►
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
BoF Deep Dive: Chaos Engineering - Sylvain Hellegouarch, ChaosIQ
Can we actually learn in production are we doomed to remain in panic mode? In this session, we will be exploring how Chaos Engineering can help us learn and improve our production system by experimenting against it from various angles. We will be using the Open Source Chaos Toolkit CLI to initially codify our experiments before using the Chaos Platform to schedule them and collaborate on our learnings.
https://sched.co/MPlV
A
Hello
feel
like
I'm,
a
teacher
of
sitting
around.
It's
really
really
disturbing
hi
everyone.
Thank
you
very
much
for
coming
today.
I,
don't
know
if
you're
here
for
the
talk
or
because
you
have
plugs
where
you
can
charge
your
laptops,
but
I'm
gonna
go
for
the
first
one
and
hope
for
the
best
anyway.
Thank
you
very
much
for
being
here.
A
We're
going
to
talk
and
I've
got
35
minutes
and
it's
it's
a
short
very
short
time
to
demo
everything
I'm
going
to
demo.
So
sorry,
if
I
rush
a
bit,
don't
is
it
that
to
come
back
afterwards
and
actually
ask
any
questions
I'd
be
pleased
to
you
know,
figure
out
for
any
any
problems
right,
so
we're
going
to
talk
about
Harrison
during
I.
Don't
know
if
you're
familiar
with
the
term.
A
Suffice
to
say
that
what
we're
going
to
agree
on
today
is
calendaring
is
not
necessary,
break
or
not
break
it's
more
about
collecting
data
to
see
if
your
witness
in
your
system
and
collecting
data
in
an
automated
fashion,
to
see
how
your
tooling,
already
like
a
tracing
in
like
monitoring,
can
help
you
out
figuring
out.
If
you
have
witnesses
in
your
system-
and
you
probably
don't
realize
that
so
we're
going
to
do
a
demo
sweet,
let's
get
rid
of
the
slides
okay.
So
it's
really
tricky
when
you
have
a
you
know.
A
Big
screen
like
that
and
you've
got
so
many
windows
to
show,
but
let's
hope
for
the
best
right.
So
what
what
we
start?
We've
got
a
very,
very
stupid,
simple
service
that
returns
when
we
talk
to
it,
I
think
that's
the
one
that
when
you
talk
to
it
simply
return
the
payload
that
says
you
know
the
version
of
the
service
running
the
timestamp
of
the
call
and
the
number
of
call
that
is
received
so
far.
So
that's
it.
A
That's
just
the
endpoint
that
we're
talking
to
you
I
kept
the
simple
thing,
because
I
wanted
to
focus
on
the
chaos
part,
not
on
the
service
side
right,
so
we've
got
that
service
and
we
see
that
we
have
just
one
pot
for
it's
at
the
top.
Now,
because
we
are
very,
you
know,
modern
everything
we
want
to
go
to
microservice,
you
know
sort
of
architecture
and
we
decide
that's
where
it
shouldn't
be,
actually
the
one
generating
the
number
it
should
be
talking
to
another
service
somewhere.
A
However,
it
does
it
and
return
the
value,
so
we're
going
to
have
a
second
service
doing
that
that
we're
going
to
deploy.
You
know
you
know
you
know
and
on
our
side
here
on
the
on
the
left
side,
so
service
to
now
is
actually
generating
data
for
us
and
we're
going
to
see
to
see
that
and
we
sing
on
the
on
the
bottom
right
that
service
to
is
responding
and
generating
a
number.
But
as
so
far
we
see
service,
one
is
not
still
talking
to
service
two.
It's
still
generating
its
own!
A
A
That
is,
you
know,
changing
the
image
of
a
pod
somewhere,
so
in
this
very
simple
set
up
we're
going
to
actually
do
that
and
we're
going
to
have.
Let
me
see
yes,
so
I'm
gonna
increase
a
bit
this
one
again.
Yeah
I
should
be
fine
right.
So
we're
going
to
see
what
that
can
look
like,
but
I'm
going
to
explain.
First,
one
thing
here
is
enduring:
if
for
fo
lot
of
people
is
we
were
just
discussing,
that
is
quite
often
just
kills
monkey
you
run
break
stuff
and
uip.
