►
From YouTube: 2022-05-19 meeting
Description
cncf-opentelemetry meeting-2's Personal Meeting Room
C
It
occurs
to
me,
I
don't
know
what
time
zone
folks
are
in,
so
you
agree
with
me
that
it's
the
morning,
so
I
take
it
here
in
the
u.s
pacific
as
well.
Yes,
that's
right!
Yes,
okay,
okay,.
C
A
Well,
this
small
group,
we
know
each
other-
I,
as
you
know,
have
been
focused
a
lot
on
the
hotel,
metrics
specs
and
have
not
been
really
paying
too
much
tension
here
we
did
get
our
spec
marked
stable,
so
that
was
a
major
milestone
right
now,
I'm
focused
on
finishing
our
exponential
histogram
work,
which
was
also
backed
up,
so
I
still
am
not
really
focused
here,
but
I
did
want
to
come
and
see
what
the
group
has
been
discussing
and
spencer.
You
put
a
question
in
slack.
We
should
probably
discuss
that
and
then.
C
Sure,
okay,
thanks
so
yeah,
I
indeed
had
kind
of.
I
don't
want
to
call
it
a
stream
of
consciousness.
I
tried
to
like
make
it
a
little
bit
coherent,
but
there
are
a
lot
of
questions
in
in
slack.
I
thought
I
think
I
identified
at
least
to
me
what
appeared
to
be
internal
inconsistencies
in
like
certain
spec
language
at
least
one.
Maybe
two
I'm
not
really
sure
and
yeah
joshua.
C
I
see
you
nodding,
so
I
think
maybe
I
actually
have
like
a
more
I'd
like
to
say
just
a
little
bit
about
like
why
I
was
even
like
reading
the
spec
in
that
way
and
what
my
objective
was,
and
so
my
objective
was
like
in
my
actual
day
job
I'm
trying
to
do
some
like
open,
telemetry,
sampling,
stuff
instrumentation
of
our
systems
in
a
way
that
is
like
as
forward
compatible
with
stuff.
C
That
I
happen
to
know
is
like
kind
of
coming
due
to
being
aware
of
this
group
and
what
what
we
are
working
on,
and
so
I
was
trying
to
make
it
like
forward
compatible
in
this
way-
and
specifically,
I
was
trying
to
so.
I
think
joshua
introduced
the
term
like
span
to
metrics
pipeline
and
what
I
was
trying
to
do.
C
My
task
at
work
was
to
sort
of
determine
how
to
part
of
our
telemetry
pipeline
is
converting
hotel
data
to
a
different
representation
where,
rather
than
the
p
values
and
r
value
that
this
group
is
defined,
rather
than
that,
the
concept
of
adjusted
count
is
stored,
not
in
trace
context
alongside
each
span,
but
actually,
as
a
span
attribute
on
each
span.
C
I
won't
get
into
like
why
that's
like
not
the
best
choice
and,
like
this
group
understood,
like
you,
know
it's
better,
to
put
it
in
trace
context,
but
this
is
just
the
system
that
I'm
working
with
where
it
like,
wants
to
put
adjusted
count
as
a
span
attribute
whose
value
is
an
integer.
That's
the
integer
n
in
the
sentence
like
one
in
n,
so
like
1
in
12
sampling,
you
would
put
this
span
attribute
as
12,
and
that
would
be
like
probability
1
over
12..
C
So
that's
kind
of
my
my
use
case.
I'm
trying
to
translate
between.
I
want
to
instrument
my
system
to
like
use,
set
p
values
and
things,
and
then
a
sort
of
downstream
converter
that,
like
at
the
last
moment
before
storing
into
my
storage
system,
would
convert
it
to
a
span
attribute
with
these
other
semantics.
This
one
in
n
semantic.
C
So
that
was
kind
of
my
setup
and
in
in
order
to
do
that
in
order
to
sort
of
specify
what
that
transformation
ought
to
be
between
hotel
p-values
and
this,
like
n,
in
one
in
n,
to
specify
that
transformation.
I
need
to
consider
all
the
possible
values
of
p
that
could
be
observed
on
incoming
data,
and
so
that's
pretty
clear
for
the
what
I'll
call
that
like
zero
through
62..
C
But
then
there's
this
like
63
value.
That
has
this
certain
meaning.
And
so
that's
how
I
got
like
very
focused
on
like
trying
to
be
sure.
I
totally
understood
like
what
it
would
even
mean
for
like
for
me
to
have
sort
of
recorded
and
exported
a
span
whose
p
value
was
63.
like
that,
might
even
be
a
contradiction
and
like
probably
wouldn't
even
ever
happen.
But
I
was
trying
to
sort
of
just
specify
like
what
should
this.
You
know.
C
A
Yeah
I
mean:
does
I've
talked
a
bit?
Do
yeah.
A
I
do
I
mean
I
think
you
probably
have
identified
some
inconsistencies
in
this
text.
We
should
look
at
that.
The
the
key
that
I
think
may
be
lost
from
the
nitty
gritty
text
into
the
high
level
idea
here
is
that
it's
meant
that
that
value
of
p
of
63
is
is
meant
to
to
be
special
and
and
because
there's
no
there's
no
one
in
n.
That
gives
you
you
know.
The
value
of
n
is,
of
course,
infinite
or
something
like
that.
A
It
would
be
an
acceptable
approximation,
though,
to
use
two
to
negative
63.
In
this
case,
it's
almost
zero,
it's
very
close
to
zero
and
that's
the
reason
why
it
ended
up
on
that
end
of
this.
