►
From YouTube: 2022-11-17 meeting
Description
cncf-opentelemetry meeting-2's Personal Meeting Room
A
A
B
C
A
C
B
B
I
do
that
last
time
we
left
off,
we
were
kalyana,
had
presented
a
topic
about
question,
question
really
asking
about
sampling
and
links
and
there's
an
issue
now
that
has
some
discussion
probably
put
that
in
the
agenda.
B
I
put
some
information
from
mine
into
that
issue
and
talk
about
that
next
on
the
list
was
one
that
somebody
asked
us
to
talk
about
in
Slack,
asking
about
science
exemplars,
although
I'm
not
sure,
there's
much
for
us
to
say
that's
about
it.
That
I
had
I
found
an
old
issue
that
we
might
eventually
want
to
discuss
again.
B
I
feel
I'm,
starting
to
feel
like
it's
time
for
some
work
on
proposing
the
non-power
of
two
probability:
sampling
that
we've
kind
of
speculated
about
based
on
WPC
case
context,
I,
think
I'm,
starting
to
think
that,
because
we've
talked
about
it
and
I
know
that
there
was
this
issue,
the
one
I
just
pasted,
finally
into
the
notes
that
we
discussed
a
few
months
ago,
maybe
six
months
ago
about
maybe
more
should
present
my
screen
for
that.
B
Well,
actually,
we
should
talk
about
this
failed
almost
a
year
ago,
so
we
should
we'll
put
that
one
last
about
Trey
State
when
you're
not
doing
probability
sampling
at
the
head.
Okay,
so
I've
put
three
things
in
the
agenda
and
I
think
we
should
start
with
me
presenting
my
screen.
B
This
is
the
one
that
I
thought
was
most
relevant
this
week,
and
so
you
can
read,
of
course,
you've
probably
read
this
already,
but
the
summary
is
that
we're
asking
about
how
a
user
might
experience
the
presence
of
both
span,
links
and
sampling
and
there's
a
link
that
that,
if
you're
following
closely,
you
may
be
familiar
with
some
just
some
debate
in
this
space
unrelated
to
sampling,
but
having
to
do
with
span
links
that
this
fellow
Johannes
tax
has
been
working
on
for
a
while.
B
So
the
links
kind
of
crossed
here
or
the
the
issues
crossed
here,
I
wrote
this
one
response.
Last
last
earlier
this
week,
sort
of
half
half
answering
what
I
think
about
this
problem,
but
also
trying
to
find
solutions
for
the
the
real
underlying
problem,
which
I
think
is
exposed
by
the
issue
of
that
that
having
to
do
with
messaging
and
spam
links.
B
B
D
Thanks
Josh,
this
is
certainly
helpful.
So
the
first
section
here,
I
just
did
a
small
proof
of
concept.
Yesterday
just
tried
to
write
that
composite
sampler,
as
you
said,
using
a
non-probabilistic
sampler
for
the
one
which
looks
at
the
links
and
then
I
just
used
a
regular
pattern
based
sampler
for
the
other,
probabilistic
sampler
and
then
compost
both
of
them
I
think
I.
Think
that's.
That
sounds
like
a
fine
way
to
go.
I
think
couple
of
follow-up
questions
I
had
number
one.
D
Is
this
something
we
imagine
would
be
like
a
whole
tip
and
we
go
a
bit
further
in
terms
of
specifying
similar
to
how
the
other
things
you
have
in
the
like
in
consistent
in
the
link,
you
have
right,
like
the
composition
rules
you
talk
about.
This
is
how
you
do
should
we
should
we
go
a
bit
further
and
say
hey
if
it
is
links,
it
should
be
like
whatever
you
have
in
this
first
section.
D
Should
that
become
part
of
the
specification
in
some
form,
that
will
probably
lead
the
way
for
sdks
to
even
support
out
of
the
box
samples
in
in
the
future
or
or
do
you
think
we
should
stop
with
like
some
kind
of
a
guidance
or
like
a
sample
sample
is
probably
the
wrong
word
here
in
this
context,
but
like
an
example
or
a
demo
thing
or
like
I'm,
just
trying
to
see
what
is
the
spectrum
of
where,
where
we
would.
B
Well,
I
think
for
my
myself,
I
I.
The
reason
why
we
haven't
got
I
guess
like
an
SDK
spec
that
has
composition
rules
included
in
it
is
that
I
think
we
suspect
in
this
group
that
users
may
ask
for
configurable
sampling,
but
what
they?
What
they
want,
is
quite
a
lot
more
sophistication
than
just
one
aspect
of
configurable
sampling,
so
being
able
to
to
specify
a
single
composable.
Sampler
is
just
another
building
block
and
people
have
been
are
asking
for
hold
Solutions,
not
not
building
blocks
I.
C
I'll
just
say
that
you
know
like
one
of
the
use
cases
we
see
sometimes
is
somebody
who
has
you
know
they
have
a
trace
and
they've
sampled.
