anything
like
that.
It's
just
all
cops
for
deployment.
We,
our
architecture,
is
basically
not
micro,
servitude,
not
micro,
so
see
or
not
to
much
much.
B
B
B
So
language
wise,
we
have
a
lot
of
node
in
typescript.
We
also
have
started
porting
a
lot
of
stuff
over
to
go
like
we're.
A
start
like
we
have
to
write
stuff
really
fast
and
kind
of
node
allows
us
to
do
that,
like
it's
kind
of
a
very
quick
bootstrap
language
like
we
can
get
up
and
running
pretty
quickly.
Well,
then,
for
the
most
of
the
wastes
or
semi-solid
in
we
come
over
to
go
yeah,
that's
kind
of
in
terms
of
how
they
communicate.
Everything
is
done
internally
with
G
RPC,
G
RPC,
good
idea.
B
Yeah
I
mean
it's
interesting.
It's
we
basically
have
everything
as
HTTP
to
start
with,
and
then
we
started
load
testing
and
realizing
that
there
was
so
much
overhead
and
the
amount
of
connectors
are
the
amount
of
like
resources
that
uses.
So
our
architecture
basically
starts
off
with
a
lot
of
micro
services,
but
we
have
one
public
API,
which
is
Baskerville
and
bath.
Girl
is
like
the
the
API
like
the
phone
client
talks
to
the
thing
about
the
graphical
client.
B
So
yeah
we
switch
to
GRDC
I
mean
it's.
It's
pretty
good,
like
we
kind
of
every
the
thing
is
really
hard,
especially
not
be
doing
Haiti
to
peel
and
is
rest.
Fbi's
are
very
hard
to
design
and
every
developer
has
a
different
idea.
So
we
basically
end
up
with
so
many
different
ways
to
communicate
some
data
that
the
service
needs
in
the
URL
some
days
in
the
post
body
like
it
very
hard
to
manage,
whereas
GLP
sees
all
kind
of
schemas
and
provo.
So
it's
kind
of
nice,
you
get
a
truth.
A
B
A
B
I'm
from
the
UK
originally
not
from
Stockholm,
not
a
Swede
but
I
watch
the
talk
that
he
gave
about
kind
of
his
cell,
like
communities
and
movie
possibilities
and
then
having
to
do
all
the
interesting
stuff
in
the
middle.
So
I
always
want
to
get
start
with
it.
We
tried
the
kid
he
won
but,
like
you
said
earlier,
it's
very
or
nothing
like
and
we
couldn't
really
get
it
working
for
us,
but
Lickety
too.
However,
it
works
super
nice,
great.
A
B
B
Is
it
working
like
it
makes
demo
earlier,
it's
very
hard
to
know
when
our
services
performing
well
just
by
our
health
checks
right,
like
we
ship
a
lot
of
stuff
to
live
every
day
like
we
want
to
know
if
something
breaks
quicker
than
just
a
health
check
like
we
won't
be
able
to
see
in
real
time.
So
initially
we
just
use.
B
We've
got
girl,
but,
like
anything,
new
that
we're
building
now
has
leaked
could
be
cycled
in
next
to
it,
so
that
we
can
track
and
monitor
that
service
when
it
comes
from
new
into
both
furthered
online
over
time
with
real
live
traffic,
like
it's
very
hard
also
for
us
to
get
an
idea
about
what
services
talk
to
what
like,
when
you're,
moving
fast
and
building
stuff
and
shipping
stuff
like
no
one
really
has
any
idea
about
what
service
to
talk
to
water
very
high
level.
I
was
a
link.
Addy
gives
you
that
network
topology
yeah.
A
That's
it.
You
know
it's
crazy,
that
that
is
such
a
common
problem,
but
you
know
yeah.
You
know
back
when
I
worked
on
her.
This
was
like
you
know.
Six
years
ago
it
was
impossible
to
keep
track
of
who
you
know
who
was
talking
to
him
like
they
make
a
spreadsheet.
You
know
after
two
weeks,
but
after
like
after
an
hour
like
that's
rich,
it
would
be
obsolete
because,
somewhere
to
chain
yeah.