A
The
spectrum,
the
the
primary
reason
that
I
saw
for
that
to
happen
came
when
we
started
talking
about
these
non
so-called
non-probability
samplers,
and
that
was
a
creation
to
accommodate,
essentially
like
the
bucket
the
sort
of
leaky
bucket
abstraction
is
close
enough
to
a
probability
sampler
in
in
sort
of
requirements
that
that's
not
a
good
example,
but
we
have
examples
that
are
more
like.
I
want
to
see
one
trace
per
minute.
That's
just
not
a
probability
at
all:
there's
not
even
a
rate
there.
A
So
if
you
have
a
composite
sampling
policy
that
says
I
need
to
select
one
per
minute
and
one
in
ten.
Then
you
end
up
with
this
case.
It's
a
corner
case,
but
it's
a
case
where
the
one
in
every
the
one
in
every
minute
selected,
but
the
probability
didn't
select,
and
so
we
wanted
to
distinguish
between
unknown
and
zero.
In
that
case,
and
so
unset
pr
unset,
p
value
means
unknown
and
you
do
end
up
with
a
known
probability.
C
Okay,
yeah,
that's
kind
of
where
I
landed,
and
do
you
also
agree
that
that,
like
in
theory
like
in
the
system
that
I've
described
where,
like
you
have
you
know,
services
that
are
emitting
spans,
like
they
probably
shouldn't
ever
emit
a
span
with
a
p
value
of
63?
That
is
sort
of?
Is
that
like
an
internal
to
like
within
a
process
like
that
value
might
arise?
But
it
is
unrealistic
for
that
value
to
be
present
on
an
exported
span.
A
You
know
if
you
didn't
sample
you,
you
wanna
record
it
so
because
it's
useful
information,
but
it
counts
for
zero,
and
that
sounds
a
little
contradictory,
but
it's
better
than
producing
an
unknown
when
you
know
that
the
probability
was
zero,
because
otherwise
you
can't
count.
I
think
if
that.
B
C
It's
it's
tricky
because
of
my
situation.
I
don't
yeah
like
I,
I
think
I'm
struggling
to
sort
of
grasp.
You
know
the
situation
where
simultaneously
we
like
want
to
record
it,
but
also
in
like
statistics
about
sampled
spans
like
you,
wouldn't
you
would
leave
it
out.
I
mean,
like
literally,
I
understand
what
it's
describing,
but
I
guess
I'm
as
like
a
user
of
it.
It's
a
it's
quite
odd.
I
guess
yeah.
A
Because
mathematically,
the
expected,
you
know
the
expected
count
has
to
be
zero
for
the
sums
to
work
out
on
those
things
that
you
didn't
select,
I'm
not
sure
how
to
help
help
with
this
intuition.
At
this
point,
yeah,
if
you
don't
have
a
non-probability
sampler,
this
just
doesn't
come
up.
A
Again,
I
do
feel
that
that
you
would
be
re
in
reason
to
use
the
value
of
2
to
the
negative
63
as
as
a
substitute
for
0
in
this
case
that
well
actually
wait.
I
mean
that's
going
to
mean
counting
it.
A
C
Another
question
I
had
is
one
aspect
of
this
spec
which
doesn't
sound
like
it's
in
question
or
anything
is.
C
I
think
I
have
like
a
really
like
10
000
foot
question
about
like
what
we
mean
by
composition,
and
my
question
specifically
is:
if
is
it,
I
may
have
confused
it
initially
for
one
one
process
that
I'm
familiar
with,
or
I
don't
know
I
was
thinking.
I
had
it
in
mind
when
I
was
reading
the
spec
and
then
I
was
like.
C
Maybe
it's
not
talking
about
this
is
the
process
where
you
have
sort
of
sequential
decisions,
and
so
this
is
the
case
where,
like,
if
you
had
you
know
a
sequence
of
services
that,
like
called
each
other,
it's
like
a
a
receives
a
request
from
the
internet.
A
calls
b,
call
c
calls
d
and
so
I'll
call
that,
like
sequential
application
of
like
sampling-
and
I
wasn't
sure
that's
what
that's
sort
of
my
that
was
my
mental
picture.
C
When
I
started
reading
the
section
on
composition
rules
and
then
I
wasn't
sure
actually,
if
that
section
described
this
sort
of
sequential
thing
or
if
it
described
a
different
notion
of
composition,
so
I
wondered,
if
like
you
could,
or
anyone
could
make
what
we
mean
by
composition,
a
little
bit
more
concrete.
D
Now
so
composition
is
when
you
have
multiple
samplers
deciding
about
whether
to
sample
a
single
span.
So
that's
what
was
meant
by
composition
and
as
josh
explained,
there
could
be
some
strange
things
happening
when
you
have
a
non-probability
sampler
coexisting
with
a
probability
sampler
that
causes
some
headaches,
of
course,
but
in
general
you
can
have
an
arbitrary
number
of
samplers
making
the
the
contributing
to
making
the
decision.
C
Okay
yeah,
thank
you
that
was
that
was
like
where
I
landed,
that
it
was
sort
of
a
yeah,
multiple
samplers
sort
of
ored
together
for
a
single
decision
like
a
joint
decision
for
a
single
context,
so
that
that
made
sense
and
what
ultimately
sort
of
helped
me
land
there.
That
was
in
the
spec
is
when
it
says
like
in
terms
of
like
combining
their
actual
results.
Just
like
take
the
logical
or
I
was
like.