The
trace
and
the
trace
has
I
saw
one
this
week
they
had
on
average
three
times
as
many
span
links
within
the
traces
that
were
spans
within
the
trace
and
they
wanted
to
sample
those
span,
links
pretty
aggressively
and
like
they
didn't
care,
they
didn't
care.
C
If
all
of
the
links
were
present
in
a
given
Span
in
a
given
Trace,
they
were
willing
to
say
I
just
want
there
to
be
less
spans,
and
so
so
that's
just
like
a
data
point
of
there
are
some
people
who
have
for
who,
for
practical
reasons,
are
content
with
tail
sampling
even
and
saying.
No,
you
know
just
just
get
rid
of
some
of
this
crap
for
me
because
I
don't
want
it
all
here
and
you
know,
then
they
were.
C
There
were
requests
for
sort
of
more
sophisticated
ways
of
specifying
that,
but
but
on
a
base,
thing
like
the
your
probabilities
should
be
stacked
so
that,
if
you
pick
the,
if
you
pick
the
trace
you're
going
to
pick
all
the
links,
that's
actually
not
what
some
people
want.
So
that's
a
like
making
that
normative
and
say
it
has
to
work.
That
way
would
probably
be
against
what
a
lot
of
people
actually
want
to
see.
C
Sorry
they
had,
they
have,
they
have
big
traces,
a
trace
with
hundreds
or
thousands
of
spans
and
and
well
in
span
is
Honeycomb
defines
it
includes
fan,
links
and
Spain
events,
and
so
so
they
just,
but
they
show
up
not
in
the
trace
waterfall
as
we
call
it
for
the
the
view
of
the
whole
Trace.
C
They
show
up
as
like
points
attached
to
the
things
they're
linked
to
and
they
just
they're
like
you
know,
these
are
all
events
in
my
system
and
and
they're
interesting
to
me
sometimes,
but
but
in
the
context
of
a
given
Trace,
they
may
not
be
interesting
to
me
and
I
would
like
to
I
would
like
to
sample
span
events
differently
and
more
aggressively
than
I
am
sampling,
other
kinds
of
spans,
and
so
so
a
a
thing
that
says
the
probability
of
keeping
the
link
should
be
equal
to
or
higher
than
the
lynx.
B
Sampled,
like
we
can't
do
much
better
than
the
than
combining
probabilities
that
you're
going
to
get
all
your
span
links
and
that,
in
my
my
belief,
I
think
this
is
what
you're
sort
of
verifying
is.
That
is
that
users
generally
should
not
expect
all
of
their
strand
links
to
be
present,
but
still
I.
Think
there's,
there's,
maybe
a
need
for
guidance
and
I
I
want
to
I
think
that,
obviously,
from
this
first
response
by
Johannes,
I
didn't
make
myself
clear
in
this
sentence.
B
This
pair
of
this
section
here,
maybe
I,
can
try
it
in
words
and
Kent
and
see
if
this
might
help
for
users.
So
the
the
problem
that
I
see
is
that
when
we
create
a
child,
the
child
refers
to
as
parent,
but
the
parent
never
gets
a
reference
to
its
child
to
its
child.
There's
no
spam
event
saying
I
have
a
child.
You
assume
that
the
child
will
be
there
to
reconstruct
the
child
parent
relation.
B
Well
in
this,
in
the
case
of
a
span
link
and-
and
that
was
the
reason
why
we
have
this
non-ascending
probability
recommendation,
which
is
to
say
that
if
you
choose
something
right,
then
all
of
your
children
will
be
there,
but
as
long
as
you
don't
lower
your
probability.
Otherwise
you
have
this
Gap
that
you
won't
literally
won't
know
about,
because
you
can't
tell
when
an
unsampled
child
appears
unless
you,
unless
you're
consistent.
B
So
the
same
type
of
problem
happens
with
Spam
links.
You
are
creating
a
new
span.
It
is
going
to
link
to
something
and
that's
something
that
you
are
linking
to
has
no
notion
that
it
was
linked
to
and
unless
you
have,
and
so
this
notion
of
completeness
is
I
when
I
hear
it
I
think
what
people
are
looking
for
is
like
both
a
way
to
know
that
all
my
span
links
whether
my
spam
links
were
or
were
not
present,
and
that's
something
we
can.
B
We
can
indicate-
and
you
can
Shuffle
probabilities
to
try
and
improve
that
those
odds.
But
we
can
just
tell
you
these
were
weren't
sampled,
but
the
opposite's
not
true
and
I.
Think
one
of
the
reasons
that
we're
holding
back
on
this
second
part,
which
is
allowing
span
links
to
be
created
arbitrarily
after
the
life
after
a
Spam,
starts.
