►
From YouTube: Policies and Telemetry WG - 2018 04 18
Description
Policies and Telemetry Working Group Discussion from April 18, 2018.
B
A
B
A
A
C
D
A
A
C
C
C
That
come
to
close
tonight,
okay,
is
anyone
there.
They
can
hear
every
yeah,
okay,
all
right
so
saying
that
we
need
to
be
able
to
extend
to
modern
the
climate
server
in
the
future,
because
the
Gateway,
the
reason
grace,
is
all
kind
of
proxies
that
might
potentially
come
up
so
having
a
string,
give
them
more
flexible
option,
and
then
we
also
propose
that
the
ID
type,
a
yeast
conducts
a
port
ID
thing
for
the
actual
name
of
the
proxy.
C
So
we
can
distinguish
between
different
kinds
of
different
servers
and
there's
a
reason
result
issue
there
with
the
names
for
VMs,
because
we
don't
know
how
to
name
them.
That
does
mean
in
the
ship
issue
in
the
past
I.
Don't
think
you
know
still
how
the
two
can
work
so
I
don't
think
it's
going
to
be
it's
an
it's,
not
just
an
issue
in
this
document.
It's
over
all
of
them.
A
E
The
reporter
ID
and
and
type
so
many
many
times.
The
thing
that
you
finally
want
to
derive
is
whether
it
was
reported
clients
either
server-side
or
intermediate.
So
I
also
look
II.
If
we
know
the
reporter,
then
we
have
to
compare
it
with
the
actual
payload
to
know
if
the
reporter
is
the
client
or
the
server
right
it.
That's
that's
the
general
idea
right,
yeah.
C
F
C
E
Get
the
other
kind
of
relief
party
client-side
telemetry
is
that
you
have
to
decide
if
we
keep
both
those
reports
open
by
default
right
I
did
it.
It
would
seem
that
a
more
reasonable
default
is
report,
telemetry,
client
side
and
do
policy
checks,
server,
side
and
users
can,
of
course
change
things,
but
I
think
we
we
have
to.
We
have
to
think
about
that
in
life
with
that.
E
D
E
It's
what
1xrp
see
but
we're
compressing
it
on
every
call
and
then
mixer
decompresses
the
whole
thing.
So
so
we
pay
the
cost
somewhere.
It's
not
continuous,
it's
in
bits
and
pieces.
The
person
is
doing
it.
Both
climate
oversight
doubles
that
cost
and
the
question
is:
is
it
reasonable
to
just
keep
that
as
the
default
or
dude
client-side
yeah
I?
Guess
it
comes
out
telemetry,
and
so
when
people
can
turn
both
of
them
on
if
they
want
to.
D
Yeah
I
think
it's
worth
looking
at.
This
said
that,
with
the
lens
of
the,
the
config
model
are
going
to
have
in
a
few
months
where
we
have
will
have
service
config
that
describes
the
behavior
or
everything
that's
coming
into
a
service
and
we'll
have
consumer
config
to
describe
what
happens
on
the
way
out
and.
D
A
Yeah
yeah
I
mean
so
I'm
worried
if
we
turn
on
only
one
or
the
other,
like
about
the
config,
that's
written,
assuming
both
they're
getting
to
be
there
right.
Like
do,
we
start
losing
server
side
like
server
requests
accounts,
so
we
depend
on
now,
or
you
know
like
what.
How
do
you
write?
Config
is
flexible,
based
on
which
thing
you've
titled.
E
On
I
think
yeah
I
think
that
is
exactly
what
needs
to
be
thought
through,
but
because,
just
by
just
by
saying
that
you're
turning
this
one
little
bit
on
you're,
doubling
its
collection
and
the
differential
between
those
two
is
I
mean.
We
know
that
there
is
a
difference,
but
it's
very
small.
So
is
that
is
holding
the
data
justified?