C
Oh
that's,
like
you
know,
it's
a
single
decision,
we're
ordering
it
that's
what
is
meant
here
by
composition,
okay,
one
thing
that
I
remain
a
little
bit
unclear
on
and
I
think
I
would
appreciate
any
explanation
is
I-
and
I
said
this
in
slack,
but
basically
like
when
you're
like
I'll
just
say:
oring
when
you're
ordering,
like
a
probability,
sampler
and
a
non-probability
sampler,
it
is
surprising
to
me
that
the
like
outcome
in
terms
of
like
adjusted
count
is
well.
C
I
actually
don't
want
to
misstate
it,
but
I
have
a
little
table
that
summarizes
all
the
different
rules
but
yeah.
So
if
you
have
mixed
mixed
probability,
non-probability
and
they
it
was
yeah
like
if
the
consistent,
if
the
probability
thing
sampler
says
sample,
then
you
take
its
probability,
whereas
if
it
doesn't,
but
the
non-probability
one
does
then
that's
63.
C
is.
Is
the
rule
that
I
have
written
here
and
it
was
surprising
to
me
that,
like
it
would
have
been
my
intuition
that
if
you
combine
like
a
probability
thing
sampler
with
a
non-probability
sampler
that,
regardless
of
each
sampler's
decision,
I
would
have
guessed
that
the
resulting
adjusted
count
is
unknown
due
to
the
sort
of
non-probabilistic
sampler
sort
of
poisoning
the
like.
C
You
know
no
ability
of
the
adjusted
count
and
that
having
a
sort
of
viral
effect
on
like
if
you
ever
or
anything,
with
a
non-probabilistic,
sampler
you're.
Now
in
the
like
unknown
regime,
that
was
my
intuition.
I
want
to
hear
why
that's
not
correct.
D
Right
so
when
you
are
together
decision
from
probabilistic,
sampler
and
non-probabilistic
sampler,
and
they
both
decide
to
sample,
because
you
cannot
just
it's
not
just
like
very
simple
boolean
or
you
have
to
look
more
carefully
at
the
cases
if
the
probabilistic
sampler
decides
to
sample,
regardless
of
whether
non-probabilistic
sampler
decides
to
sample
or
not.
You
know
the
adjusted
count.
A
Thank
you
peter.
I
was
gonna
say
roughly
the
same,
you,
you
know
the
probability
from
the
probability
sampler
and
you
know
nothing
from
the
non-probability
sampler
and
when
you
combine
no
ability
with
nothing,
you
get.
No
you
get
known
and
and
that
that
p63
value
is
the
corner
case
to
make
it
all
work.
A
Meaning,
if
the
probabili
probability
sample
rejects
that
it's
zero
probability
so
zero
just
to
count
and
nothing
a
non-probability
sampler
can
do
we'll
change
that
all
the
gnome
probability
standpoint
can
do
is
change
whether
you
record
the
span
or
not,.
A
This
is
a
conceptual
leap,
so
I
understand,
and-
and
the
other
thing
is,
I
think
that
perhaps
lacking
or
sort
of
taking
away
from
clarity
here
is
that
the
rules
of
probability
that
we
learned
in
school,
like
for
for
normal
probabilistic
composition
or
not
being
followed
here,
like
you,
don't
when
you,
when
you
have
two
probability,
events
and
you
want
to
end
them.
Usually
you
multiply
probability.
That's
the
common
case
that
we
know
that
what
kind
of
intuition
that
we
have
but
for
consistent
decisions.
A
D
Well,
so
the
rules
do
not
apply
because
these
are
not
independent
decisions
right
so
let's
say
you
have
two
probabilistic
two
consistent
probabilistic
sampling,
sample,
samplers
or
together
or
ended
together,
but
they
have
different
probabilities.
D
D
C
C
C
Halfway
sort
of
understand
the
the
notion
that,
like
oh,
like
one
of
these
decided
to
sample
so
the
other
one
is
like
immaterial
and
I
sort
of
recalled
actually
I'm
gonna
post
in
this
in
the
zoom
chat.
How
I
think
what
my
hang
up
is
looking
at
this
from
like
a
more
like
analytic
angle.
So
I
I
shared
a
message,
but
this
is
when
I
was
trying
to
sort
of
rationalize
this
to
myself
earlier
I'm
trying
I
was
trying
to.
C
Just
from
like
a
first
principle
say
like
okay,
I
have
two
things:
what's
the
probability
that
at
least
one
of
them
comes
out
a
certain
way,
and
so
this
is
not
what
I
wrote
here
is
like
not.
C
It
doesn't
consider
like
the
actual
decisions
produced.
This
is
like
this
is
considering
the
like
parameters
of
the
two
sample,
the
p1
and
the
p2.
Only
one
of
those
is
not
even
probabilistic,
but
I
think
I
was
trying
to
come
at
this
from
an
angle
of
like
what
is
the
sort
of
like
equivalent
joint
probability.
Well,
you
can't
say,
because
you
don't
know
one
of
the
p's,
and
so
that's,
I
think,
how
I
sort
of
arrived
at
mine,
and
I
think,
I'm.
C
I
think
I
still
struggle
with
like
understanding
why
this
sort
of
angle
of
approach
leads
me
to
the
wrong
conclusion.
E
E
With
the
non-probability
sampler-
and
I
mean
for
that-
I
think
the
combination
is
not
really
defined.
So
then
I
mean,
if
you
combine
both,
then
you
just
take
the
sampling
decision
of
the
consistent
one
and
derive
the
p-value
from
it
that
so
it's
the
p-value
of
that
one
and
the
other
one
does
not
have
any
impact.
E
E
So,
but
if
you're
just
combining
consistent
samples,
I
mean,
as
peter
already
said,
it
is
a
consistent
or
I
mean
not
independent
sampling
decisions.