A
Well,
that's
one
way
of
doing
it,
no
doubt,
but
it's
more
protocol.
If
you
break
something
or
you
change
something
in
your
system
without
even
breaking
it,
but
you
don't
have
an
idea
of
what
you
need
to
collect
to
analyze
after
the
fact
you'll
know
better
than
if
you
didn't
break
anything
because
you
don't
know
what
to
look
for
you
don't
know
what
to
read
in
your
system.
A
At
the
end
of
the
day,
we
have
a
lot
of
data
that
we
can
generate
a
report
and
analyze
from,
but
that's
that
you
still
haven't
actually
checked.
If
your
issue
in
there
in
the
right,
the
whole
idea
of
kills
and
wrong
is
to
start
from
an
AI
possesses
your
question:
if
I
do
this,
if
I
do
a
rollout
of
a
version
that
is,
is
safe,
what
should
I
see?
I
should
expect
that
you
know
service,
one
is
deployed
and
service
to
it.
No
sorry
Swiss
deployed
a
service.
A
One
is
happy
bunny,
but
what,
if
trampled
you,
you
think?
Ok,
what
happens
if
I
release
a
new
version?
That
is
an
Elfie.
You
probably
don't
know.
What,
if
you
release
something
that
has
latency
now
answer
is
to
is
responding.
It's
all!
Fine!
Your
testing
was
fine,
you,
it's
all
Berean,
you
know
functionally,
it's
is
working,
but
maybe
your
system
is
under
C
for
it
because
because
it's
slower
and
you
don't
necessarily
know
how
to
service
one
or
fan
service,
one
is
expecting
your
response
for
service
too.
A
It's
not
happy,
so
you
see
just
that
little
command
of
rolling
out
a
new
version
already
starts
expanding
into
lots
of
questions
and
those
questions
should
be
turned
into
experiments.
And,
honestly,
they
just
look
like
a
bunch
of
declaration
that
we
am
going
to
demo
here
that
you
can
run
automatically.
So
what
does
this
look
like?
Alright,
we're
going
to
go
for
the
method.
It's
a
bit.
You
know
over
the
top,
perhaps
the
naming,
but
the
method
is
basically
a
bunch
of
actions
or
probe
that
you're
going
to
run
to
change
the
system.
A
So
we
are
we
going
to
say
and
I'm
going
to
show
you
one.
It's
a
probe,
saying
well
I'm,
going
to
run
its
Python
thing
running.
In
the
background,
doesn't
really
matter
so
we're
going
to
run
that
thing
that
is
called
set
deployment
image
and
we
say
for
the
deployment
service
to
release
version
1
and
that's
going
to
be
our
new
lab
purchases.
I'll
explain
that
later,
what
we'll
do
in
half
that
is
we're
going
to
run
some
blood
tests
doesn't
have
to
be
high
impact,
just
something
that
you
can't
really
run
against
your
system.
A
So
here
this
one
is
going
to
simply
talk
to
the
stamp
every
free.
Second
for
20
seconds.
Don't
you
know
that
at
the
end
here
we
we
have
a
probe
that
probably
simply
collecting
data
here,
we're
going
to
look
at
the
kubernetes
pod
status
or
deployment
status
actually
and
tell
us
that
will
tell
us
whether
or
not
we
actually
shift
shifted
or
not
change
the
image.
So
all
the
data
is
there,
but
you
need
to
collect
it
as
we
run.
A
Otherwise
you
can
actually
compare
or
or
correlate
sorry
I
might,
if
I
remember
the
command
better
you're
welcome.
So
we
do
that
and
doesn't
really
matter
that
looks
basically
it
so
semi.
It's
always
the
same
idea.
Either
you
run
a
Python
function
or
you
run
a
process
or
you
run
an
HTTP
endpoint.
It
doesn't
really
matter,
you
could
go.
You
know
all
the
things
she
finally
says.
You
run
that
secretly
and
you
look
at
what
what
happens,
but
now.
