►
From YouTube: 2022-11-03 meeting
Description
cncf-opentelemetry meeting-2's Personal Meeting Room
A
A
A
B
C
C
A
D
A
A
E
Thank
you,
I
mean
my
usual
place
of
of
dialing
and
to
listen
and
not
having
much
to
say.
E
I
did
see
a
comment
from
last
week's
last
session
about
our
value
and
user
front-end,
traces
and
stuff
I.
Wonder
if
there's
anything
concrete
development
developing
or
anything
we
I
can
help
with.
C
I
missed
the
last
the
last
meeting,
but
but
listen
to
the
recording,
so
I
think
there
was
a
discussion
about
the
linking
choices.
Right
and
I
mean
consistent
sampling
if
choices
across
different
races,
it's
just
possible
if
red,
if
the
the
root
span
of
the
other
choice
is
at
the
linkage
point
right.
Otherwise
it's
it's
not
really
possible.
D
Our
sometimes
spend
a
lot
of
time
in
a
queue
and
it's
impractical
to
have
them
in
a
single
trace,
because
we
want
traces
to
complete
fairly
quickly.
These
are
technical
limitations
in
The
Collector.
Very
often
we
collect
all
spans
of
of
the
trace,
and
we
cannot
afford
to
wait
for
hours
or
days.
That's
why
they
invented
this
idea
of
linking
traces,
so
they
can
be
continued
in
some
way.
E
Yes,
I'm
hearing
just
from
some
of
the
my
stock
customers
as
well.
They
really
want
us
to
have
an
answer
and
I
feel
like
the
there's.
A
traditional
tracing
vendor
position,
which
is
probably
light,
steps
which
is
traces
should
be
hierarchies
and
you
should
not
do
anything
fancy
with
them
and
it
doesn't
always
work
and
then
customers
come
in
saying.
Let's
do
linking
and
then
the
question
is
how's
sampling
going
to
break
that
I.
Don't
have
much
I.
C
No,
if
I
I
don't
know
if
this
is
the
common
case
that
the
other
tree
starts
right
at
the
linkage
point.
So
even
if
you're
Peter,
you
mentioned
this,
this
example
where
you
have
a
queue
or
things
are
put
into,
and
another
thread
result
takes
fix
that
from
the
queue,
but
does
this
other
tree
is
really
start
where
it
picks
up
the
the
item
from
the
queue
and
I
mean
the
tests.
You
know
some
history
already,
so
it
the
root
span
will
be
somewhere
else.
E
C
Yeah,
but
if,
if
just
already
is
going
on
for
that
yeah
you
you
hit
would
have
you
would
start
a
new
choice,
but
it's
already
I
know
traced.
C
E
D
D
It
is
much
more
evolving,
some
some
some
real
life
artifacts
and
in
this
case
the
shipping
can
can
even
wait
for
days
because
the
warehouse
can
be
short
on
stuff
and
and
still
we
want
to
trace
the
whole
history
of
a
purchase
that
forces
us
to
to
link
traces,
because
it's
no
longer
really
a
single
trace.
I
think.
E
Yeah,
what
I
told
my
you
know,
team
at
lightstep
that
asks
when
they
ask
questions
from
sort
of
hotel's
perspective?
Is
that
there's
sort
of
nothing
to
see
here?
If
you,
if
you
decide
to,
if
you
decide
to
use
links,
they
are
parts
of
they're
referring
to
other
traces,
and
sampling
is
just
what
it
is.
So
if
you
have
a
link
and
you're
doing
100
100
sampling,
100
100
times
you're,
going
to
find
your
link
and
or
something
like
you
know
like
there's,
you
there's
a
probability
argument
that
can
be
made.
E
So
if
you
intend
to
use
links
intent,
you
should
also
expect
to
find
them
missing
as
much
as
you're
sampling
and
and
just
keep
searching
until
you
find
the
example
that
has
both
the
link
to
span
as
well
as
the
other,
and
then
now
you
have
an
opportunity
to.