So
that's
why
the
formula
you
described
here
in
the
chat
would
not
apply.
C
E
That
and
that
I
follow
yeah
yeah
and
I
mean
a
use
case
for
that
is,
for
example,
you
have
you
know
one
one
sampler
which
samples
one
percent,
or
I
don't
know
or
yeah
one
quarter
of
full
spans
right
and
then
you
have
an
additional
sampler
which
samples
every
maybe
every
span
with
an
error
with
50,
and
you
won't
want
to
combine
that,
for
example,
and
so
you
can
define
those
two
samples
and
combine
that
with
an
or
for
example,
and
then
you
end
up
with
the
sampler
that
samples
all
the
errors
always
is
50
on
the
remaining
spans
with
25.
C
Yeah
I
mean
I
will
say
I
am
I'm
very
clear
on,
like
you
know,
taking
the
smaller
probability
of
two
consistent
probability
samplers
that,
because
they
are,
they
share
their
source
of
entropy,
so
they're
very
correlated
that
all
makes
sense
to
me
yeah.
I
think.
C
I'm
like
trying
to
convince
myself
that
it's
not
like
an
arbitrary
like
philosophical
choice,
to
like
I,
I
am
still
like,
focused
I'm,
I'm
still
framing
it
in
terms
of
the
like
parameters
of
this
two
samplers
themselves
and
I'm
struggling
to.
Like
I
hear
you
and
peter
saying
you
know
it
actually
depends
on
their
outcomes
and
like
what
their
specific
realizations
for
like
would
would
sort
of
change.
What
the
intrinsic
like.
C
I
think
my
contention
is
like
there
is
an
intrinsic
adjusted
count
for
this
pair
of
samplers,
and
what
you
are
saying
and
the
spec
says
and
like
I
think,
I'm
just
not
yet
grasping-
is
that
there
is
no
intrinsic
adjusted
count
for
when
this
pair
selects
or
doesn't
it's
sort
of
it's
a
function
of
their
specific
decisions
and
there's
like
a
two
by
two
sort
of
matrix
of,
like
you
know
the
possible
decisions
that
a
pair
of
samplers
could
produce.
E
I
mean
in
general,
the
sampras
are:
are
allowed
to
choose
a
sampling
rate
for
each
individual
span,
so
it
sampling
rates
or
sampling.
E
Probabilities
do
not
have
to
be
the
same
so
basically
for
every
span:
you're
free
to
choose
a
sampling
probability
based,
maybe
on
span
attributes
or
whatever
yeah,
and
from
that
you
get
if
the
span
gets
sampled
it
the
span
gets
its
individual
p
value
right,
and
so
you
can
also
implement
the
sampler
could
implement
a
very
complicated
conditions
or
formulas
how
to
derive
the
sampling
probability
for
this
specific
span
and
yeah
with
by
with
the
composition.
Basically,
you
can
express
some
more
complicated
samples
in
a
easier
way.
E
A
This
reminds
me
that
you
know
there
were
other
ways
we
could
have
gone
forward
with
a
specification
about,
say,
multiple
sampling
policies,
at
the
same
time
being,
let's
say,
operational
and
or
independent,
like
a
lot
more
just
describe
the
situation
where
you
have
a
sample,
that's
10
percent
of
all
spans
and
50
of
all
errors,
and
we
have
a
way
to
define.
We
think
that
makes
sense
to
combine
those
into
one
sample.
A
That's
a
combined
sample
but
like
we
could
have
also
had
a
conversation
about
how
to
report
multiple
samples
so
that
this
span
can
be
reported
as
two
adjusted
counts.
It
has
an
adjusted
count
for
the
one
sampler
and
has
a
just
account
for
the
other
sampler.
As
long
as
you
don't
combine,
those
two
counts.
You
you
have
accurate
counts.
You
have
a
count
of
errors
and
you
have
a
count
of
spans
and
they're
different
counts.
A
So
at
some
point
I
think
there's
an
option
to
to
expand
the
scope
here
and
have
a
way
of
multi
of
carrying
out
multiple
sampling
operations
that
are
not
combined,
that
are
maintained
as
separate,
but
that
would
mean
having
essentially
multiple
adjusted,
counts
or
multiple
trace
states,
essentially
for
each
independent
experiment
or
sampling
procedure.
That
you're
doing.
We
want
to
throw
that
out.
C
Yeah,
that's
interesting.
I've
not
considered
that
yeah
and
I
I
will
say
I
am
decently
acquainted
with
like
the
world
of
like
choosing
a
p
for
a
specific
decision
sort
of
based
on
span,
attributes
or
stuff
like
that,
but
yeah.
I
think
I
need
to
like
go
to
a
desert
and
meditate
on
this,
like
there's,
no
intrinsic
adjusted
count
for
a
pair
of
samplers,
but
it
depends
on
like
the
outcomes
of
the
two
samplers.
C
What
the
adjusted
count
is.
I
think
that
is
sort
of
at
the
core
of
my
misunderstanding,
so
yeah.
If
I
come
up
with
a
different
way
to
phrase
it
or
if
I
have
any
insight
that
might
be
useful
as
like
a
from
like
a
pedagogy
perspective
for
like
future
people
reading
this
stuff
I'll
follow
up,
but
I
think
I
think,
for
me,
this
thread
is
for
to
take
up
all
of
your's.
Time
is
kind
of
at
its
end.
C
So
thank
you
for
for
taking
a
half
hour
to
try
to
explain
stuff
to
me.
I
appreciate
it.