E
D
D
D
E
Many
times
client
is,
in
my
view,
more
useful,
because
if
all
the
clients
of
the
service
are
having
poor
performance,
then
the
fact
that
the
service
itself
says
but
I'm
like
I'm
great
he's.
Well,
it's
still
a
it's
still
a
good
debugging
tool.
But
what
you
really
want
is
how
it
is
perceived
by
its
clients.
G
D
E
It's
I
think,
that's
I,
think
that's
our
fallen
low
because
the
client
may
even
go
up
because
of
network
or
because
the
client
is
sending
the
data
in
a
particularly
bad
way
to
the
server.
That's
why
only
latency
perceived
by
that
client
has
gone
up,
but
because
now
the
server
is
doing
more
work.
Just
on
on
that
back
part,
so
I
think
ownership.
If
we
can
keep
it
completely
separate
that
would
this
will
be
discipler.
D
A
D
E
A
D
A
I
well
so
I
would
argue
that
if
you
take
a
server-side
metric,
you
can
know
who's
calling
you
and
you
can
look
at
from
that
caller.
The
request
that
you
receive
right
but,
like
you,
don't
want
to
see
faults
injected
on
the
client
side
for
your
service
right.
That's
not
something
that
you
care
about
in
some
sense
right,
I,
don't.
D
Also,
I
think
what
you
said
before,
though,
in
terms
of
a
scope
or
some
kind
of
sphere
of
responsibility
as
the
producer.
If
all
of
my
metrics
look
good
and
in
one
particular
flight
is
having
trouble
talking
to
me
and
it
doesn't
show
up
in
my
metrics,
then
their
problem
and
I
can't
do
anything
about
it.
So
I
like
then
I
I,
don't
care
about
their
metrics.
It's
their
problem,
so
yeah
I
kind
of
like
that
answer,
because
it
simplifies
the
data
model.
Quite
a
bit.
G
C
D
If
I
want
to
see
what
my
clients
are
are
doing,
I
would
need
new
attributes
that
that
is
somehow
client
related
attributes
that
I
can
see
and
apply
policies
against
and
log
and
and
while
now,
I've
I've
and
what's
the
expression
across
the
Rubicon
or
whatever
I
have
I
now
have
access
to
the
client
data.
If,
if
I
don't
have
these
attributes
and
I
can't
see
client
data,
then
there's
no
good
place.
There's
no.
We
don't
have
a
console
for
the
mesh
operator.
It's
effectively.
We
have.
We
have
service
views.
D
F
A
E
E
It
is
subject
to
some
refactoring,
but
we
decided
to
put
it
in
because
it
was
kind
of
blocking
the
other
PR,
and
then
there
is
a
API
PR
that
was
much
today,
which
just
flashes
out
the
value
types.
So
previously
value
types
were
used
as
a
just
as
a
metadata
or
as
a
marker,
and
then
code
gen
translated
that
to
actual
go
types
directory,
but
now
these
are
going
to
go
over
wire.
So
now
these
are
actually
concrete
types.
So
the
the
values
are
now
concrete
types
and
they
actually
contain
data.
H
So
I
have
a
question
related
to
the
auto
profit
after
work,
so
you
guys
can
hear
me
yes,
so
earlier,
the
the
adaptors
and
templates
were
not
namespaced,
so
they
were
compiled
into
into
mixer
and
there
was
only
one
adapter
with
name
from
me:
here's
a
one
at
one
tempted
with
named
metric
and
now
in
the
outer
profit
after
work.
We
are,
we
are
doing
this
work
where
this
information
of
water
adapters
mixer,
can
talk
to
our
templates.
It
supports
our
coming
through
custom
resources.
H
So
do
we
want
them
to
be
namespaced,
or
should
they
be?
There
can
only
be
one
template
called
metric
in
this
in
the
universe
of
mixer,
or
there
can
only
be
one
adapter
called
parameters
or,
or
that
or
with
the
other
option
is
we
have
matrix
inside
namespaces,
do
default
and
then
other
matrix
inside
some
of
the
namespace
or
Prometheus
for
that
matter.
