►
From YouTube: Policies and Telemetry WG Meeting - 2020-05-20
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
B
C
D
A
A
A
B
E
B
B
A
B
B
B
Ok,
so
I
think
so
I
think
we
should
keep
metrics
customization
separate
and
spell
that
out.
The
reason
is
that
there
are
additional
reasons
why
matrix
customization
is
more
awkward
and
even
after
we
fix
the
the
the
top
line,
API,
the
bootstrap
part,
and,
like
all
that
crap
we
still
orthogonal
attack,
still
still
needs
to
be
resolved.
D
You
can
put
Jacob
down
for
this
okay,
so
this
fits
outside
the
bucket
of
extensions
as
of
now,
because
we
configure
tracing
directly
through
on
Y
built-in
features.
As
far
as
I
know
so
I
know
everyone
wants
to
put
everything
in
the
same
bucket,
which
is
confusing
so
I,
don't
know
what
more
can
we
do
to
make
it
better
right
now.
A
It,
maybe
that's,
maybe
that's
the
ultimate
decision
we
come
to.
Is
it
it's
fine
all
the
thing
in
proxy
config
and
that's
the
API
that
we
want
to
have
from
this,
but
I
think
it'd
be
nice
to
have
one
more
look
at
what
we
can
do
to
cover
tracing
and
logging
and
Peter
at
Peter
was
mentioning
rate
limiting
to
which
I
don't
know.
B
So
rate
so
rate
limiting
definitely
belongs
in
the
extension.
Api
is
and
like
the
first
step
of
that
in
the
looper
is
doing
simply
to
show
how
envoy
native
rate
limiting
can
be
used
with
this
too,
and
then
pulling
it
and
showing
examples
of
how
it
can
be
used
with
the
new
extension.
Api
would
would
be
I
think
that's
a
requirement.
The
it
does
replace
functionality
that
we
lost.
D
D
D
B
B
D
G
G
D
B
So
I
mean
that,
and
actually
thanks
for
bringing
that
up,
Rob
as
well,
but
I
think
that's
exactly
if
if
someone
actually
goes
through
that,
like
produces
a
talk
and
says,
hey
look,
this
is
how
we
can
do
it
and
therefore
we
don't
need
to.
We
don't
need
any
other
API
and
I
think
that
could
be
adequate
and
then,
if
we
realize
there
are
gaps
then
then
we
will
create
an
API.
B
D
A
C
A
D
A
B
Don't
have
owners
in
mind,
but
okay,
so
what
but
I
can
I
can
tell
you
why
I
added
that
and
that
and
that
and
the
CNIB
is-
is
kind
of
related
right.
The
the
issues
that
we
found
late
in
one
six
cycle
were
a
because
CN
I
wasn't
quite
tested
it
with
telemetry
in
the
same
way
and
be
because
we
didn't
have
enough
negative
tests.
That
said,
oh
I
found
the
edges
that
I
was
looking
for
or
I
found.
B
H
Mean
honestly,
probably
not
I
hate
to
say
that,
because
of
the
amount
of
workload
that
I
have
already
so
I
mean
I
was
happy
to
be
able
to
bring
those
up.
I
know
it
was
so
late
in
the
cycle.
I
apologize
for
that.
But
this
is.
This
is
a
problem
that
we're
trying
to
solve
in
general
with
regards
to
getting
a
key
ally
out
with
an
sto
version.
H
So,
as
you
guys,
you
know
come
up
with
1.6
release,
you
know
we're
trying
to
get
a
key
ally,
a
compatible
key
ally
version
out,
and
we
have
trouble
testing
that
until
very
late
in
the
cycle.
Because
of
all
the
churn,
you
know,
even
even
the
first
test
day
for
is,
do
one
six
was
really
kind
of
too
early.