A
What
you
do
is
you
define
a
set
of
probes
again
and
you're,
going
to
say
I'm
going
to
ask
my
system
for
that,
and
that
system
is
going
to
just
simply
talk
to
my
monitoring
tool
and
say:
did
I
see
any
500
errors
from
the
service
one
now,
usually,
what
you
would
want
here
is
not
something
too
technical.
That's
a
demo!
That's
a
bit
low
level!
You
would
want
your
SSL
over
there
because
you
really
want
to
know
from
Pearson
during
I've
an
impacted.
A
My
users
have
I,
actually
hurt
my
business
and
I
mean
not
in
a
very
friendly
sense.
Have
I
heard
my
you
know
my
business
by
doing
this,
you
know
already
by
doing
testing
that
from
a
technical
bit
you
know
you
can
roll
out
that
you
know
that's
all
good,
but
sometimes
you
forget
about
looking
from
the
know
very
far
away
and
see
how
your
users
are
impacted
by
doing
that.
So
usually
your
your
things,
they're,
your
probes,
are
simply
know
I'm
going
to
talk
to
my
endpoint,
something
in
my
system.
A
That
tells
us
tell
me,
am
I
doing
right?
Am
I
looking
good
for
my
users?
Do
you
collect
data
in
you?
What
you
have,
in
addition,
is
a
tolerance.
That's
simply,
this
one
is
going
to
go
through
the
JSON
return
and
look
for
a
bit
of
data.
That's
it
and
this
one
is
going
to
save
for
the
tensile
last
ten
seconds.
I
didn't
see
any.
You
know
500
now.
A
That
thing
is
interesting
in
itself,
but
it
runs
twice
and
that's
what
makes
it
very
interesting,
overall
focus
and
ring
what
it
does
is
when
you
start
running
it
will
run
the
first
time
before
the
method
and
said
before
I
do
any
action.
Am
I
actually
already
doing
fine,
because
if
I'm
not
there
is
not
no
point
in
continuing
I
already
know,
I'm
broken
or
in
a
state
that
I
can
actually
read
after
the
fact,
so
it
will
Bailey
on
the
experiment
immediately.
A
The
second
time
it
runs
is
after
the
method
and
basically
it
will
tell
you-
have
I
deviated
if
I
see
500
errors.
Now,
admittedly,
you
know
you
could
have
errors
coming
up
for
all
sorts
of
other
reasons,
but
for
the
purpose
of
this
demo
we
know
that
it's
the
only
way
you
could
have.
That
is
because
what
we've
done
right,
if
you've
got
errors,
then
you
know
that
when
you
do
that
specific
action,
when
you
roll
out
that
version
or
a
specific
event,
you
know
in
in
some
and
some
conditions-
you've
broken
broken.
A
It's
not
a
really
the
right
word,
but
you've
you
haven't
kept
your
promise
towards
the
user,
so
cure.
Centering
is
really
to
collect
data,
collect
evidence
that
you're
doing
fine.
It
doesn't
say
the
bricks
are
doing
fine.
You
may
have
actually
a
suicide
broken
somewhere.
It
doesn't
really
matter
what
matters
is
all
it's.
Your
system
is
how
use
your
users
perceive
your
system
doesn't
really
matter
if
you
have
actually
lost.
You
know
half
of
your
system
if
you
can
still
provide
as
sub
service
right,
let's
say
my
Amazon
or
something
like
that.
A
Be
that
okay,
if
I
people
come
by
but
can
still
search
for
product
and
put
that
in
the
basket,
am
I
still
doing
good.
So
that's
the
sort
of
question
you're
looking
for
in
your
system
and
calendaring
is
there
to
collect
data
towards
finding
events
and
sirs
toward
that
right
quickly,
because
I,
don't
want
goes
straight
away
to
that
quickly.
Here
is
it
right?
A
We
have
got
rollbacks
rollbacks
I
just
want
to
go
quickly
on
that
rollbacks
and
not
about
setting,
at
least
in
your
tool,
kit
and
caring
not
to
put
back
your
system
where
it
should
be.
It
can't
know
that,
because,
if
you've
discovered
something
you
didn't
know,
how
could
you
actually
prevented
that
anyway
or
put
it
back
right?