A
A
Well,
thank
you
for
that
spencer.
I
think
we
may
have
reached
the
end
of
your
topic
and
there
is
not
anything
on
the
agenda.
Would
anyone
else
like
to
raise
a
topic
or
discuss
and
then,
as
always,
I
hope
to
be
able
to
do
more
on
this
in
the
future
and
keep
going.
F
A
A
All
right,
if
anybody
else
has
a
topic
last
minute
here
here
we
are
otherwise.
I
think
we
should
move
on
and
keep
working.
C
I
I
did
have
another
thing
that
I
was
trying
to
figure
out
how
relevant
it
would
be
to
our
previous
discussion.
I
don't
think
it
was.
It
was
super
relevant,
which
is
why
I
didn't
bring
it
up,
but
I
was
sort
of
working
on
a
adjacent
problem
to
what
I
was
just
talking
about,
and
I
think
I
might
have
like
independently
arrived
at.
C
We've
talked
about
like
sc
value
in
this
group
before,
and
I
think
I
might
have
like
not
fully
understood
it
when
we
previously
talked
about
it,
and
then
I
was
like
thinking
about
some
concrete
problem
I
had
and
what
I
came
up
with.
I
was
like.
Oh,
I
think
this
might
be
like
that
c
value
we
were
talking
about,
so
I'm
actually
going
to
share
just
a
small
set
of
notes.
C
So
I
was
trying
to
understand
what
or
you
know,
reason
about
what
adjusted
count
would
be
for
a
sequence
of
decisions
on
not
a
single
context
but
again
yeah
like
context
through
multiple
separate
decisions,
samplers
and
where
I
kind
of
landed.
Was
that
all
actually
I
mean
depends,
do
you
want
to
start
abstract
or
concrete?
Abstract
is
up
here
concrete's
down
here,
but
I'll
give
you
a
couple
minutes
to
sort
of
read
read
this.
I
think
it's
all
on
the
screen
now.
A
A
Well,
you
have
these
independent
probabilities,
which
you're
multiplying
together
right,
and
that
was
the
that
was
what
we
were
talking
about,
how
none
of
the
consistent
sampling
composition,
rules
that
we
have
are
designed
to
handle
that
type
of
scenario.
C
E
Yeah
I
mean
for
this
scenario:
you
actually
need
you
know
to
collect
the
another
kind
of
p
value
for
the
independent
samplers
right
and.
E
Yeah,
if
it's
in
japan
yeah,
I
think
c,
was,
I
think,
related
to
yeah
table
sampling
are
probably
also
independent
yeah.
You
could
also
interpret
maybe
a
c
value,
but
the
point
is
that
you
need
two
different
values.
E
You
you
have
to
to
you're
not
allowed
to
mix
those
values,
because
otherwise
you
cannot
estimate
correctly.
C
Right
that
was
my
conclusion
as
well:
yeah,
okay
yeah.
I
guess
my
next
question
then
is
like,
and
this
may
be
directed
at,
like
the
people
with
actual
customer
experience
so
like
joshua
and
ben.
Like
I
mean,
obviously,
this
group
prioritized
like
a
scheme
for
transmitting
like
consistent
sampling
over
like
inconsistent
or
independent
or
whatever,
and
I
guess
I
was
curious,
like
can
anyone
think
of
like
a
case
where
someone
would
like
reasonably
want
to
be
like?
C
Oh
no,
I
want
each
node
in
my
system
to
make
an
independent
decision.
Is
that
a
thing
that
anyone
can
conceive
of
demand
for,
or
was
this
just
like
an
academic
exercise
on
my
part.
D
A
Don't
we
get
benefit
by
in
in
the
common,
simple,
straightforward
case,
where
you're
sampling
at
the
root?
I
think
that's
the
case
where
it
is
possible
to
con
to
mix
independent
probability
decision
with
a
consistent
probability
decision,
and
this
is
where
I
think
spencer
has
pointed
out.
We
could
add
a
c
value
that
can
be,
I
think,
maintained
in
parallel
correctly.
I
think
we
would
have
to
spell
out,
as
rules
is,
that
what
we're
getting
at
here.
E
Another
thing
which
have
to
be
considered
if
you
allow,
for
example,
yeah
independent
sampling
in
the
context
of
traces,
is
that
you
know
the
extrapolation
can
be
very
expensive
for
the
the
estimation
can
be
very
complex
if
you
have
to
consider
for
those
independent
sampling
decisions,
especially
if
you're,
for
example,
counting
where
you
want
to
count
the
number
of
traces
which
you
know
called
some
series
a
and
series
b.
C
I
don't
I'm
not
sure
I
followed.
Why
is
that?
Is
that
saying
that,
like,
like,
I
feel
like
I
just
heard
sort
of
beneath
the
surface
of
what
you
just
said,
something
quite
profound,
which
is
that,
like
throughout
a
system
there's
no
there's
no
measure
that
can
be
taken
so
that,
like
locally
local,
like
isolated
parts
of
the
system,
can
like
decide
how
to
do
things
on
their
own?
Rather,
the
entire
system
has
to
sort
of
participate
in
a
coherent
sampling
design
in
order
for
the
whole
system's
output
to
be
validly
countable.
C
E
I
mean-
or
I
mean
maybe
say
it
in
a
different
words:
the
the
number
of
possible
outcomes.
If
you
have
in
independent
sampling,
you
know
grows
exponentially
with
the
number
of
sampling
decisions
right
and
because
you
know,
every
independent
sampler
has
its
own
decision,
which
is
completely
independent.