You,
know
dynamically
control
sampling
rates
to
try
and
find
the
thing
that
you're
trying
to
find
balanced
with
all
the
other
demands.
I
guess.
E
C
I
mean
I
understand
that
if,
if
really
the
auditory
starts
at
the
linkage
points,
is
it's
possible
to
to
make
the
sampling
decisions
consistent
again?
But
if,
if
the
trade
is
already
started
before
before
reaching
the
linkage
point,
then
you
just
have
inconsistent
sampling
with
all
the
downsides.
So
and
if
there's
not
just
one
linkage
point
but
multiples,
then
it's
already
impossible
to
get
yeah
to
get
all
of
them
in
in
the
same
yeah
to
capture
all
of
them
simultaneously.
So
it's
it's
like
inconsistent
sampling
and
with
all
the
downsides,
then.
F
Yeah
to
add
to
that
I'm
also
of
the
opinion
that
customers
have
started
using
a
lot
more
modeling
scenarios.
There
are
many
scenarios
where
links
are
required
to
model
that
use
case
so
telling
them
that
hey.
This
is
not
going
to
work
well,
I,
don't
think,
is
that
going
to
be
acceptable
for
many
of
those
customers
so
I
think
last
week
or
the
two
weeks
back,
we
had
discussed
that.
F
Maybe
people
could
write
a
custom
sampler
where
they
can
look
at
whether
the
span
context
has
links
and
if
so,
follow
the
position
of
the
linked
Trace
I.
Think
I
haven't
explored
that
further
and
and
see
whether
it
can
work
reliably.
But
if
that
can
be
done,
then
it
becomes
more
like
a
guidance
I
think
that
was
the
discussion
two
weeks
back,
but
but
overall
I
do
feel.
We
need
some
at
least
a
guidance
or
a
recommendation
on
how
people
can
get
little
bit
more
consistency
across
links
when
it
comes
to
sampling.
C
I
mean
if
you
make
the
sampling
decision
later
on,
when
you
already
know
that
the
choice
is
linked
and
we're
talking
about
tail
based
sampling
right
I
mean
if
we
still
be
sampling,
it's
it's
maybe
easier,
but
if
you
have
to
decide
up
front,
if
you,
for
example,
sample
The
Roots
benefit
rays,
which
you
don't
know
if
this
Trace
will
be
involved
in
and
or
it
will
be
linked
somewhere
right.
So.
F
So
in
the
example,
we
were
just
discussing
right
like
the
producer
consumer
kind
of
a
scenario.
If
the
consuming
span
happens
to
be
the
root
span
of
a
new
Trace,
then
it
could
potentially
do
a
and
if,
let's
say
the
customer
is
using
head
based
sampling,
then
potentially
it
could
be
used
to
that
sampling
decision
is
made
up
front
at
that
time.
F
Right
like
when
that
root
span
is
being
created,
but
I
I
see
your
point
like
there
could
be
other
cases
where
the
span
that
is
linking
to
another
one
is
already
like:
it's
not
a
root
span,
and
is
it
it's
already
like.
There
is
some
already
some
sampling
decision
made
by
its
ancestors,
so
even
if
it
decides
to
now
follow
the
links
spans
sampling
decision,
it
could
become
only
a
partial
Trace
is
that
is
that
what
your
saying?
Okay,
maybe
a
partial
Trace,
is
okay
in
that
case
right.
F
If
it's
it's
better
than
nothing,
I
guess,
at
least
it
says
that
hey
from
the
point
where
the
linking
happened
and
further
Downstream,
if,
if
let's
say
in
the
example
of
parent-based
sampling,
then
at
least
there
we
get
I
would
imagine
that,
having
that
partial
Trace
might
be
better
in
some
situations
than
not
even
having
the
predictability
of
having
that
partial
race.
E
So
Peter
it
sounds
like
you're
saying
we
can
do
better,
but
it
requires
you
to
essentially
essentially
resume
a
trace.
That's
already
been
started
at
the
point
where
you
would
otherwise
link.