So
what
it
means
is,
if
we're
having
a
namespace,
then
every
instance
config
that
the
operator
is,
writing
will
have
to
reference
to
the
right
template
names
of
when
I'm
writing
a
request.
H
A
H
So
templates
were
so
the
yes.
If
you
want
to
namespace
them,
then
there
is
going
to
be
a
custom
resource
definition
for
adding
templates.
So
you
can
say
the
kind
is
template,
and
here
is
the
base64
encoding
of
its
definition
and
its
name
is
metric
and
same
for
adapters.
You
can
say
the
kind
is
an
adapter,
and
here
is
the
name
of
the
adapter
called
Prometheus,
and
here
is
the
base64
encoding
office.
Adapter
conflate,
as
well
as
the
basics,
depend
coding
of
the
templates
I
support.
D
H
H
Mixer
mixer.
No,
so
there
are
two
or
three
C
IDs
that
are
defined
by
a
mixer
framework.
One
is
a
adaptor
C
or
D,
and
a
template
C
or
D,
and
the
operator
not
the
operator
but
with
the
with
the
build
of
mixer
itself
will
be
defined,
will
be
having.
So
if
there
are
four
building
templates,
we
can
stop.
We
have
some
custom
resources
referencing
to
the
the
CR
e
you
can.
You
can
call
it
default
configuration
before.
D
E
E
Not
no,
this,
like
a
command-line
tool
that
that
we
will
that
we
will
provide,
will
brought
some
scripts
or
something
like
that,
this
part
of
their
build
right.
They
need
to
build
their
stuff.
Can
it
produce
the
transitive
closure
and
then
we'll
give
them
a
little
tool
that
packages
it
as
a
CRT
and
then
that
they
can
shove
it
in
yeah.
E
H
So
so
Mandarin
I
talked
about
this
thing,
so
we
can
as
an
air
for
the
adapter
developer.
There
will
be
a
weekend
prom,
and
this
weekend
be
a
doctor
doctor
or
operator
mean
mainly
for
that.
Whatever
you
say,
you
run
this
tool
and
here
is
my
photo,
which
defines
my
operator
adapter
config,
and
here
is
the
proto
for
the
templates
and
that
to
eventually
just
duns
the
transitive
closure,
every
the
combines,
everything
and
finally
gives
you
a
gamma
that
you
can
stick
on
your
website.
D
E
E
We
have
an
option
they're,
equally
unique
name
for
a
template,
or
does
templates
name
also
depend
on
its
anarchy
in
your
particular
config.
So,
for
example,
if
one
of
my
templates
is
like
well
that
I
am
got
something
right
it
in
in
some
ways
it
would
make
sense
for
that
to
have
a
globally
unique
name
right
because
I
like
it,
it's
the
way
that
you
talk
to
Google
I
in
it
shouldn't
depend
on
very
place
that
please
the
most
in
working
together
and.
D
E
Think
I
think
templates
is.
We
could
definitely
start
off
with
the
position
and
the
template
like
adding
a
template
is
a
mesh
level
thing
right,
because
you're
actually
adding
a
whole
new
functionality
itself,
so
the
mesh
operator
gets
to
do
that.
But
the
question
still
remains
that
the
name
is
that
the
name
of
the
resource,
or
is
that
the
name
that's
claimed
by
the
template?
Okay,.
D
E
So
again
my
example
of
like
Google
a
tie
and,
let's
see
that's
the
template
yeah,
which
one
of
those
things
would
be
that
because
it
shouldn't
no
matter
whether
it's
in
Istria
names
any
other
namespace,
the
template
is
called
Bhuvan
will.I.am
and
any
other
configuration
that
that
any
operator
is
going
to
write
is
going
to
be
against
the
Google.
What
I
am.
H
So
are
the
currently.
What
the
model
we
have
is.
The
name
of
the
template
is
what
you
define
in
your
profile.