A
B
So
we
have
both
the
lickity
dashboard
on
the
graph
on
our
dashboard
on
all
times
like,
especially
for
graph
QL
like
that
will
tell
us
when
something
downstream
is
broken
like
something
that's
mission,
critical
or
like
the
is
the
tier
0
service
or
whatever
that
is
used
by
graph
QL
and
graphical
hotels.
Myself
is
broken
by
using
the
dashboards
that
link
if
he
provides
like
it's
very
hard
to
kind
of
it's
very
hard
to
understand
when
the
dashboard
or
when
something
breaks.
B
A
B
So
I
can
give
two.
So
one
is
we
just
a
few
weeks
ago
we
deployed
two
services
concurrently
that
both
talk
to
each
other
or
that
basically
how
to
communication
with
each
other,
but
they
were
deployed
concurrently
Jenkins.
So
we
use
Jenkins
for
CI,
so
we
have
pipelines
that
run
and
they
could
run
concurrently.
So
we
made
a
change
at
the
same
time.
B
Obviously
we
test
everything
on
CI,
but
they
were
only
testing
with
the
latest
version
of
each
service
which
was
changing
currently
and
then
that
ended
up
getting
all
the
way
into
production
and
basically
there's
a
schema
incompatibility.
Well,
a
probe
of
incompatibility
between
the
two
versions,
so
that's
very
hard
to
test
with
the
health
check,
because
you
don't
see
the
health
checks
failing
because
the
schemas
finally
like
it's
returning
properly.
But
you
don't
see
the
problem,
so
yeah
open
up
link
a
new
dashboard
and
we
can
find
out
exactly
what
the
problem
is
straightaway.
B
A
B
When
we
switched
over
to
G
RPC,
we
always
do
a
lot
of
stuff
written
in
node
and
the
documentation
of
the
Geo.
Pc
client
isn't
great,
like
it's
getting
there,
but
we
just
totally
forgot
about
timeouts.
So
we
were
calling
all
these
services
would
not
release
ever
timeout
for
the
dependent
service
dependent
service.
So
I
think
it
defaults
to
30
seconds
and
that's
a
long
time
when
graph
gives
way.
Thirty
seconds
will
reply
to
violently
so
yeah
I
mean
linka.
B
D,
obviously
was
Susan
putting
kadhi
and
we
could
tell
what
the
problem
was
because
Lincoln
D
then
changes
the
timeout
to
ten
seconds
by
default,
and
you
can
configure
that
so
that
kind
of
helped
us
out
a
hell
of
a
lot
and
when
we've
got
four
services
going
through
and
updating
the
clients
for
each
one
of
them
downstream.
So
it's
kind
of
a
big
task.
So
this
helps
us
out
a
lot.
B
B
A
Be
in
2.1,
which
should
be
out
before
cube
time,
we're
gonna,
introduce
this
notion
of
service
profiles
or
you
can
actually
on
a
service
by
service
level.
You
say:
here's
the
you
know:
here's
the
retry
behavior!
I
want
for
this.
Here's
the
timeout
behavior
that
I'm
watching
us,
here's
the
circuit,
breaking
behavior!
You
can
start
parameterizing
things
like
service.
A
That's
gonna
set
us
up
to
give
you
even
a
better,
even
better
level
of
configurability,
good.
Okay,
then!
Well,
thank
you
very
much
for
joining
us.
Actually,
I,
don't
think
you're
free
of
webinar
jail
yet
because
machi
is
now
gonna
do
a
Q&A
and
there
may
be
questions
for
you.
We'll
see
we'll
see
who
has
said
the
craziest
things.
C
Is
it
a
favorite
part
I
actually
really
like
the
the
kind
of
fireside
format
so
kudos?
Putting
that
together
and
hearing
from
production
uses
are
fantastic,
so
we've
had
a
inundation
of
questions.
There
are
18
in
the
queue
I'm
gonna,
try
and
summarize
some
that
are
similar
and
I'll.
Give
you
the
opportunity
to
answer
them,
but
just
keep
in
mind.
I
will
try
and
get
through
as
many
as
possible,
and
obviously
you
can
ask
your
questions.
A
Currently,
no,
but
that
is
an
interesting
feature
request
and
we
you're
not
the
first
to
ask
for
it
right
now.
The
the
behavior
the
house
starting
point
for
us
is:
we
don't
want
to
break
anything
by
default.