H
We
hit
a
blocker
right
off
the
bat,
and
so
it
becomes
this
incredible
rush,
four
key
Ollie
at
the
very
end
of
the
cycle
and
that
turned
into
a
rush
for
you
guys
on
Thursday
Friday,
when
I
finally
got
a
chance
to
run
with
the
CNI
and
I'm
like
oh,
oh,
but
the
negative
tests
would
be
very,
very
important
because
that
seems
to
be
where
we
get
hit
in
general.
Historically.
Is
that
there's
extra
telemetry
being
generated
that
you
don't
expect
and
then
it
shows
up.
You
know
when
we
look
at
the
graph.
H
B
A
A
Think
what
we
need
here
is
like
exactly
what
what's
gonna
happen
day,
one
of
one
that
a
really
stable
you
know
are.
We
are
we
moving
this
to
some
other
pieces,
some
other
repository,
and
we
are
we
pairing
off
an
old
is
steal
proxy
with
it.
That's
the
kind
of
thing
I
think
we
need
to
have
a
full,
a
full
plan
for
and
then
in
the
working
group
leads
meeting
yesterday.
A
D
D
B
B
B
What
we
do
is,
after
every
five
or
ten
minutes,
we
essentially
clear
out
the
state
and
off
of
the
metrics
that
we
have
collected
of
the
matrix
entries
that
we
have
collected
inside
mixer.
Without
that
what
happens
is
if
you
look
at
it
at
the
end
of
the
day,
you'll
basically
see
a
list
of
every
single
metric
combination
that
was
encountered
during
the
lifetime
of
that
process
right.
B
So
what
it
means
for
a
proxy
is
that,
at
the
end
of
the
day,
the
proxy
will
show
all
the
different
error
codes
and
all
the
different
methods
and
everything
that
that
it
encountered
throughout
the
whole
day.
So
it
becomes
like
a
cumulative
thing
of
these
things
I've
seen
now.
None
of
those
counters
are
change.
They're
still
the
same.
Okay,.
I
D
H
H
A
D
F
D
Reporter
so
what
that
means
is,
we
are
talking
here
about
two
things.
One
is
when
the
user
goes
to
the
Prometheus
query
server
and
queries
the
current
metrics.
They
will
see
everything
with
the
counters
which
are
not
moving,
but
if
they
convert
that
to
a
rate
or
some
other
query,
then
you
will
get
the
right
things
right.
You
can
filter
them
out
easily,
so
some
part
of
that
can
be
easily
fixed
by
filtering,
whether
it's
in
dashboards
or
other
places,
and
then
the
other
piece
is
saving
space
I'm,
not
getting
both
the
problems
right.
Yes,.
C
B
C
B
They
do
yeah
I
mean
we,
we
just
systematically
add
a
lot
of
metrics,
so
any
overage
we
have
is
again
systematically
amplified
by
all
the
proxies
and
that's.
Why
that's
why
these
things
are
important.
Prometheus
and
any
other
metric
system
is
tape
full
of
handling
like
metrics
that
go
away
and
then
handling
researched
and
all
that.
So
correctness
is
not
not
a
problem
here.
This
is
only
about
efficiency,
I.
D
A
B
B
Especially
the
open,
API
expect
thing
right,
I
think
that's
the
if
you,
if
you,
if
you
look
at
that
issue,
there
are
like
10
steps
out
of
which
three
are
done:
okay,
yeah,
so
so
there
is.
There
is
a
lot
to
be
done,
but
this
is
t
of
CTL
and
open
API.
Spec
consumption
was
like
one
of
the
main
UX
related
things
there.
So
that's
the
part
that
that
is
nicely
capable
for
an
intern.
So.
D
B
D
B
C
B
D
B
D
B
A
B
C
B
D
E
I
I
could
do
that.
Okay,
one
thing.
We
also
like
you,
sync
your
phone
at
seven.
They
have
some
new
features.