So
it's
the
package
name,
the
last
segment
of
the
package
name
right
so
either
we
ditch
that
concept,
and
always
the
name
of
the
template
is
the
one
that
the
custom
resource
has
claimed
in
its
engines
in
its
metadata.
Or
we
say
we
don't
really
care
about
what
name
you
have
specified
in
the
metadata.
D
Nope,
look
if,
if
Oliver,
if
how
the
the
stuff
is
in
one
name,
one
name
space
all
the
time
they
see
our
these
are
in
a
single
a
space,
then
there's
no
there's
no
problem.
There's
not
going
to
be
any
I
want
to
go
is
to
say
that
look.
If
you
have
templates
that
are
in
multiple
namespaces,
a
namespace
is
what's
supposed
to
identify
them.
If
we
have
this
all
through
your
name,
that
needs
to
be
unique
across
namespaces,
you
just
can't
broke
the
resource
model.
D
D
D
D
E
There
I
want
to
make
sure
that,
when
someone
tries
to
do
this,
go
through
a
whole
bunch
of
like
replace
stuff
just
to
get
it
get
it
to
work
right,
you
you!
So
if
it's
just
one
place
that
then
I
think
I
think
it's
okay
right,
but
essentially
like
we
define
a
template,
change
the
name,
but
any
other
configuration.
That's
based
on
the
template
right
that
that
is
in
the
examples
will
have
to
change
once
we
once
they
start
using
it.
E
H
So
so
are
we
saying
the
name
that
is
inside
embedded
inside
the
that
itself?
We
need
to
use
that
at
least
to
make
the
so
the
the
the
PA
definition
is
derived
from
that
right.
So
what
the
adapter
is
implementing
is
using
the
name.
That
is
not
that
that
has
nothing
to
do
with
the
resource
name
correct,
which
is
the
right
split.
H
Okay,
so
what
it
means
is,
as
an
operator
he
gives
the
name
of
the
temp.
The
the
template
is
named
by
him
by
his
through
his
the
source
name,
and
he
refers
in
his
in
his
in
his
universe.
He
just
references
when
he's
writing
rules
and
means
writing
whatever
uses
or
instances.
He
uses
the
name
of
the
district
of
the
the
resource
name
and
then,
and
then
we
can
link
stuff
via
stuff
as
based
on
just
that
information.
H
But
when
you're
making
calls
over
the
wire
we
use
the
name
or
the
function
function
that
we
have
to
call
over.
G
RPC
will
be
based
on
the
name
of
the
the
the
template,
because
that
is
the
API
that
the
other
guy
has
implemented,
which
which
is
namespace
than
itself.
No,
no,
no
because
I
mean
that
one
right
now
the
template
name
is
just
the
last
segment
of
the
proto
package,
name,
okay,
so
that
that
we
would
need
to
fix
that.
H
A
A
A
So
let
me
just
say
the
most
generic
possible
way
right
what
what
customer
facing
functionality
is
talking
to
another
customer
facing
functionality,
and
you
order
the
transaction
rates
where
the
error
rates,
what
are
the
latency
distributions
and
to
be
able
to
draw
that
gradually
and
to
be
able
to
you
know,
annotate.
The
edges
with
this
with
the
traffic
data
is
a
super
important
part
of
the
solution,
because
you
know
we
talk
about
easy,
oh
I
think
II.
It
has
a
few
different
value.
A
That's
absolutely
a
value
prop
to
the
kind
of
deployment
side,
but
there's
a
few
benefits
the
operation
side,
which
is
that
if
everything
was
kind
of
standardized
at
service-
and
you
can
look
from
the
customer
facing
the
customer
facing
understanding
the
world
on
down-
and
you
can
you
can
kind
of
follow
everything
from
just
a
bunch
of
infrastructure
to
define
while
working
together
and
knowing
exactly
how
they're
connected
and
how
they're
talking
to
one
another.
The
benefits
for
the
customer
are
tropical
to
you.