So
if
you
have,
I
know
10
samplers,
then
there
are
2
to
the
power
of
10
outcomes,
of
what
you
could
get
from
a
trace
right
and
with
consistent
sampling.
E
The
number
of
possible
outcomes
is
quite
limited.
It's
basically
the
number
of
distinct
p-value
which
have
you
chosen
within
the
trace.
So
it's
much
better
to
to
analyze
actually
and
for
estimation.
You
usually
have
to
yeah
consider
all
the
possible
outcomes,
and
you
know
and
use
corresponding
extrapolation
factors
and
things
like
that,
and
so
it's
it
becomes
invisible.
If
you
have
independent,
independent
sampling
decisions
or
if
you
have
many
of
those.
C
Okay,
I
think
that
is
a
statement
against
what
I
had
presumed
to
be
true,
but
it
may
not
be
true
that,
like
I'd
imagined,
I
could
have
like
a
graph
of
you
know:
services
or
processes
or
whatever
and
like
those
could
all
be
configured
to
be
doing
like
independent
sampling,
and
my
presumption
is
that
there
exists
some
way
of
like
reporting
that
on
span
contexts
and
things
such
that,
like
accurate
estimates,
valid
estimates
are
possible.
C
E
B
E
Sampling
decisions
yeah:
this
is
actually
what
you
did
here
right
and
but
if,
if
you
wanna
estimate,
yeah
more
complex,
yeah
quantities
from
a
trace,
then
it's
not
so
obvious
anymore.
C
I
think
one
thing
that
might
be
illustrative
is
like:
if,
if
we
considered
like
a
system
where,
like
service
a
selects
with
probability
one
quarter
and
service
and
that
it
calls
service
b,
that
selects
with
probability
1
16.
C
and
is,
is
your
statement
that
you
know
sets
traces
for
such
a
system,
which
I
guess
I
guess
I
would
ask
like.
What's
an
example
of
a
quantity
that,
like
would
be
in,
you
know,
difficult
to
estimate.
I
guess
in
that
system,
where
a
calls
b
each
doing
independent
sampling?
What
is
such
a
quantity.
E
Yeah
an
example
is,
if
you
want
to
know
how
many
traces
you
are
called
a
and
b
so
yeah,
and
this
means
that
you
need
to
see
a
connection
between
a
and
b
within
a
trace
and
if
you
have
yeah,
if-
and
you
know
this
happens,
this
connection,
you
only
see
if
you
know
all
the
spans
in
between
are
sampled
so
and
if
you
have
independent
sampling
decisions,
yeah,
it's
it's
your
first.
You
see
that
you
have
a
very
low
probability
to
see
this
connection
at
all.
C
E
Even
if
the
the
estimation
would
be
simple,
you
get
the
high
variance
of
high
statistical
error
for
estimating.
C
E
E
E
C
I
appreciate
the
the
concrete
examples
of
like
quantities
that
are
either
impossible
to
validly
estimate
or
just
like
merely
difficult,
and
the
reason
I
appreciate
them
is
because
I
I
think
I
first
of
all,
I
think,
no
vendor
in
the
observability
space
like
supports
such
quantities
today,
like
the
average
depth
of
a
trace
or
something
like
that,
and
so
it
I
think
those
don't
like
occur
to
me
as
like
possible
things
that
could
be
estimated,
but
I'm
sure
I
have
no
doubt
that,
like
actual
research
on
tracing
systems
like
has
thought
up
some
pretty
interesting
quantities
that
could
be,
you
know,
said
of
a
tracing
system
that
aren't
yet
like
in
my
intuition
as
things
that
I
would
even
care
about.
F
F
I'm
wondering
if
we
are
going
to
get
pushed
to
have
a
stronger
way
of
representing
those
two
pieces
coming
together
and
coming
up
with
a
useful
probability
rate
or
some
sort
of
number
that
vendors
can
then
use
in
the
kind
of
trace
to
metrics
pipeline
other
than
it
feels
like
now
and
I'm.
I
will
totally
admit
that
I'm
only
on
like
50
on
the
details,
but
it
does
feel
like
now
that
we
we
lose.
F
We
lose
some
granularity
or
some
detail
in
that
or
have
they
have
the
possibility
of
losing
that
where,
if
we
end
up
in
the
space
of
this
63,
you
know
p63
number.
We
kind
of
we're
kind
of
games
over
at
that
point,
there's
not
much
to
do
to
come
back
out
of
that.
F
E
I
I
mean
till
by
sampling
and
and
this
kind
of
sampling
there
I
think
they
can
be
combined
they're,
quite
independent.
The
thing
is
that
tail-based
sampling
is
usually
sampling,
the
whole
trace
it
composes
the
trace
first,
and
then
you
have
one
consistent
decision
across
all
spans
of
the
trace,
so
it's
just
a
furthest
sampling
stage
and
of
course
you
have
to
regardless,
which
sampling
probability
you
choose
for
the
trace.
E
E
You
do
not
have
additional
possible
outcomes
right,
so
it
it's
it's
quite
and
what
to
go
now
approach
so
they
can
be
combined
can
be
used
together.
The
first
stage
you
have
span
sampling
using
consistent
sampling
and
then,
if
you
have
the
full
trace
already
composed,
you
can
add
an
another
sampling
decision
which
yeah
which
acts
on
the
whole
trace,
and
so
I
do
not
see
any
any
problems
for
tail
based
sampling
here.
F
Yeah,
I
think
maybe
that's
where
I'm
getting
at
is
if
there
is,
if
that
becomes
an
easier
use
case,
to
focus
on
to
add
any
additional
parts
of
the
spec
that
might
be
necessary,
or
maybe
maybe
not.