What
you
do
is
release
you
undo
what
you've
done
so
in
this
one
we'd
simply
do
a
rollback
of
the
version
we
rolled
out
right,
that's
it!
So
how
does
this
look
right
here?
We've
got
now.
A
We've
got
service
one
service
to
distill,
Epiphanes
yeah.
So
if
you
need
to
move
service
one
to
service
to
actually
right,
so
we're
going
to
change
the
image
of
this
one.
So
if
that
that's
a
manual
our
rollout,
so
what
we
do
here,
if
you
look
at
this
one,
we
see
with
changing
that
version
so
that
certainly
this
one
will
match
this
one,
and
we
saying
basically
version
one
should
be
talking
to
service.
One
should
be
talking
to
service
to
some
obviously
demo
goodies
with
me
yeah.
A
So
now
we
know
that
our
service,
one,
the
one
that
users
see,
even
though
they
don't
know
it's
called
service
one.
But
what
people
talk
to
is
talking
to
the
sponsor
is
not
yeah.
We've
got
to
microservice
architecture.
So
what
is
it
brittle?
Okay,
what
are
we
going
to
do
here
is
going
to
run
that
thing
that
I
showed
you
the
experiment
in
a
chaos
engineering
way
and
the
first
one
is
we're
going
to
say
we're
going
to
roll
out
the
same
version
here
that
we
already
have
running
now.
A
What
would
Cuba
needs
to
do
for
that?
Well,
it
won't
actually
do
anything,
because
the
image
is
the
same.
So
nothing
happens.
That's
all
new
light
possesses
says
that
so
baseline.
If
already
something
was
broken
there,
either
communities
was
in
trouble,
I
wouldn't
bet
on
it
all.
We
would
have
actually
a
problem
with
a
setup.
So
in
this
case,
what
we
see
is
simply
saying:
okay,
we
knew
were
already
on
version
one.
So
Cuba
entities
didn't
do
anything.
There
was
not
changed
so
to
speak,
it's
good
it!
Well
it
well,
it
went
through
now.
A
Let
can
we
actually
do
a
roll
out
of
a
version
to
which
further
the
purpose
of
this
demo
is
a
version
that
is
supposed
to
work.
So
the
the
service,
it's
exactly
the
same
is
responding
the
same
and
it
won't
actually
break
in
theory,
it
shouldn't
break
the
the
service
that
users
are
talking
to,
because
service
to
is
actually
Sun
and
working
fine.
So
we
all
see
that
the
pod
have
been
flipped,
and
now
we
took
in
to
version
two
and
we
see
those
things
but
already
what's
interesting
here,
one
we
checked
for
five
hundreds.
A
We
could,
we
could
argue,
we
broke
users
promises
because
we
reset
the
cognitive
value
so
already
either
we
decide
that
our
kills
experiment
needs
to
account
for
that,
and
so
that's
part
of
our
SLO
or
SLA
with
permissive
value.
That
is
complimented
not
going
back
in
in
time
right.
So
you
could
argue,
that's
a
broken
thing,
even
though
all
services
have
been
aptly
switched
over,
but
for
the
for
the
purpose.
Here,
we're
not
going
to
deal
with
that,
but
that's
a
question.
You
would
ask
yourself,
and
maybe
you
wouldn't
have
she
didn't.
A
Do
you
know
chaos
experiment?
Now,
let's
go
for
service
I'm
going
fast
here,
just
because
I
want
you
to
showcase.
That's
not
very
interesting.
It's
the
same
experiment
with
a
variety
of
images
that
are
free,
can't
that
do
something
I
wanted
Phil
demo
on
that
version.
Three,
it's
a
same
service,
but
now
the
else
probe
is
going
to
return
an
error
immediately.
A
So
what
we
see
here
and
we're
going
to
be
able
to
see
that
yeah
I
think
is,
we've
got
two
parts
and
we've
we
see
version
three,
that's
good,
but
we
see
we
don't
we.
We
were
reverted
through
the
rollback
to
version
1,
and
normally
we
shouldn't
actually
go
to
version
3,
because
PB
only
fees
won't
allow
it,
because
the
probe
is
is
saying
I'm,
not
you
know
in
a
good
good
position.