A
The
education
space
like
who
does
people
knowing
what's
actually
running
environment
and
how
will
and
the
second
benefit,
obviously
kind
of
probably
triage-
is
a
huge
benefit
like
it's
being
able
to
go.
Oh
here,
here's
the
graph
they're
the
points,
the
graph
when
you
go
solve
that
and
customers.
You
know
we
talk
to
this
and
I.
Don't
make
this
about
commercial
because
in
a
great
developer,
focus
but
they're
all
of
us
work
for
company
freight
and
real
life
customers-
and
you
can
we
talk
about
customers
about
this.
A
They
kind
of,
like
you,
know,
literally
get
excited
you
don't
get
out
of
their
chair
and
and
they're
like
one
who
we
had.
How
could
we
have
that?
That's
what
we
want,
and
so
with
all
that
preamble
I
felt
a
little
wordy
really
like
to
your
honor
to
make
sure
that
the
data
model
and
the
staff
model
support
us
being
able
to
draw
that
wrap.
Now
it
does
there's
gonna,
be
some
levels
and
directions
right:
I'm,
not
gonna
dictate
that
haven't
look
exactly
in
the
data
model.
A
The
way
I
want
to
put
it
on
the
screen
right.
So
there's,
like
you,
know,
indirection
between
the
service
in
the
workload
or
other
concepts.
As
long
as
it's
something
that
we
can
programmatically
collapse,
you
know
we're
good
with
that,
but
we
need
to
people
find
out.
If
I
was
a
service,
we
need
to
be
able
to
identify
what
traffic
is
going
from,
which
service
to
which
other
service,
if
we
have
those
two
things
in
kind
of
any
way,
shape
or
form,
yah-hoo
programmatically,
the
old.
A
The
rest,
that's
kind
of
a
high-level
issue,
I
think,
was
something
that
we've.
You
know
you
know:
Google
has
been
thinking
about
from
the
beginning
again,
I
was
Wendy's
fielding
a
couple
years
ago.
This
is
a
kind
of
a
solution
of
Lewis
had
in
mind
and
and
now
we're
getting
to
a
point
where
the
technical
lesson
both
like
okay,
okay,
can
we
do
it?
How
do
we
do
it?
You
know.
D
D
D
That's
the
thing:
it's
not
a
decision,
it's
a
it's
just
a
state
of
things
from
a
practical
standpoint,
there's
a
single
binary,
that's
running
that
implements
multiple
services
and
that
single
binary
can
make
calls
in
open
sockets
to
an
arbitrary
number
of
outgoing
services.
And
you
just
don't
know.
What's
the
relationship
between
what
came
in
to
the
workload
and
what
come
out
came
out,
it's
one
binary!
So
it's
not
about
how
we've
architected
this
tio
is
just
a
fact
of
that's
how
it
works.
That's
how
it
works.
It's
in
Google
as
well.
D
Just
that
that
kind
of
information
just
does
not
exist,
so
they
are
the
only
way
to
extract.
This
is
if
the
service
binary
itself,
when
it's
sent,
make
issuing
a
request.
If
that
code,
in
there
said
okay
issued
this
request
on
behalf
of
service
one
or
on
behalf
of
service
to
then
we
could
tell,
but
without
changing,
the
binary
is
just
stuffs
coming
out
of
that
binary
and
we
can't
really
figure
out
on
behalf
of.
A
Those
servers,
I,
guess
I'll,
take
it
from
the
operational,
but
then
there's
probably
arguments
to
be.
You
know
why
not
be
such
a
thing,
maybe
as
complexity
maybe
give
it
as
other
other
challenges.
One
things
from
the
architectural
respect
from
the
operational
perspective
is
basically
the
potentially
eliminates
one
of
the
gravest
operational
benefits
and
mm-hmm
friendly
again.
This
is
one
of
the
things
that
make
customers
happy
where
there's
lots
of
ways
that
we
can
address
this
issue
right.
We
can
address
it
by
assigning
a
canonical
service
for
workloads
for
telemetry
purposes.