It
doesn't
sound
like
there's
any
fundamental
way
to
work
around
the
sampling,
a
decision.
That's
already
happened
without
continuing
that
decision,
or
at
least
I
didn't
understand,
yeah.
D
D
D
It
has
some
some
issues,
of
course,
really
because
later
when
you,
you
cannot
assume
that
that
all
spans
in
a
trace
have
the
same
r
value
which,
which
complicates
a
little
bit
span
processing
you
can
so
yeah,
that's
inconvenient.
C
F
So
when
we
talk
about
so
we
are
talking
about
consistent
probability,
sampling
and
links
right,
but
but
going
a
step
back
just
looking
at
regular
parent-based
sampling
right
there
could
the
way
link
base
sampler
could
be
written
as
by
looking
at
whether
so
in
the
in
the
shoot
sample
of
the
custom
sampler
somebody
could
go
and
say:
hey,
like
does
my
current
activity
that
I
am
that
started.
F
Does
it
have
links
and
if
it
has
links,
look
at
each
link
and
look
at
the
context
of
the
link
and
whether
it
has
the
recorded
the
flag,
like
basically
it's
kind
of
like
a
form
of
parent-based
sampling,
except
that
we
are
looking
at
the
sampled
flag,
whether
it
is
set
in
the
context
of
at
least
one
link
and
if
so
then
decide
to
sample
in
the
current
span.
That
is
linking
to
one
of
those
things
which
has
already
been
sampled.
F
So
I
don't
know
if
this
is
going
to
be
give
us
a
full,
reliable,
like
a
consistent
thing
across
links
that
part
I
need
to
think
a
bit
more
about,
but
doing
something
like
that.
I
mean
this
is
just
like
a
extension
of
parent-based
sampling
right.
There
is
no
in
this
it's.
There
is
no
r
value.
P
value
impact.
I
mean
this
is
just
how
could
somebody
get
consistent
sampling
across
links
if
they
are
using
parent-based
sampling?
That's
the
problem.
I
was
so.
C
So
what
I
meant
maybe
before
is
you
could
reset
the
p-value?
Yes,
the
Prevail
here
so
at
the
linkage
point
you
could
set
it
to.
You
know:
100
sampling
probability,
and
then
everything
is
is
sampled
because
of
that
so
from
the
linkage
Point
Downstream
and
I
mean
this
is
some
kind
of
parent-based
sampling,
but
this
would
be
consistent,
at
least
because
then
you
wouldn't
have
to
reset
the
R
value
of
the
other
trees,
because
if
you're
sampling
everything
you
would
get
it
but
I
mean
then
you're
not
sampling.
E
But
they'll,
but
when
you
search
for
the
link
to
spans
and
traces
you'll
find
them,
so
this
would
be
a
and
if
they,
the
link
to
span
was
not
sampled,
you
would
have
a
zero
adjust
account
in
that
case.
So.
E
I,
like
these
here,
I'm
sorry
good.
It's
supposed
to
be
meaning
to
say
that
you
can.
You
can
make
a
decision
to
collect
something
without
changing
its
probability
and
in
other
words
and
then
sort
of
the
the
closet.
The
effect
of
the
links
to
would
be
essentially
inherit
the
probability
of
the
linked
from
the
span.
E
Let's
see
if
I
said
something
meaningful,
so
the
idea
is
I'm,
gonna
have
a
root
trace
and
then
I'm
gonna
have
a
result,
a
continuation
which
is
going
to
be
linked.
E
My
root,
my
root
Trace
is
going
to
have
some
sampling
logic.
So,
but
so
at
the
moment,
when
I'm
creating
this
link,
I
have
my
own
trace
and
its
own
sampling
probability
are
known,
and
this
new
Trace
comes
in
with
an
independent
sampling
probability
and
I
create
the
link.
At
the
same
time,
I
can
decide
to
start
recording
the
everything
in
the
linked
to
trace.
E
How
do
I
do
that
I'm
sort
of
thinking
on
my
feet
right
now.