B
The
reason
why
we've
held
off
on
that
is
that
there's,
no
good
sampling
story
and
I
think
we
could
fix
the
sampling
Story
by
effectively
saying
that
links
are
bi-directional,
like
you
create
a
link,
so
you
you
are
pointing
to
something.
That's
something
you're
pointing
to
has
an
event
that
links
back
and
if,
if
either
side
is
sampled,
you
can
save
that
reference
to
see
that
there
was
something
whether
or
not
it's
sampled
is
independent.
B
So
so
it's
basically
a
summary.
That's
saying
that
if
we
added
to
our
data
model,
saying
anytime,
there's
a
link,
there's
a
there's,
a
reverse
link
receiving
it,
it's
a
reference
from
reference
from
type
of
event,
then
you
could
record
the
reference
from
event
in
the
parent
or
the
linked
to
context.
Everybody
can
sample
consistently
and
do
what
they
like
and
you'll
see
when
there
was
something
missing.
That's
that's
kind
of
the
the
critical
missing
feature
right
now
is
is
that
we
don't
have
a
way
to
tell
when
you're
missing
something.
C
B
Right
right,
it
does
assume
that
and.
C
Which
is
not
really
the
case
in
like
say
a
Kafka
environment
where,
okay,
you
know
I'm
going
to
embed
in
the
packet
that
I
shoved
through
Kafka,
my
Trace
ID
and
then
meanwhile
off
goes
that
my
producer
trace
and
then
sometime
later,
my
consumer
sucks
that
thing
up
and
goes
hey.
Here's
a
link
to
the
producer,
Trace
I
got
no
way
to
modify
that
Trace.
It's
long
gone.
B
Yeah
right,
what
makes
me
wonder:
I
mean
how
ambitious
we
all
feel
and-
and
this
is
really
already-
this
issue
is
bigger
than
sampling.
So
we
can.
We
can
just
sort
of
offer
opinions,
but
but
it
suggests
to
me
that
potentially
we
look
for
other
ways
to
record
those
incidental
contacts.
B
So
if
you're
recording
a
span
because
it's
sampled
and
you
have
some
links
that
is
effectively
recording
the
event
in
the
reverse
direction,
if
you
find
a
span
your
database,
it
matches
that
link
to
context.
You
can
create
a
link
from
that
span
or
to
that
span
from
the
thing
that
you
that
you
have
in
your
hand.
B
So
if
you're,
not
sampled,
but
you
start
a
span
and
you're
linked
to
something
that
is
sampled,
but
that
sample
thing
was
already
flushed
from
memory
now.
That
was
the
problem
that
you
were
describing
and
like
one
answer
would
be
to
like
write
another
span
record
or
span
continuation
record
or
span
addendum
record
or
log
saying
there
was
an
unrecorded
span
for
a
sample
context.
B
Unrecorded
span
link
for
sampled
contacts,
reference
record
I,
don't
know
like
I'm,
just
trying
to
say,
like
all
we
have
to
do
is
record
record
that
somehow
not
saying
we
want
that,
but
I
do
think
it
would
help
resolve
this
asymmetry
of,
like
the
the
sampling
sampling,
can't
guarantee
you
completeness,
but
what
we
really
really
want
from
sampling
is
to
know
when
something
is
incomplete
and
and
right
now
we
don't
know
when
you're
missing
events,
if
they
are,
if
they
happen
after
your
span
was
created,
which
is
the
problem
with
these
messaging
scenarios.
B
D
Thanks
Josh,
just
to
summarize
what
Kent
and
you
mentioned,
I
think
we
are
saying
that
hey
there
may
be
different.
Customers
may
want
different
things,
and
there
is
no
Universal
scenario
or
agreement
that
the
links
should
be
sampled
consistently.
Hence
the
specification
will
not
really
prescribe
a
little
bit
more
in
terms
of
hey.
This
is
how
you
should
do,
and
it's
up
to
people
to
build
their
own
custom,
Samplers
that
put
together
these
kind
of
probabilistic
and
non-probabilistic
thing
and
and
Achieve.
What
they
want
correct
is.
Is
that
a
fair
summary.
D
Yeah
the
I'm
I'm,
specifically
talking
about
the
first
section
in
your
response,
which
is
links
created
at
the
like.
The
link
information
is
available
at
the
at
this
time
of
spam
creation
right
in.
B
That
case,
yeah,
I
I,
think
I
can
frame
the
question
then,
and
then
let's
maybe
ask
someone
else
to
try
and
answer
it.
I
have
two
parents
I,
let's
put
suppose
that
I'm
a
new
span.
I
have
two
parents
that
are
not
root,
that
are
not
parents
and
I'm
going
to
create
links
to
them,
and
they
are
each
sampled
to
50
and
I
have
to
I.
B
Think
I
have
to
know
that
I
always
have
two
parents,
and
then
I
can
say
something
like
well
I'm,
creating
a
new
span
and
my
parents
are
sampled
at
50
each,
and
so
they
either
have
a
just
account
of
zero
or
two
and
somehow
I
can
combine
my
two
parents
probabilities
to
compute.