A
Love
the
Pens
right.
You
could
also
first
stuff
that
comes
in
to
a
service.
You
can,
we
can
pass
information
along
and
if
the
context
is
propagated
from
the
call
to
through
to
the
other
call
we
can
trace,
treat
for
the
service
calls
there
as
well
right.
So
it's
not
entirely
without
a
workaround
that
gets
its
money.
A
A
I
just
want
to
basically
make
sure
that
you
know
in
you
know
it
isn't
next
month
that
we
will
be
able
to
wire
out
something
that
that
kind
of
proves
that
the
picture
is
possible
to
build
and
that
the
traffic
is
possible
Saturdays
and
that
we
can,
you
know,
Telegraph
to
the
world.
Yes,
because
higher
level
of
operational
overview
is
yeah.
D
So
so
so
the
document
I'll
send
later
will
have
actual
pictures
in
it.
The
point
is
the
in
the
end
the
view
of
services
of
circle
Center
services,
and
that
is
traffic
services.
This
is
now
what
we
can
achieve.
Something
that's
very
similar
delivers
the
same
benefit
to
the
customer,
so
I
I'm
gonna
paint
that
picture
so
just
step
away
from
the
cliff.
It's
not
yet.
It's
not
worth
jumping
yet
I.
Think
I.
Think.
A
D
So
it's
just.
We
need
to
look
at
this
from
the
the
holistic
view
of
things.
What
what?
How
do
we
configure
the
system
in
a
world
of
of
consumers
and
service
producers,
where
consumers
are
not
themselves
services?
How
do
we
visualize
metrics?
How
do
we
apply
policies
in
that
context?
It
all
needs
to
work
together
and
the
canonical
service
thing.
D
A
The
sender
is
not
a
service
or
worth
the
receivers.
Not
a
service
I
totally
totally
get
that,
but
we
should
at
least
make
sure
that
the
sender
and
receivers
or
services
we
have
a
good
solution
right
because
I
mean
customers
don't
have
to
adopt
everything
that's
available
to
them,
but
if
they
do
about
it
they
should
get
a
really
nice
out-of-the-box
experience
and.
D
I
think
the
if
the
model
a
time
I'll
be
proposing
is
if,
if
you
follow
best
practices,
where
you
have
one
service
per
workload,
one
security
principle
per
service,
then
beautiful
place.
Everything
works
great,
but
the
key
is
that
we
need
to
support
where
that's
not
the
support
properly,
where
this
is
not
the
ideal.
What
the
ideal
is
not
realized,
because
I
think
it's
going
to
be
very,
very
common,
that
the
ideal
is
not
realized
in
our
customer
deployments.
A
Smooth
transition,
pads
I
think
we
need
a
way
they
take
us
from
where
we
are
today,
which
is
a
situation
that
doesn't
support
multiple
work.
What
are
mobile
services
for
workload
for
all
of
these
things
into
and
timeline
for
this
transition
into
the
system
that
it
allows
more
complex
expressions
there.
A
F
A
The
only
other
thing
I
had
on
the
agenda
was
just
to
talk
about
some
of
the
stuff.
That's
coming
up
soon.
I
expect
to
see
some
of
you
at
KU
con
in
two
weeks,
I
think
in
the
intermediate
time
we're
going
to
do
some
more
design
work
around
the
direct
API
and
figure
out
what
we
want
to
do
with
direct
API
and
how
that
might
work.
So
look
for
just
some
documents
related
to
that.
A
We're
obviously
have
some
more
news
about
that
reporting
and
the
impacts
on
metrics
and
dashboards
moving
forward,
as
well
as
we're
starting
to
push
more
into.
How
do
we
push
signals
into
analytics
pipelines?
So
there
should
be
some
new
stuff
to
talk
about
in
the
coming
weeks
there
as
well
I,
don't
know
if
anyone
else
has
other
things
that
they
want
to
talk
about.
They're
coming
up.