I
can't
allow
you
to
actually
go
through
that's
good.
At
least
we
know.
Now
we
have
evidence
at
QB.
A
Aunties
does
the
right
thing
if
you've
set
up
your
else
probe
and
readiness
probe,
you
know
accordingly
now
what
if
and
that's
the
one
that
is
interesting
because
so
far
you've
got
seen
on
in
green.
So
you
like
its
effect,
the
effect
data
right.
So
what
if
we
now
release
a
version
that
else
probe
is
going
to
be
fine,
but
it's
a
little
slower
so
when
service
one
talks
to
it,
so
it's
too
is
too
slow
to
respond.
What's
going
to
happen,
so
in
theory
we
should
see
ya
version.
Four
has
been
flipped
here.
A
It
should
come
back
in
a
minute.
Yep
we've
switched
to
the
saw
before
and
we
see
internal
server
purely
because
my
serious
want
to
deal
with
it.
I
have
no
absolutely
no
error
check
for
that,
so
it
broke
it
breaks.
So
at
least
we
know
now
we
do
have
evidence
that
when
we
really
is
a
slower
version
of
version
of
service
tool,
we've
actually
broken
a
promise
to
your
user.
Even
though
I'm
sure
your
test
would
say
you
know
it's
all
fine
I
mean
there
was
nothing
broken
of
service.
It
was
working
right
did
respond.
A
Yes,
no,
but
that's
fine!
So
here
with
Cassandra,
you
simply
looked
at
the
system
overall
and
decided
to
say:
okay,
how
am
I
impacting
my
users
throughout
instead
of
looking
at
the
discrete
block
and
maybe
through
some
testing
some
values
in
a
violent
of
testing?
You
would
have
looked
at
that,
but
the
I,
like
this
approach
of
chaos,
engineering
and
saying
I've,
got
a
question.
I
have
no
idea
what
I'm
going
to
actually
find
out,
but
I'm
going
to
ask
anyway.
A
Now
that
says:
that's
good
you've
experimented
that
and
we
roll
back
and
service
is
fine,
so
at
least
we've
proved
as
well.
Roll
back
does
work
when
I
think
you
know
things
matter
an
eye
that
stage
we
are
in
banking
system.
So
it's
pretty
good
because
we
have
covered
with
a
very
simple
experiment:
we've
covered
quite
a
few
scenarios.
We
proved
that
we
can
roll
out
the
same
version.
A
No
impact
can
prove
we
can
roll
out
a
new
version
that
works
sort
of
no,
depending
on
how
we
look
at
it,
either
no
impact
or
actually
proper
impact
because
we
reset
it
the
VCG
the
value.
In
that
case,
you
could
argue,
then
maybe
in
its
Upstate
and
that
sort
of
thing,
but
that's
another
discussion,
version
three.
When
the
probe
says
I'm,
not
MC,
fine,
you
don't
change
anything
and
version.
Four
is
functioning
but
to
slow
doesn't
actually,
you
know,
work
within
the
boundaries
of
expectations
and
we
broke
a
promise.
A
A
All
right,
so
we've
generated
a
report
from
those
two
slugs,
and
what
we
have
here
is
a
PDF
report
could
generate
HTML
that
basically
tells
you.
How
do
you
dealt
so
I'm
going
not
going
to
go
through
those
those
diagrams
because
doesn't
only
matter
for
now,
but
what's
interesting
is
we
do
see
if
I
come
back
to
you?
Sorry
yeah
this
one,
the
first
version
we
see,
we
didn't
impact
or
users
encounter
that
we
proved
we
collected
generated
something
gradually.
A
A
Now,
we've
got
to
the
other
one
where
we
simply
change
version
working.
We
see
the
drop
all
right.
So
that's
interesting
because
you
are,
you
could
really
argue
you
know.
Am
I
doing
right
when
that
happens
am
I
actually
breaking
my
system
and
my
users
or
good
this
one
is
another
one
that
is
version.
3
will
remember
the
else.
Prob
was,
you
know,
not
happy.