Something
probabilistic
about
myself
is
that
is
that,
where
we're
heading.
D
No,
no
I
was
kind
of
going
into
the
non-probabilistic
option
right
that
you
recommended
so
I.
Think
in
the
first
section
is
okay.
Whenever
you're
looking
at
links,
it
becomes
a
non-probabilistic
sampling
approach
and
so
the
adjusted
count
becomes
zero
and,
like
you,
you
set
the
P
value
of
six
and
all
that
stuff
right,
which
is
probably
okay.
I
mean
that
I
think
the
trade-off
is
you're.
Saying
Hey
I
want
that
more
consistency
across
my
linked
traces,
but
I'm
trading
off
that
ability
to
extrapolate.
D
And
if
you
were
to
do
like
a
purely
probabilistic
approach,
you
would
have
been
able
to
do
better
on
the
estimation
from
the
like
all
these
plan
to
metrics
and
all
of
that
right
so,
but
but
the
second
part
I'm
I'm
kind
of
trying
to
still
wrap
my
head
around
like
the
like.
Today.
The
model
doesn't
support
that
creating
links
outside
of
start
or.
B
Or
you
can't
just
kind
of
refer
to
the
biggest
reason.
Why
not,
which
is
that
like
I
may
be
linking
to
something
that
just
finished
and
it's
flushed
out
of
memory
already
and
and
I
think
you
raised
this
question
at
least
verbally
last
week
or
last
time
two
weeks
ago,
which
was
sort
of
like
yeah
I
mean
I'm,
linking
you
something
that's
perhaps
in
Flight,
perhaps
already
finished
and
I'm
sampled
and
for
some
reason,
let's
suppose
through
configuration.
I
want
to
say,
like
this
has
now
become
this.
B
This
thing,
I'm
linking
to
has
now
become
so
important
that
I'd
like
to
turn
on
tracing
all
like
retroactively
as
much
as
possible,
meaning
if
it
is
alive
and
it
start
continuing
to
make
child
contacts
I'd
like
to
begin
sampling
that
halfway
through
or
something
like
that,
I
think
that
was
what
was
question
was
whether
that
could
be
done.
And
my
my
answer
to
that
is:
let's
make
span,
links
bi-directional
and
you
can
add
a
link
to
a
span
after
it
starts
now.
B
I
might
be
able
to
add
a
link
from
a
sample,
expand
to
an
unsampled
span
and
and
detect
that
I
should
start
half
sampling
or
whatever
like
half
tracing.
It
would
still
have
zero
just
to
count
in
my
opinion,
but
you
could
turn
on
a
sampling
bed.
So
you
could
turn
on
non-probialistic
sampling
of
a
link
to
span
somehow
that's
kind
of
what
I
feel
like
we're
asking
about
or
looking
for.
D
I
see
kind
of
like
a
deferred
sampling
like
some
form
of
basic
tail
sampling
right.
C
B
That's
that's
what
I
hope
to
say
here
is
all
I'm
gonna,
try
and
I
will
without
the
action
here,
at
least
for
me,
is
I'm
going
to
follow
up
and
try
and
answer
johann's
question,
because
I
think
I
was
trying
to
to
give
him
a
piece
of
something
he
could
use
and
I
think
he
hasn't
seen
it
yet
so
I'll
work
on
that
I
I,
don't
feel
I
feel
like
this
is
quite
a
esoteric
conversation.
B
You
know
the
the
idea
of
turning
on
sampling
part
way
through
a
trace,
because
you
find
it
to
be
interesting
but
I,
but
this
is
like
not
uncommon,
like
if
you've
been
in
this
space.
Long
enough,
you
know
someone
who's
going
to
say:
I,
don't
want
to
trace
Everything
But
as
soon
as
that
fan
crosses
two
seconds
like
it's
become
errant
and
like
I
want
to
do
anything.
I
can
to
collect
anything
about
that
span,
but
only
when
it
crosses
a
threshold.
B
In
time,
for
example,
I
know
I've
seen
that
request
a
bunch
and
then
you
can
say
and
then
everything
from
then
on
will
be
traced.
It's
become
an
error
situation
anyway,.
D
One
last
comment
so
I've
seen
that
request
in
in
a
regular
Trace
right
for
getting
links
for
a
minute
like
a
regular,
hey,
my
Downstream
call
failed
and
I
want
my
that
thing
to
be
propagated
up
back
to
the
call
chain
and
all
that,
but
across
links,
I,
wonder
whether
that's
a
real
use
case,
because
links
by
definition
imply
different
life
cycles
right
or
different
lifetimes
of
those
traces
so
having
this
kind
of
a
something
that
is
linking
to
me
like
if
I'm,
depending
on
that
to
make
my
sampling
decision
I.