The
the
whole
point
of
this
zero
adjusted
count
mechanism
that
we've
talked
about
just
briefly
now
is
that
you
can
record
something
without
changing
a
sampling
probability.
Even
when
something
wasn't
sampled,
you
can
record
it
with
zero.
F
Got
it
so
if
I
understand
you
correctly
you're
saying
this
little,
even
though
you
may
get
consistent
sampling,
it
may
not
play
well
with
the
span.
2
Matrix
pipelines
and
all
those
right
like
the
adjusted
count
is
not
going
to
be
possibly.
E
Well,
I
think
the
address
account
would
be
unaffected
by
whatever
the
adjusted
counter
the
linked
to
span
was
would
would
be
unaffected.
You
can
decide
to
collect
something
without
changing
its
count.
In
other
words,
but
in
that,
in
that
sense,
I'm,
just
treating
the
the
link
to
span
is
as
a
part
of
the
links
from
Trace
effectively.
E
E
F
Got
it
yeah
I'll
paste
a
small
code
sample
of
of
a
like
a
small
link
based
sampler,
that
kind
of
describes
whatever's
exploring
as
a
potential
option
for
links?
Actually,
a
teammate
of
mine
wrote
this
code,
but
this
is
basically
the
one
one
potential
option
of
kind
of
saying
that
hey
look
at
my
links
and
if
any
of
my
links
linked
traces
has
already
decided
to
sample,
then
I
also
decide
to
so
I.
Think.
F
The
main
thing
is
that
looking
at
the
recorded
flag,
which
is
basically
the
sampled
flag
in
the
context
of
any
of
the
links
but
but
I
can
see
that
this
may
not
work
that
well.
In
the
consistent
probability,
sampling
option,
I
mean
if
somebody
is
using
plain
parent
based
sampling.
It
feels
like
this
could
potentially
be
easy
to
explain.
But
I
haven't
thought
through
what
this
really
means
in
the
hope.
E
Can
we
make
an
issue
to
a
GitHub
issue
to
discuss
this
I
feel
like
I?
Don't
have
the
sitting
on
sitting
on
my
feet
right
now,
I
don't
feel
like
I
have
time
to
to
meaningfully
contribute
to.
This
sounds.
F
E
I
need
another
cup
of
coffee,
but
I
think
that
this
is
actually
I
mean
I'm,
not
trying
to
like
get
us
out
of
a
meeting.
But
I
do
think
that
this
conversation,
maybe
is,
is
I'm
not
going
to
be
able
to
contribute
to
meaningfully
in
the
next
half
hour.
But
I
would
like
to
see
this
discussed.
E
I,
I'm
and
I
see
your
snippet
and
I'm.
I
I
feel
like
what
I
want
to
try
and
clarify
as
I.
E
Is
I'd
like
to
see
you
be
able
to
do
what
you're
doing
here
without
corrupting
the
interpretation,
the
probability,
interpretation
and
I'm,
not
sure
that
this
is
going
to
do
it,
but
I
think
that
we
can
construct
a
way
to
do
it
as
long
as
you.
E
But
what
but,
but
then
now
there's
two
ways
of
counting
that
link
to
trace.
When
you
come
in
from
the
link
you're
going
to
use
the
effective
probability
of
the
span
that
you
came
from
when
you
came
in
when
you
come
in
from
the
root
of
the
linked
to
trace
you're,
going
to
find
its
own
independent,
consistent
probability
and
that's
as
far
as
I
got.
E
I
will
move
that
we
end
this
meeting.
Otherwise
I
don't
have
much
I
I
can
tell
you
I'm
distracted
by
many
things
right
now,
so
I'm
gonna
I'm
gonna
not
burden
you
with
my
distraction,
anymore.
I,
like
you
guys
a
lot
and
I
love
this
conversation.
So
I'll
see
you
two
weeks
from
now
and
I
look
forward
to
that
issue.
Depiano
we
can
discuss
that
on
GitHub
I
think
that'll
be
a
better
place.
Thank.