A
We
didn't
change
anything
fine
and
it's
interesting
because
in
this
one,
if
you
know
again
when,
when
you
see
that
that
thing
here,
it
really
tells
humanity
tells
you
I
tried
to
actually
roll
out
what
I
couldn't
I
try
to
switch,
but
I
couldn't
so
I
ended
up.
You
know
not
doing
it.
That's
that's
interesting
to
look
at
if
you
have
money
trying
to
by
the
way
and
the
final
one
is
obviously
the
one
where
you
know
we
remember
it
was
slow.
A
So
we
see
at
some
point
we
don't
actually
get
a
new,
a
new
value
because
we
can't
sew
through
very,
very
simple
experiments.
You've
learnt
a
lot
about
your
system
and
simple
things
like
rolling
out
a
new
version
and
while
I
was
doing
the
demo,
I
realized
I
learned
a
lot,
but
you
know
things
on
communities.
I
didn't
know
about
yeah,
read
them
on.
Books
actually
saw
the
documentation
with
finisher.
A
Yes,
I,
don't
know,
but
yeah
what's
interesting
here
is
trying
to
really
think
about
I.
Think
I
know
when
you
said
to
your
colleague,
ya
know:
what's
going
to
happen,
you
the
verb
here
is
actually
I
assume
I.
My
my
own
experience
or
I've
done
that
manually
somehow
is
telling
me
this
is
going
to
happen,
but
you
have
any
proof
for
that.
Then
you
have
evidence
or
think
the
other
way.
A
Sometimes
you
know
something
is
broken
and
you
want
to
actually
achieve
something
or
change
your
you
know
make
sure
that
it's
changed
and
you
want
to
go
to
talk
to
your
manager.
I
have
data
for
that.
That's
going
to
be
more
than
you
know,
I
think
it's
broken.
So
that's
the
thing
sort
of
thing
you
want
to
do
with
chaos
and
right
now
before
we,
the
the
talk
is
taught
least,
it's
called
from
one
to
many.
So
if
I've
done
everything
on
my
own
on
my
laptop
and
everything,
it
doesn't
have
to
be.
A
A
So
if
you
have
the
right
header
and
think
that
Trace
that
you're
running
from
Cassandra,
you
can
decide
to
have
a
different
behavior
within
your
application
or
you
know
alerting
you
could
decide
not
to
actually
page
everyone
when
you
know
it's
a
chaos,
engineering
experiment
that
has
broken
something
right.
So
really.
The
idea
here
is
to
say
from
one
to
many,
not
just
many
experiments,
but
many
as
I'm
not
running
that
on
my
own.
There
is
no
one
chaos
engineering
in
your
organization,
you're,
all
chaos
engineering,
because
you
all
care
for
reliability
resiliency.
A
What
matters
is
that
you
are
visible
when
you're
doing
it.
That
really
is
important.
So
that's
the
sort
of
thing
that
kills
toolkit
the
open
source
tool
actually
will
provide
you.
It
gives
you
a
protocol
to
define
your
case
engineering
experiment,
but
it
also
allows
you
to
talk
to
you
the
rest
of
the
that's
quite
interesting
and
that's
quite
powerful.
Just
one
thing
here
and
you'll
notice,
and
not
in
this
one.
A
Unfortunately,
but
there
was
one
where
it
went
from
not
working
to
work
in
to
not
working,
which
shows
as
well
how
the
rollout
happens
and
how
actually
community
switched
from
one
service
to
the
other
from
a
network
perspective.
So
you
learn
again
a
lot
by
doing
so
as
a
sort
of
thing
and
when
you
know
when
things
work,
fine,
it's
all
good
and
you
you
know
it's
basically
all
good.
All
the
the
calls
that
you
make
in
the
background
to
the
system.
You
know
sure
that
they
respond
very
quickly.
So
it's
really
good.
A
Now
again,
like
I,
said:
I
run
that
on
my
machine.
Probably
you
would
want
to
run
that
in
your
CI
CD,
although
don't
necessarily
run
that
ty
you're
catering
to
your
deployment
run
that
part
you
know
in
orthogonally.
Sometimes
you
want
to
roll
out
something
and
you
want
to
run
some
aspects
of
care
centering,
but
you
should
always
ask
us,
and
during
that
runs
continuously
constantly,
because
the
prime
is
doing
with
role
at.