D
Think
I,
don't
know
because
usually
like
these
are
all
for
async
or
like
these
kind
of
messaging
scenarios,
where,
as
Kent
was
saying
like
like
the
lifetimes,
are
completely
different.
So
I
don't
know
whether
natural
fit
to
to
kind
of
depend
on
something
that
is
of
a
different
lifetime,
whereas,
unlike
in
a
in
a
single
trace
context,
they
are
within
the
same
lifetime.
D
You
know
that
because
the
Upstream
one
hasn't
finished
yet
until
at
least
in
a
synchronous
case,
so
I'm
not
sure
yet
whether
that
is
worth
solving
the
the
back
propagation
of
the
decision
across
leads.
B
I'll
go
through
here,
yeah
calling
back
propagation
is
probably
a
good
word.
That's
probably
one
I've
heard
before
as
well
awesome.
Thank
you.
Thank
you.
Okay,
any
any
comments
on
that
topic
before
I.
Try
to
move
us
back
to
the
agenda.
C
The
only
thing
I
will
just
note
so
I
don't
forget
it
later
is
because
we're
going
to
talk
about
the
power
of
two
problem
is
the
the
yeah
actually
never
mind
I'm
Gonna
Save.
Until
we
have
that
conversation,
oh
I
will
remember.
Okay,.
B
All
right
well,
then,
there's
this
one
that
popped
up
in
slack
and
the
summary
quickly
is
that
metrics
providers,
especially
the
Prometheus
ecosystem,
have
been
kind
of
waiting
for
open
Geometry
to
provide
these
span
exemplars
for
quite
a
while
when
you
get
down
to
it
or
something
a
little
tricky
about
it.
B
When
you
start
sampling,
because
exemplars
are
a
form
of
sampling,
of
course,
and
let's
see
foreign
this,
if
if
ever
we've
come,
you
know,
for
example,
Ottmar
has
worked
on
a
sampler
that
combines
an
exporter
and
a
processor,
because
you
need
to
do
some
sort
of
hard
rate
limit.
B
This
request,
I,
believe
runs
into
essentially.
The
same
type
of
thing
is
that
you
need
to
have
coordination
between
the
export
and
processor
and
buffer
things
differently
than
you
would
in
an
ordinary
Trace
export
pipeline,
because
this
is
basically
saying
that
when
the
histogram
chooses
an
Exemplar,
for
example,
you
might
want
to
okay,
okay,
start
tracing
this
or
you
might
prefer,
instead
to
just
say
limit
your
histogram
example
ours
to
trace
spans.
B
That
is
actually
the
position
that
someone
below
is
taking
is
don't
do
that
it
comes
it's
very
similar
to
what
we
just
discussed,
though,
is
like
you
know,
during
the
middle
of
your
untraced
span,
a
really
long,
histogram
event
happened.
The
histogram
wants
to
choose
it
as
an
Exemplar,
and
now
you've
got
a
thing
saying:
you've
got
an
unsampled
span,
but
I
want
to
record
it
and
maybe
I
want
to
like
start
tracing
it.
B
You
know
that's
that
type
of
thing,
but
basically,
and
if
you
read
down
you
know
this,
the
final
final
word
from
Josh
here,
other
Josh,
is
that
you
know
this
can
be
built
into
an
exporter.
You
can
buffer
your
spans
if
they
were
chosen
as
a
histogram,
Exemplar,
I
I
haven't
actually
dug
in
on
this
one.
I
just
want
to
make
sure
that
people
see
it.
B
In
case
you're,
aware
of
any
user
requests
involving
exemplars
basically
says:
if
you
want
to
do
tail-based,
sampling
and
you're
buffering,
then
you
can
do
this.
We've
seen
that
before
with
you
know
hard
great
limited
Samplers,
for
example,.
E
I
mean
you
can
do
that
also
consistently.
So
if
you
yeah,
if
you
know
that
a
Spam
is
Exemplar,
then
you
can
just
pick
yeah
simply
probability,
one
which
corresponds.
What
is
the
p-value
or,
as
you
get
zero
right,
yeah
and
then
it's
consistently
samples
for
sure,
and
that's
all
I
mean
this
is
a
valid
approach
and.
E
Actually,
in
consistent
sampling,
you're
free
to
choose
the
p-value,
you
can
also
use
attributes
for
your
choice
so,
depending.
E
No,
the
R
value
is
fixed,
so
it
is
a
constant
which
is
defined
at
the
root,
and
you
cannot
change
that
and
you
are
not
allowed
to
choose
the
p-value
dependent
on
the
r
value,
because
this
will
break
the
consistent
sampling
decision.
But
what
you
can
do
you
can
freely
choose
the
p-value
for
every
single
span,
so
you
can
choose
a
different
sampling
probability
for
every
spin
and
this
different.