So
we
try
to
proxy
everything
as
hard
as
we
can,
but
adding
that
kind
of
policy
around
like
this
is
allowed
and
it's
not
allowed
that's
definitely
on
the
roadmap.
So
I
would
encourage
you
to
file
a
github
issue
on
the
link
ready
to
repo,
and
then
they
critique
a
project
and
tell
us
exactly
what
you
want.
A
No,
it
supports
HTTP
just
as
easily
and
in
right
now
it's
just
on
the
it
is
just
on
the
headers,
so
you
don't
get
the
request
body.
What
is
definitely
coming
up
soon,
because
I
think
this
is
a
super.
Important
feature
is
the
request
body
as
well
and
of
course
there's.
You
know
that
when
you
combine
that
with
TLS
everywhere,
it
becomes
really
really
interesting.
A
It's
probably
more
a
question
of
GRP
seed
than
anything
else.
I
think
what
gr
PC
gives
you.
You
know
I'm
like
a
10%
expert
on
this.
It's
two
things.
What
if
it's
been
touched
on,
which
is
okay,
you
know
you
now
have
using
protobufs,
rather
than
you
know,
kind
of
JSON,
whatever,
which
means
a
you're
kind
of
forced
to
standardize
like
you
have
this
form
or
like
here
is
the
API?
Here's
a
request
that
I
expected
here's
a
response
that
I
expect
you
get
some
type
checking
and
stuff
like
that.
A
In
the
you
know,
in
the
in
the
code,
that's
really
helpful.
It
also
reduces
the
serialization
and
deserialization
cost
dramatically
over
something
like
JSON
and
then
finally,
I
think.
The
other
benefit
is
because
G
RPC
it's
built
on
HTTP,
and
this
is
kind
of
where
Laker
D
starts
coming
in,
because
it's
built
on
HTTP.
You
end
up
with
dramatically
lower
TCP
overhead,
because
h2
is
really
good
at
keeping
one
connection
open
between
any
two
pods.
You
don't
have
to
have
12
connections
or
50
and
multiplexing
requests
that
can
really
reduce
your
tail.
A
Latency
is
especially
when
combined
it
with
things
like
negative
acknowledgement
and,
like
you
can
say,
like
hey
I,
don't
want
this
request
right
now,
but
but
the
implication
like
kind
of
the
side
effect
is
that
layer,
3
layer,
4
load-balancing,
which
is
done
at
the
connection
level,
starts
breaking
down.
So
because
you
know
you
just
maintain
this
one
connection
open
alright
and
we
have
a
blog
post
coming
out
of
that
about
that's
pretty
soon,
but
one
of
the
biggest
drivers-
and
this
might
have
been
the
driver
for
Ben
as
well.
A
B
C
A
What
we
wanted
to
do
first
was
get
telemetry
like
totally
nailed
down
and
then
can
there
get
on
get
on
some
of
the
reliability
features
like
retry,
this
and
circuit
breaking
stuff
like
that,
and
then
after
that,
we'll
probably
add
routing
I.
Think
for
us.
Our
experience
with
link
would
be
one
that
XO
one.
That
ex
you
know
supports
everything.
In
fact,
one
dead
X
has
this
very
powerful,
complicated
language
called
D
tabs.
That's
like
the
fact
tracking
hierarchical.
A
You
know,
kind
of
mapping
of
you
know
concrete
names,
a
semantic
name
and
like
old
stuff,
it's
very,
very
powerful.
You
can
do
all
sorts
of
crazy
routing
stuff.
That's
also
it's
really
difficult
to
understand
and
we
had
to
build
all
these
tools
for
how
you
you
know,
Express
that
and
you
know,
and
how
you
understand
what
your
D
tags
doing.
So
one
thing
that
we
don't
want
to
do
is
have
it's
like
repeat
that
experience
where
you
have
all
this
power,
but
it's
incredibly
hard.
A
So,
that's
all
to
say,
I
would
love
to
you
again,
so
you
get
hub
issues
from
you.
You
are
from
all
of
you
saying
exactly
what
you
want
from
you
know.
If,
in
the
routing
there
is
it
Bluegreen
deploys,
is
it
Canaries?
Is
it
you
know?
How
does
it
compare
to?