Maybe
what
I'm
hearing
is
that
it
actually
could
kind
of
work
today,
but
that
might
be
a
nice
place
to
start
from
in
expanding
the
the
stretch
of
this
or
the
scope
of
this,
the
spec
that
we
have.
A
I
I
want
to
say
I
think
what
otmar's
has
has
proposed
I've
seen
the
code,
and
I
think
I
get
the
basic
idea
of
the
algorithm
to
do
consistent,
tail-based
sampling.
On
the
other
hand,
I
want
to.
A
E
Reservoir
sampling
of
spans
right,
which
happens
consistently.
So
this
is
a
different
thing:
okay,
so
it's
before
which
which
can
still
yeah.
The
problem
is
that
you
have
the
same
place
to
do
an
immediate
sampling
decision
right
and
so
in
the
in
the
in
the
processor
yeah
you
can,
you
know
you
can
buffer
the
spans
and
then
drop
spins
and
then
adjust
the
adjusted
counts
correspondingly,
but
this
is
still
before
the
the
spans
are
combined
with
each
other
or
before
the
trees
is
built.
E
Actually
because
the
this
is
what
I
consider
as
tail-based
sampling
is,
when
you
already
have
all
spans
of
a
trace,
yeah
physically,
combined
and
and
then
do
a
sampling
decision
on
the
whole
thing.
So
this
is
just
an
additional
stage
this.
What
I've
proposed
is
an
additional
stage
between
the
the
sampler
and
and
yeah
and
the
exp
yeah.
A
Pipeline,
I
think,
is
what
you're
describing
so.
What
I
think
perhaps
I
can
speak
for
spencer
now,
is
to
try
and
like
if
stepping
back,
what
we
want
is
a
span
of
metrics
pipeline.
We
know
that
there's
going
to
be-
or
this
is
one
of
the
things
we
want.
We
know
that
there's
going
to
be
span
sampling,
that
happens
in
the
sdks
in
context,
and
we
know
now
we've
specified
how
to
carry
out
that
consistent
sampling
and
then
collect
those
spans.
A
This
is
a
new
thing,
come
up
with
a
new
way
to
say
what
the
adjusted
count
of
the
trace
is
and
but
what
I
think,
what
we're
hoping
for
is
that
we
can
just
sort
of
pass
through
those
spans
and
and
make
sure
that
the
count
on
the
spans
reflects
both
stages
of
sampling,
and
I
think
it's
true
that
we
know
how
to
combine
consistent
sampling
parameters
so
that
we
could
combine
p-value
and
modify
a
p-value
in
the
second
stage.
A
But
I
think
what
we're
getting
at
here
is
that
everyone
understands
independent
sampling,
flipping
coins
is
natural
and
the
outcome
is
so
simple
that
we'd
like
to
just
perhaps
have
another
parameter,
which
is
the
combined
independent
sampling.
That's
happened
and
that
could
be
expressed
as
c
c
value
and
it
might
allow
us
to
do
both
to
do
sort
of
the
straightforward
form
of
tail
sampling,
which
is
limit
spans
to
adjust
the
traces
you
want
and
then
continue
setting
expands.
A
E
E
We
could
have
independent
sampling
on
a
span
level
and
also
on
on
a
trace
level,
so
which
would
be
tail
based
sampling,
so
so
this
would
require,
I
think,
two
additional
values
on
on
which
you
have
to
store
in
order
to
be
able
to
estimate
correctly
because
the
one
the
tail
based
approach
you
know,
makes
this
decision
for
all
the
spans
of
a
trace.
So
this
is
also
some
kind
of
consistent
sampling.
E
F
If
I
understand
correctly,
I
I
think
what
I,
what
I'm,
what
I'd
be
interested
in
and
maybe
we're
moving
towards,
is
starting
with
the
consistent
sampling
and
process
and
the
and
flexibility
I'm
going
to
be
a
little
ambiguous
in
the
tail
sampling
and
it
we
would
have
to
be
clear
on
kind
of
where
these
different
pieces
can
be
used
and
how
they
can
be
used
properly
together.
F
I
think
there's
enough
demand
for
this
kind
of
tail
sampling,
where
you
can
capture
all
the
errors
and
drop
all
the
other,
interesting
things
that
are
less
interesting
and
so
forth,
enabling
that,
in
a
focused
way,
that
gives
it
an
ease
of
implementation
and
that
spanned
the
metrics
pipeline,
without
kind
of
going
to
the
full
suite
of
of
cross
product
of
all
the
different
ways
of
kind
of
combining
these
things
at
both
span
and
trace
levels
in
process
and
tail
and
so
forth.
A
To
me,
this
leads
to
a
conclusion
that
there
should
be
perhaps
another
probability
number
that's
just
like
we
know
people
want
to
do.
Randomization
and
probability
falls
out
of
it,
but
it's
not
consistent
anymore
and
it
seems
to
me
fairly
straightforward
to
you
know.
If
you're
applying
one
and
three
sampling
at
a
collector,
then
we
can
just
put
a
c
value
of
three
on
those
spans
and
when
we
come
to
the
end
of
expanded
metrics
pipeline,
you
do
the
p-value
calculation
and
multiply
it
by
the
c
calculation
and
that's
your
total
adjusted
count.
A
I
think
I
I
think
what
I'm
hearing
from
band
and
probably
spencer
as
well,
is
that
there
would
be
value
in
having
that
type
of
thing
specified.
You
recall
that
there's
a
probabilistic
span
sampler
in
the
collector
and
right
now
it's
doing
one.