E
This
choice
of
the
p-value
can
depend
on
span
attributes
which
are
available
when
the
sampling
decision
has
to
be
done
and
if,
if
it's
known
that
a
span
is,
is
an
example
when
the
decision
has
to
be
made,
then
you're
free
to
choose
a
p-value
of
zero,
which
corresponds
to
100
sampling
probability.
B
C
B
B
F
B
Yeah,
because
there's
the
sort
of
convention
is
to
collect
one
Exemplar
per
bucket
of
a
histogram
per
collection
cycle,
so
that
if
your
collection
cycle
is
30
seconds
long,
you
may
have
seen
a
thousand
histogram
observations.
Some
portion
of
those
would
be
sampled
and
you
can
have
it
you.
Basically,
the
idea
is:
there's
a
fixed
reservoir
one
per
per
histogram
bucket,
and
then
you
can
do
atmar's
fanciness
to
do
a
fixed,
Reservoir,
consistent
sample
I.
Think.
E
B
Yeah,
so
let's
Suppose
there
is
there's
an
R
value
and
you
want
a
sample
of
one
just
just
as
a
special
case,
because
that's
basically
what
Prometheus
does
want
is
the
special
case
of
a
consistent
sample
for
a
reservoir
of
one
just
choose
the
one
with
the
high
star
value.
E
I
have
to
think
about
it.
If
you
have
to
keep
if
there
are
multiple
with
high
star
value,
maybe
you
have
to
keep
either
all
of
them
or
not.
I
have
to
think
about
it
in
order
to
be
to
have
the
possibility
to
extrapolate
that
on
in
a
non-biased
way.
So,
but
I
have
to
think
about
that.
So
you
cannot
answer
that
immediately.
B
That
sounds
fair
well,
I've
put
it
I've,
put
a
link
to
this
issue
and
I
I
at
the
very
least,
will
try
and
say
something
about
consistency
and
support.
Josh's
other
response
here,
which
I
think
is
correct.
Does
that
sound
like
we've
done
enough
here?
If
you
have
any
thoughts
on
that,
maybe
we
can
Circle
back
next
time.
B
All
right,
I'll
make
sure
to
follow
up
with
after
the
meeting.
Just
briefly.
B
Okay,
so-
and
this
is
the
one
where
Kent
had
a
topic
ready
to
go
and
I-
have
something
to
say
as
well.
I
know:
I
I,
put
a
link
to
an
old
issue.
I
might
as
well
open
it
just
to
show
you
what
we're
talking
about.
B
This
was
right
after
we
got
our
probability,
stuff
kind
of
landed
and
and
finally
merged
and
I
I
made
this
half-baked
idea,
which
turns
out,
was
pretty
Half,
Baked
yeah,
the
idea
being
that,
if
you're
doing
probability
sampling
in
a
tail
tail
sample
or
in
the
collector
that
you
could
invent
our
values
after
the
fact
or
sorry
for
different
P
values
and-
and
there
were
lots
of
reasons
not
to
do
that,
I
think
Peter's
and
not
Mars.
Both
responded
with
gut
reaction.
B
This
is
a
bad
idea
and
I
think
it
led
to
the
suggestion
that
we
probably
want
to
maintain
independently,
like
if
you're,
if
you're
doing
after
the
fact,
random
sampling.
That's
that's
like
a
different
type
of
adjustment
than
a
consistent
sampling
adjustment
and
they
might
want
to
be
kept
separate.
I
raised
this
topic
in
connection
with
just
generally
this
desire
and
I.
Don't
have
an
issue
about
it,
but
it's.
B
The
issue
is
essentially
following
through
on
the
w3c,
the
proposal
to
add
seven
bits
of
definite
Randomness
which
gives
us
this
opens
the
door
to
non-power
of
two
probability
sampling
this
issue
here
that
I
just
showed
you
was
like
the
point
being
here-
is
that
these
can
these
collector
configure
these
collectors?
B
That
are
it?
Let's
suppose
you
know,
one
of
three
sampling
is
happening.
It's
after
the
fact,
whether
there's
a
p-value
or
some
other
valued
indicate
after
the
to
indicate
random
sampling.
B
The
the
the
idea
is
that
if
this
you
can't
when
you're
doing
head
sampling-
and
you
want
one
and
three
sampling
you
can
you
can
Pro,
you
can
probabilistically
choose
between
the
nearest
powers
of
two
one
half
or
one
quarter,
to
get
one
and
three,
but
when
you're
doing
this
sampling
in
a
collector
pipeline
as
a
tail
sampler,
you
cannot
do
that
anymore,
but
you
could,
if
there
were
seven
random
bits
of
our
value
baked
into
your
Trace
ID,
begin
to
do
consistent
sampling
at
one
and
three.