We
might
use
I
like
in
particular,
and
they
have
this
thing
where,
if
you
have
multiple
panels
with
the
same
query
but
like
slightly
different
presentations,
they
can
share
the
same
query
so
that
may
help
with
some
of
our
dashboards
Ruby.
You
know,
get
all
the
requests
multiple
times
awesome,
who.
C
A
D
A
Is
this
so
we
used
to
have
stats
on
like
say,
number
of
virtual
services.
Remember
reservation
rules
that
kind
of
thing,
but
they
were
powered
out
of
galley
that
right
and
when
it
was
consolidated
Zod.
It
started
going
through
a
different
code
path.
So,
while
our
tests
were
still
passing,
if
you
exercise
the
old
gala
code
pad
that
we
no
longer
get
stats
on
like
number
of
virtual
services
number
of
destination
rules,
so.
C
C
D
C
C
D
C
A
I
think
I
think
this
is
support
over
the
tests
and
I
think
there
was
concern
that
some
of
the
new
models,
the
multi
cluster
telemetry,
is
just
running
through
the
new
models
and
making
sure
we're
generating
the
right
thing.
So
yeah
testing,
both
manual
and
automated
so
now,
I
want
to
take
a
look
at
the
VM
stuff.
I've
been
looking
at
it
a
little
bit
if
anyone
wants
to
help
I'd
be
happy
to
have
help.
So
to
my
care,
if
you
don't
wants
to
hell.
D
I
A
B
D
D
B
I
D
H
D
Interesting
some
of
the
customers
that
we
talked
to,
they
use
other
dedicated
egress
gateways
which
are
not
our
own
voice
because
of
their
there,
they've
already
bought
into
it,
and
they
don't
want
to
get
another
half.
So
this
is
interesting.
I
will
be
interested
in
learning
and
our
and
J
about
your
use.
Cases
like.
B
So
not
every
day,
so
at
least
so.
The
specific
cases
that
I
was
thinking
about
are
to
reach
like
one
or
two
specific
destinations
right.
So
so
you
have
some
old
like
some
old
outside
the
mesh,
TCP,
endpoint
and
and
actually
this
is
TCP-
it's
not
HTTP.
It
is
the
ones
that
that
I've
seen
firsthand
so
instead
of
the
whole
base
remote
access,
this
TCP
endpoint
they
just
let
the
Eco
state,
we
access
the
TCP
endpoint
and
then
it
goes
else
goes
with
the
aggressor
gateway.
A
D
J
D
Protest
now
or
hold
your
peace
kind
of
thing,
but
we'll
we
can
come
back
and
say
if
we
do
sign
up.
Some
I
was
saying
that
we
might
have
some
more
work
in
the
security
and
networking
space
this
quarter.
If
we
do,
we
can
come
back
and
say
we
can't
do
this,
hopefully
quicker
earlier.
We
do
better
for
planning,
okay
yeah.
Let's
take
a
look
yeah.
A
D
B
B
Extensions
right
seems
like,
though,
the
weights
headed,
which
is
similar
to
how
kubernetes
in
grace
increases
all
also
to
Annette
right.
So
you,
you
define
a
point,
and
then
you
delegate
to
some
forward
difference
from
from
then
on,
and
then
that
thing
can't
contain
details
about
what
happens.
If
you,
if
you
look
at
the
dark
that
we
had,
we
actually
have
laid
out
the
requirements
right
like
what
all
you
want.
The
system
to
be
able
to
do,
which
is
provide
defaults,
have
a
well-defined
way
to
do
overrides
exclusions
and
up-
and
things
like
that.
B
So
so
now
we
have
to
reconcile
that
with
the
forward
reference
forward
references
and
then
just
come
up
with
a
with
a
concrete
proposal
and
again
if
they
like,
the
initial
doc,
does
list
forward
differences
as
one
of
the
one
of
the
options.
But
at
that
time
it
was
I
think
mentioned
as
an
option
like
a
distant
third
or
something
like
that.