You
can
do
Bluegreen
today
with
kubernetes
by
doing
label
flipping?
Why
is
that
not
sufficient
for
your
purposes,
the
more
of
that
kind
of
input
that
we
have
the
easier
it's
going
to
be
for
us
to
make
a
solution?
C
C
B
Yeah
you
get
incoming
and
outgoing
services
for
one
service.
I
think
that
we
there
used
to
be
the
ability
to
see
everything,
that's
connected
everything
at
all
times,
but
that
gets
very
messy
very
quickly,
especially
when
they've
got
a
graphic.
A
well
like
that's
very
big,
to
show
on
page
I
do
know.
There's
a
good
tool
to
do
this
there,
which
is
like
weed
I,
think,
is
probably
about
a
badass.
B
A
And
we,
you
know
we
actually
have
gone
back
and
forth
on
this
a
bit
because
we
did
have
that
default
like
a
full
topology
like
you
know,
you
know,
boxes
and
arrows
graph
that
was
drawn
I.
Think
it's
still
in
there
somewhere
homage
graphs
are
awesome
for
demos,
but
as
soon
as
you
have
40
services,
it
just
looks
like
a
mess.
A
So
what
we've
done-
and
we
kind
of
saw
on
the
demo-
was
if
you're
looking
at
an
individual
service
will
show
you
all
the
incoming
connections
and
they'll
show
you
all
the
outgoing
connections
right
on
a
per
service
basis.
So
we
give
you
this
little
slice.
We
call
the
octopus
graph,
it's
like
there's.
You
know:
here's
your
service
in
the
middle
of
ears,
that
here's
the
arms
and
on
both
sides
and
that's
been
the
most
kind
of
same
view.
C
A
A
C
A
Know
I'm,
probably
not
the
right
person
to
ask
there's
there's
nothing.
That's
super
top
of
my
the
super
high
priority
for
us.
No
one
has
really
asked
us
for
that.
Yet,
if
Simon
did
like
I
guess
we
would
take
a
look.
You
know
why
not
I
think
my
sense
is
most
of
the
most
of
the
users
that
we've
seen
for
whom
linker
D
has
been
valuable.
A
Don't
have
the
the
kind
of
incredibly
high
volumes
of
network
traffic
that
bps
would
make
useful
it's
you
know
most
of
our
adopters
have
been
folks
like
then,
who
does
you
know
who
handles
a
high
volume
of
traffic?
This
is
wildly
popular
app
but
doesn't
need
to,
like
you
know,
stream,
video
or
whatever
so
no
immediate
plans.
But
if
that's
something
that's
really
interesting
to
you,
you
know
we're
certainly
open
to
it.
A
A
That's
always
the
question
and
and
reading
so
yeah.
You
know
the
goals
of
the
two
projects
are
pretty
different
in
a
lot
of
ways:
link
ratibor,
not
access
pretty
similar
to
SEO
is
pretty
similar.
Do
one
dot
X
you
know,
and
the
goal
is
we
want
to
give
you
every
possible
feature
that
you
could
ever
imagine
and
to
us
it
doesn't
matter
what
the
complexity
is.
At
least
this
is
this.
Is
my
I
don't
want
to
I
want
to
speak
for
the
ISTE
Oh
team
right,
but
this
this
is
my
impression.
A
Is
that
for
them
the
cost
and
complexity?
It's
not
that
bad.
It's
more
important
that
have
a
huge
feature
set
for
link
ready.
The
focus
you
know,
certainly
with
$2,
was
very
different
right
where
we
really
care
heavily
about
the
con
about
the
complexity.
You
know
at
the
expense
of
it
not
having
thousand
different
features.
So
I
think
the
two
are
gonna
end
up
kind
of
in
very
different
places
and
I
think
which
one
is
right
for
you
really
depends
on.
A
C
Alrighty,
that's
all
the
questions
that
we
have
time
for
and
thank
you
very
much
for
all
the
attendees
that
asked
questions.
I
think
we
had
a
total
of
23.
We
made
it
through
15,
if
you,
if
there
are
any
that
I
left
dangling
or
we
didn't
get
time
to
answer,
I
urge
you
to
go
over
the
slack
channel
and
ask
them
or
any
other
open
forums
to
ask
those
questions.