B
Baked
into
the
trace,
ID
I
believe
so
that
I
think
we
can
start
the
the
that
this
issue
is
independent,
but
I
think
people
would
like
to
be
able
to
encode
probability
to
encode
all
the
adjustments
that
happen
to
their
counts
as
they
collect
their
traces,
and
this
probabilistic
sampling
processor
in
The
Collector
is
going
to
continue
getting
use.
B
I
would
like
to
see
us
have
a
solution
here,
as
well
as
for
propagating
our
values
that
lets
us
avoid
propagating
our
values,
because
the
trace
ID
has
seven
bits
of
randomness,
and
then
we
only
need
p-value.
We
talked
about
this,
maybe
three
or
three
meetings
ago,
once
you
once
you
have
r
value
baked
into
the
trace,
Trace
ID
and
you
have
seven
bits
with
reliable
order.
You
can
then
have
seven
bits
worth
of
sampling
probability.
At
that
point,
your
p-value
could
be
replaced
by
something
that
has
I.
B
B
C
C
C
So
in
other
words,
we
do
an
integer
of
one
two,
three
four
up
to
ten
thousand
or
whatever
number
you
want
to
be
so
that's
the
model
honeycomb
has
and
has
been
using
forever
and
when
we
send
to
the
trace,
we
stamp
the
trace
with
the
sampling
over
the
field.
We
call
sample
rate,
which
is
a
bad
name
for
that
inverted
probability
effectively,
because
then
we
can
multiply
by
that
number,
the
the
traces
we
have
to
show
statistics
in
you
know
in
in
the
graphs
and
it
it
works
pretty
well.
C
Most
of
the
time
I
mean
there
are
places
where
it
misleads,
but
in
general
works.
Okay,
except
when
you
start
getting
into
aggregations
things,
get
weird
anyway
point
is
that
model
is
more
sophisticated
than
the
powers
of
two
thing,
but
the
powers
of
two
things
fits
neatly
into
it,
because
it's
just
one
two,
four,
eight
and
I
guess
the
question
is:
are
we
talking?
B
Yes,
I
think
so.
I
I
mean
one
of
the
reasons
that
led
me
to
post
that
that
issue
before
was
that
75
sampling
is
like
really
a
desirable
number
for
some
people.
If,
if
you
know,
if
they're
just
putting
a
sampling
in
to
cap
their
bandwidth
and
they
want
as
much
traces
as
they
can
get
somewhere
between
50
and
100,
because
sometimes
the
answer,
and
so
so
I
and
I
was
I,
mean
I
was
surprised,
but
that
was
literally
the
first
user
response
to
like
hey.
B
We've
got
probably
two
stamping
now:
okay,
what
if
I
do
75
so
yes,
this
idea
of
a
t
value
is
is:
is
the
idea
that,
among
seven
bits,
you
can
produce
a
threshold
that
accepts
80
by
by
choosing
the
value
that
is
80
of
2
to
the
56th
or
student
56,
plus
one
or
something
like
that?
B
And
what
I
mean
57
bytes?
So
it's
56
bits
y
sub,
two
yeah
six,
okay,
and
then
we
don't
need
an
R
value
and
your
and
your
T
value
is,
which
is
the
non-power
of
two
p-value,
would
would
convey
both
integers
as
well
as
floating
Point
numbers
arbitrarily.
B
With
56
bits
of
resolution,
it
does
create
that
that
glitch
that
people
are
worried
about
some
of
the
time,
which
is
that
your
your
adjust
counts
can
be
non-integers
like
with
75
percent.
Your
adjust
account
is
you
know,
it's.
B
C
B
Yeah
so
I'm
sure
that's
such
a
problem,
but
but
I
I
do
think
that
so
so
another
alternative
to
the
T
value
which
can
give
you
these
binary
numbers
is
to
have
a
c
value.
This
is
hypothetical,
but
C
value
would
be
the
literal
count.
So
if
your
C
value
is
two,
it
means
you
did
one
and
two
your
C
value
is
three
one
and
three
see
values.
Four,
you
did
one
and
four.
This
makes
your
your
sent.
Your
adjusted
counts,
always
an
integer,
but
it's
a
little
bit
more
mechanic.
E
I
mean
it's
it's
nice
to
have
I
mean
if
the
reciprocal
value
is
an
integer,
then
it
just
accounts
always
seemed
Tetris
right.
So
but
but
one
thing,
if,
if
you,
if
you
want
to
have
all
the
just
accounts
to
be
an
integer,
then
you
are
limited
to
sampling
rates
like
one
1
divided
by
2
1
divided
by
three
here
is
I,
can't
told
us
so,
but
if
you're
limiting
to
that,
then
you
have
already
a
huge
gap
between
one
and
fifty
percent
yeah.
E
C
I
mean
I
could
totally
imagine
that
number
that
we
currently
have
as
an
integer
becoming
a
float,
and
so,
if
you
specify
you
know
the
value
of
1.333,
that's
a
75
right
and
so
that
you
know
that
would
require
us
to
modify
our
plumbing
so
that
we
use
that
floating
Point
math
rather
than
integer
math.