But
now
it
seems
like
done
it
the
preferred
option.
So
we
need
to
yes
more
and
then
consider
its
so.
D
I
D
Can
have
other
mechanisms
to
just
think
that
I
am
struggling
with
which
I
think
I
brought
a
few
times,
and
you
agree
is
whether
we
do
forward
reference
or
embedding.
How
do
we
propagate
the
correct
context
so
that
the
thing
that
we
are
instantiating
or
referring
to
gets
the
gets
in
the
right
slot
in
the
filter
chain
or
in
whatever
else?
And
how
does
it
get
the
right
information
and.
B
We're
still
right,
not
sure
so
so
yeah.
So
then,
if
you
attack,
if
you
look
at
the
attached,
PR
and
in
the
so
yeah,
so
that
that's
the
PR
that
that
I
think
Daniel
Paul
was
talking
about
earlier.
So
there
is
a
there
is
a
companion
PR
to
this
talk
which
just
dealt
with.
Do
you
contextualize
a
filter
placement
and
you.
A
B
Yeah
yeah,
so
in
that
we
did,
we
did
define
like
four
or
five
different
stages
which
actually,
if
you
look
at
nginx
traversal
stages,
that
they're
very
similar
right.
There
is
like
three
are
thought,
telemetry
stage
and
and
then
what
we
see
is
that
we
will
place
you.
You
have
to
declare
which
one
of
those
you
are
and
then
we
will
place
you
in
one
of
those
buckets
I
see
so
its
classification
and
then
relative
ordering
within
a
classification
right.
B
B
They
are
enough,
except
for
the
case
when
there
is
more
precise
dependency
between
between
two
button,
but
in
those
cases,
what
we
have
said
is
that
you
just
specify
you
just
specify
them
together
in
the
order
that
you
want.
So
we
will
not
disturb
the
input
order
right.
If
so,
if
you
think
that
there
are
two
preauth
plugins
which
depend
on
each
other,
then
you
list
them
one
after
the
other,
in
the
exact
order
that
you
want,
and
the
system
won't
try
to
second-guess
that.
B
D
B
D
B
So
if
you
have
it's,
it's
not
actually
it's
not
really
for
xcd
to
maintain
died.
It
it's
like.
If
it's
one
document-
and
you
have
two
filters
in
it-
then
that's
just
that.
That's
just
the
text
textual
integrity
right
as
long
as
you
maintain
nothing
okay,
so
we're
not
they're
not
expecting
any
ordering
or
between
different
xcd
or
different
API
server
objects.
B
We
are.
We
are
saying
that
those
two
things
are
related
did
buy
some
rules
like
what
which
phase
you
belong
to,
and
there
is
I
think
part
of
it
is
currently
not
fully
defined,
which
is
whether
namespaces
first
or
root
namespaces,
first
or
things
like
that.
But
once
you
go
through
through
those
steps,
we
will
not
have
any
other
interests
orse
of
dependency.
B
D
A
B
I
B
Okay,
okay,
so
then
so
then,
okay,
so
I
just
see
virtual
services
services.
That's
it
but
I,
don't
see
any
I,
don't
see
us
or
sample
yeah.
So
gallery
staff
like
a
generic
thing
for
every
single
or
all
the
resources,
yeah
yeah
and
then
the
ingestion
rate
is
also
not
there.
So
so,
okay,
so
we
we
also
need
to
like
config
changes
are
not
yeah.
B
K
C
K
K
I
see
Jimmy
blob,
please
do
not
drop
folks.
This
is
the
only
time
we
can
Lulu.
While
you
are
writing
I
can
give
you
the
feedback
from
one
point:
seven
retrospective
in
some
D
when
I
collected
and
I
wrote
it
out
in
a
document.
Two
things
which
came
out.
One
was
the
testing
and
other
was
a
doc.
You
know
for
testing,
we
started
operation
excellence
for
doc.
We
already
are
in
a
good
place.