Is
your
system
leaves
between
your
deployments,
a
divorce
right,
so
keep
changing
the
system,
as
you
run.
A
A
Yeah,
that
would
be
you
can
create
a
cube
job
from
job
or
whatever,
something
as
simple
at
that
and
I'm
hopeful.
This
is
going
to
work
right,
so
we
see
that
we've
got
a
job
now
and
that
job
runs
the
exact
time
and
I
ran
on
the
console
and
tries
to
create
survey
3.
We
remember
that
see
an
LC
one.
So
know
me.
Normally
this
one
here
she
didn't
change,
and
that
shows
us
basically
what
how
you
automate
that
so
we
went
from
doing
things
manually,
not
even
actually
thinking
about
doing
them.
We
know.
A
Don't
we
to
encoding
through
protocol
the
experiment
of
akio-san
during
you
know,
experiment
saying
we're
going
to
change
that
we're
going
to
collect
that
data,
that's
the
one
that
matters
to
us
to
make
sense
of
what
we're
doing
and
we're
going
to
see.
If
we
deviate
it,
when
we
did
it
right
toward
the
end
where
we
know
to
make
that
which
cube
job.
A
But,
however,
you
want
to
run
them,
you
know,
create
a
document,
and
that
runs
and
to
the
way
that
I
run
on
my
own
to
the
fact
that
I
can
actually
talk
to
everyone
through.
You
know:
Jaeger
Prometheus.
What
have
you
you
know
that
sort
of
thing
you
can
send
logs
through
your
cloud.
You
know
your
logs
provider
whatever,
but
be
visible
about
it.
That's
the
key
Luke.
It
is
a
platform
in
a
sense
where
it
integrates
nicely
with
what
you
have
instead
of
running
in
parallel,
so
chaos
I'm
doing
should
be
really
about
that.
A
So
next
time
you
think
about
chaos
and
rain.
Don't
worry
if
you,
if
you
use
words
like
braking
or
you
know,
yeah,
braking,
basically
or
killing
machines,
opening
notes,
even
though
it's
little
weird
because
it's
a
bit,
you
know
dangerous,
but
the
idea
is
really
think
that
you
need
to
be
visible.
You
need
to
talk
to
the
whole
team.
You
need
to
be
part
of
the
general
pipeline
of
healthiness
and
resiliency
and
reliability,
and
at
that
stage
I
think
that's
going
to
be
all
I.
A
Don't
want
to
go
a
little
too
far
down
the
road,
because
there
are
plenty
of
other
things
on
the
kill
to
kill,
but
that
would
take
a
long
way
out
of
that.
So
yeah
I
went
fast
because
I
really
wanted
to
show
you
all
the
starting
from
nothing
to
all
the
way
to
actually
having
the
experiment
running
and
seeing
you
know
all
the
various
experiments,
but
take
some
time
trying
to
drain
the
whole
circuit.
A
You
know
slack
or
simply
read
the
docs
and
you,
hopefully
that
will
make
more
sense
if,
if
I
went
too
fast,
yeah
I
think
I'm
good
I
want
to
teach
you
to
go
first,
so
that
you
have
time.
If
you
wanted
to
ask
questions
about,
kills
injury
in
general,
known
as
the
disease
demo
I
had
some
something
to
ask
about.
So
for
that
I
should
say.
Thank
you
for
attending
and
and
then
following
and
don't
is
it
that
if
you
have
any
question
that
be
able
to
take
them.
A
A
Well,
full
disclaimer
I
wrote
to
kill
school
kids,
so
I'm
gonna
say
yes,
it's
a
vegetable.
No,
it's
it's!
It's
a
different
tool.
Motul
that
you
have
you
see
are
there
are
about
40
injection
and
they
are
necessary.
Close
turkey
doesn't
do
the
40
injection
itself.
It's
actually
using
that
it's
it
orchestrates
those
tools
to
inject
fault
and
collect
data
as
it
is
so
the
KO
circuit
is
going
to
be
the
one
giving
you
the
protocol
for
that,
an
experiment
that
you
can
share,
but
it
really
depends
how
you
want
to
see
it
trampled.