All
the
way
through
our
system.
C
That's
work,
but
you
know,
might
be
the
right
answer
here.
So.
B
In
the
statsd
protocol,
which
is
kind
of
a
relic,
but
it's
still
alive
today,
we
have
an
example
where
the
open,
Telemetry
collector
has
support.
There's
a
receiver
written
and
is
receiving
sampled,
the
there's
a
fraction,
it's
a
decimal
fraction.
That's
that's
included
in
so
it's
a
floating
Point
number,
so
you
can
say
at
zero
point.
I
think
the
leading
zero
is
presumed,
but
you
can
say
at
point
one
and
it
means
one
in
ten
sampling,
but
the
the
hotel
receiver
has
no
way
to
express
integer,
non-integer
counts.
A
B
The
histogram,
for
example,
so
that
when
you
receive
a
histogram
sampled
with
a
non-integer
reciprocal,
it
is
it
has
nothing
it
can
do,
and
it's
basically
just
a
readme
statement,
saying
don't
do
that!
Stop!
Don't
configure
stat
C
with
not
integer
reciprocal
sampling
rates,
because
the
entire
system
down
the
stream
can't
handle
it
right
now.
B
I
think
that
that's
good
enough,
basically
saying
if
you
think
that
that
non-integer
sampling
rates
are
a
problem.
Don't
don't
do
that,
but
maybe
that's
not
enough.
B
But
the
proposal
is
that
we
we
move
forward
with
some
kind
of
document
that
says:
there's
a
non-integer
or
a
threshold-based
adjusted
count
mechanism
and
it
relies,
and
it
assumes
that
you
can
stop
sending
the
R
value
because
it's
baked
into
the
trace
ID
just
that
that's
the
the
whole
scope
of
this
and
and
then
I
guess
it's
a
separate
question
as
to
whether
a
tail
sampler
that's
doing
well.
B
Hotel
sampler
is
doing
consistent,
probability,
sampling
and,
and
it
sees
the
R
value
it
can
generate
its
own
p-value
at
this
point,
I
believe.
So
the
idea
is
that
if
you
have
that
seven
bits
of
bytes
of
Randomness
in
your
Trace
ID,
now
that
the
collector
pipeline
is
just
doing
consistent,
tail
sampling
and
it
can
output
a
t
t
value
to
convey
it
to
just
count
for
non-integer
sampling
rates,
which
is
exactly
what
I
wanted
to
get
on
this
conversation.
B
If
that
sounds
reasonably
believable
to
you,
then
I'm
I'm
content
and
I
and
I'm
not
saying
that
I
I'm
trying
to
assign
this
to
anybody.
But
it's
available
and
I
might
consider
this
more
important
as
the
months
go
on.
B
Cool
I
think
we
could
leave
it
there
for
now
at
least
I
don't
have
I
can't
commit
myself
and
I'm
and
I
have
to
think
about
this
a
bit
more
anyway.
B
D
The
Josh
would
that
be
a
Otep
or.
B
I
I've
lost
confidence
in
the
Otep
process,
although
it's
just
because
there's
there's
not
a
lot
of
success
like
people
are
there's
a
little
bit
of
a
stall
right
now
and
I
think
more
contributors,
core
contributors
like
me,
need
to
get
in
and
do
work
and
then
I
I
actually
have
held
today
to
do
catch
up
on
open,
Telemetry
I'm,
going
to
try
and
write
some
specs,
but
I'm
not
going
to
do
this
to
other
things
that
are
higher
priority
for
me.
B
But
if,
if
it
stalls
out
I'm
tired
of
it
and
I
want
I
would
propose
that
we
could
just
start
updating
this
spec
with
you
know
if,
like
this
is,
this
is
an
incremental
change
to
expect
that
we
already
have,
and
it's
not
so
it's
not
so
huge
and
it
you
know
it
does
presume
some
things
about
w3c
that
are
out
of
our
control,
and
so
we
can
just
say:
look
w3c
wrote
this
thing
here.
We
are
updating
our
spec
for
it.
Maybe
it
doesn't
need
no
tip.
B
Yeah
yeah
I'm,
trying
to
rescue
some
of
the
oteps
that
are
currently
stalled
and
I'd
like
to
see
some
of
them
merge
before
I
recommend
more
oteps.
B
Most
of
my
time,
I'm
working
on
this
Arrow
Apache
Arrow
column
representation.
We
have
a
really
nice,
Otep
and
I'm
trying
to
get
it
merged
and
it's
like
no
one.
No
one
wants
to
read
it.
So
I've
stopped
recommending
oteps,
but
that's
that
I
think
everyone
wants
the
sampling
stuff
and
time
is
coming.
B
All
right,
shell
tidbit,
let's
step
finally
out
of
support,
adding
support
for
p-values.
Finally,.