11.5.
There
are
a
couple
of
improvements
which
is
made
so
trust
me.
These
retrospectives
are
very,
very
hopeful.
D
A
D
Should
make
sure
people
are
recognized
and
I
think
we
all
you
all
did
really
amazing
work
whatever
would
we,
whatever
the
working
group
signed
up
for
it,
was
delivered
and
mostly
without
any
scramble?
So
it's
it's
remarkable
and
I
just
wanted
to
give
a
shout
out
to
Mandar
I
know
the
api's
couldn't
be
finalized
until
the
very
last
moment,
but
you
were
able
to
deliver
the
feature
nonetheless,
so
amazing
work.
Last.
B
A
Figure-Eight
know
there's
a
lot
of
hard
work.
You
know
they
required
a
lot
of
sort
of
consensus
work.
You
know
John
Howard
with
the
add-ons
and
the
Prometheus
stuff
and
Jacob
with
a
tracing
API
like
mr.
several
weeks
to
get
you
know
sorted
in
agreement,
so
we
should.
We
should
say
thank
you
for
that
work,
and
that
brings
us
what
went
well.
We
got
those
features
right.
We
got
tracing
configurability
and,
like
I
said
there
was
no
real
scramble,
I
feel
like
1
5.
A
A
A
So
we
shouldn't
have
to
rely
on
interpreting
manual
testing
of
the
Cali
team
to
find
key
issues
right.
We
should
be
able
to
catch
them
early
and
prevent
them
from
dropping
out
and
then
I
think
a
continuing
theme.
It's
just
seems
like
we're
scrambling
to
get
Docs
or
get
the
right
docs
or
edit.
The
docs
and
I
know,
even
in
the
the
release
test
to
ramp
up
to
release
testing,
you
know
just
trying
to
find
people
to
do.
The
testing
was
tough
so
that
that
I
put
in
there
okay
I'm
better.
You
know
it's.
B
So
so
I
think
that
the
related
thing
thing
in
the
next
section
is
that
the
first
draft
of
a
new
feature
should
be
should
be
completely
done
by
the
end
team,
so
part
of
what
what
went
well
this
time
is
that
we
started
the
docs
early,
but
four
features
but
I
think
the
weight
we
should
change.
That
is
when
we
begin
work
on
the
feature
the
there
should
be
a
draft.
B
That
is,
that
that
fully
describes
the
user-facing
stuff
and
that
should
be
handed
over,
and
that
should
be
the
starting
point,
because
that
way,
the
the
technical
details
are
established
like
upfront,
and
there
is
no.
There
is
no
back
and
forth
of
all
that
and-
and
that
gives
us
a
much
better
way
to
work
with
the
detect
writers.
B
L
Maybe
so
it's
like
here's
your
document
and
then,
if
you
have
any
steps
you
know
applies
or
you
know
whatever
it
is
that
you're
doing
they
actually
have
the
integration
this
year,
io
automation
they
grace
one
of
along
with
that
I
think
that
would
be
one
of
another
look
forward
for
us
now,
if
you
say
like
here's,
my
document,
what
I've
also
tested
this
as
well
for
that
one
things
do
break
in
the
feet
here
you
know
other
aggressions.
Do
then
we'll
be
able
to
fetch.
You
know
things
earlier
with
your
novel
yeah.
D
K
D
D
B
K
So
just
one
thing
I
want
to
bring
it
I
know.
We
initially
said
that
we're
going
to
have
likely
81
interns
that
actually
have
one-third,
which
is
strangely
calm
down.
So
we
will
not
get
any
support
for
the
test
cases,
which
means
I
will
run
around
again
for
1.7,
so
the
bow
could
bring
in
the
retrospect
it
is.
He
can
take
at
least
10
to
15
test
cases
over
a
few
recordings.
It
will
not
be
a
big
thing,
but
it
will
help
us
with
the
test
case
of
automation,.