You
can
configure
it
for
one
and
three,
but
we
have
no
way
to
record
that
right
now
and
I
I
had
briefly
proposed
the
idea.
Well,
we
could
just
if
it's
a
power
of
two
we
can
multiply
or
add
the
logarithm
to
the
p
value
and
that's
the
right
thing
to
do.
A
C
So
I
think
I
gathered
that,
like
there's,
probably
not
a
use
case
for
like
many
like
it
would
be
strange
or
uncommon
for
somebody's
system
to
like
have
multiple
independent
sampling
stages.
But
probably
one
at
the
end
is
likely
to
be
a
pretty
common
sampling,
design.
E
I
mean
you
mean
in
the
you
mean
a
last
stage,
which
makes
a
sampling
decision
already
for
the
whole
trace
right,
yeah
so
like,
whereas.
C
In
my
sort
of
example,
in
my
notes,
I
had
like
two
stages
of
independent
sampling
and
I
think
that
kind
of
may
have
like
misled
people
to
be
like
that's
kind
of
weird,
but
then
when
we
were
like.
Oh,
if
there's
just
one
at
the
end,
that's
eminently
normal,
and
I
think
I
gathered
like
the
same
mechanism
that
same
c
value
would
be
used
for
both
it's
just
a
question
of
like
what
will
people
actually
like
want
to
do
in
practice?.
D
We
are
over
time
here,
but
my
personal
choice
would
be
to
get
rid
of
this
independent
sampling
entirely
and
try
to
do
everything
using
consistent
sampling,
for
example.
If
you
want
to
keep
all
the
errors,
that's
not
an
issue.
A
All
so
one
development
that
we
might
bring
in
here-
and
I
know
we
are
over
time-
I
think
I
got
to
go
to-
is
that
the
w3c
has
moved
and
released
a
update
with
a
new
chase
flag.
That
was
meant
to
help
us
out
and
that-
and
that
is
to
the
idea
that
we
can
put
stop
having
an
r
value.
You
can
start
putting
that
flag
setting
that
flag
on
the
trace
id
so
that
we
know
that
55
bits
or
60.
A
Some
some
number
of
bits
are
random,
random
enough
and
that
you
can
use
that
to
make
your
decision.
We
probably
need
to
update-
and
I
we
were
kind
of
I
was
waiting
for
that,
because
without
that
you
have
to
have
an
r
value
on
every
unsampled
request
and
that
gets
pretty
expensive.
I
think
what
to
take
peter's
point
a
little
further.
I
think
we
can
probably
leverage
that
and
and
and
truly
build
pipelines
that
take
advantage
of
that
randomness
to
do
consistent
sampling
so
that
we
don't
need
to
go
to
a
c
value.
A
However,
however,
we
still
have
people
who
are
going
to
say
I
am
doing
one
and
three
sampling,
so
power
of
two
consistent
sampling
is
is
a
limit,
but
but
now
that
we
have
this
random
data,
that's
that's
truly
random.
We
can
actually
do
non-power
of
two
sampling
as
well.
I
think,
although
I
don't,
I
know
that
that's
gonna
cause
some
concerns,
but
you
know
as
long
as
those
bits
are
ordered
we
can
we
can.
We
can
flip
coins
in
between
the
powers
of
two
as
well.
A
I
think
I
just
I'm
not
sure
that
that's
going
to
be
a
popular
idea.
I
know
atmar
in
the
paper
which
we
need
to
link
for
in
the
notes.
Basically,
you
would
like
to
not
have
arbitrary
probabilities
just
because
it
makes
the
calculations
intractable,
not
because
of
any
other
property.
As
far
as
I
know,.
E
Yeah
I
mean
for
I
am
in
durham.
I
listed
a
couple
of
reasons.
One
is,
of
course,
that
you
know
if
you
want
to
estimate
integer
quantities
and
you're
using
so
and
if
you're
using
powers
of
two
sampling
rates,
then
it's
guaranteed
that
the
state
estimates
will
also
be
intentional
threat,
which
is
yeah.
I
think
yeah
good,
because
if
you
want
to
show
the
estimate,
then
you
do
not
have
to
round
it,
because
I
mean
I
know
if
someone
wants
to
see
an
estimate
of
three
dot.
E
People
here
so
so
this
is
one
thing,
then
the
other
thing
is
that
it
limits
the
number
of
possible
outcomes
of
which
you
can
see
which
allows
you
to
simpler
estimate,
more
complex
quantities,
as
described
in
my
paper
yeah.
C
E
Of
course,
the
p-value
can
be
represented
in
a
compactor
way,
so
it's
because
you
only
need
to
store
the
exponent
instead
of
a
floating
point
yeah
this
this.
I
think
the
main
arguments.
A
Yeah
I
see
that
as
a
fairly
motivating
reason,
just
the
compact
representation,
but
I
think
it's
I
mean
we
have
to
recognize
that
c
values
are
integers
so
that
first
objection
kind
of
doesn't
work
for.
There's
no
objection
there
like
as
long
as
c
values
are
integers,
then
your
estimates
are
going
to
be
integers
too
so,
like
in
the
stats
d
sampling
right
now,
we
like
you,
have.
E
A
Note
saying
please
use
integer
inverse
integer
reciprocals,
or
else
we
will
be
rounding
and
you
will
not
know
the
difference
so
anyway,
we
are
over
time.
This
has
been
constructive.
I
do
think
we're
we're.
We
can
see
the
work
ahead
for
this
sort
of
tail
sampling
work,
stay
tuned,
keep
coming,
we'll
keep
keep
at
it.
Thank
you
all.