►
From YouTube: 2022-04-07 meeting
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).
A
A
Over
the
last
two
weeks,
spencer,
wilson,
who
was
present
in
the
last
two
meetings,
did
dm
me
some
rather
insightful
questions
about
adaptive
sampling
on
slack,
it
sounded
like
he
was
curious
enough
to
keep
working
on
it
independently.
A
Likewise,
amazon
is
pushing
on
x-ray
and
the
thing
that
brought
me
to
this
group
originally,
which
was
to
get
probability
sampling
in
to
place
so
that
we
could
count
those
spans
still
not
happening
and
until
it
gets
prioritized
for
me,
it
won't
happen
for
me
the
the
work
there
would
be
to
swap
in
probabilistic
samplers
for
the
parent
based
and
the
trace
id
ratio
based
samplers,
and
that
would
be
up
next
from
you
know
somehow.
A
Then
I
see
you
have
joined
us
and
you've
been
here
in
the
past.
Is
there
anything
that
you
particularly
came
to
talk
about.
B
Today,
I
didn't
I'm
I'm
kind
of
hanging
out
just
to
keep
an
ear
out
on
sampling.
It's
certainly
something
that
we're
pretty
interested
in
and-
and
I'm
excited
to
see
that
some
of
the
folks
that
have
been
using
honeycomb
sampling
pieces
have
come
here
and
and
trying
to
have
their
voice
involved
in
the
conversation,
yeah
spencer,.
A
B
Pretty
important
yeah
no-
and
I
can't
remember,
there's
someone
else
who
came
a
month
or
two
ago
who
was,
I
think,
also
a
honeycomb
user
as
well.
So
I'm
I'm.
B
We
are
looking
to
add
more
headcount
into
kind
of
our
sampling
space
and,
in
the
meantime,
I'm
kind
of
a
placeholder
here
to
just
keep
an
ear
out.
But
I
don't
have
enough
bandwidth
to
kind
of
come
with
my
own
agenda
just
yet.
A
I
my
my
company,
for
its
part,
has
been
got
its
head
down
thinking
about
what
it
really
wants.
I've
we've
finally
kind
of
recognized
that
the
led
step
story
has
been
no
sampling
for
a
long
time
and
we
we
know
that
that
doesn't
work
for
some
of
the
customers.
We
really
know
that,
and
that
doesn't
mean
that
we've
figured
out
our
story.
So
while
that's
happening,
I'm
enthusiastic
that
at
some
point
after
the
metrics
work
is
stable,
that
we'll
push
in
that
direction
again.
A
The
other
thing
that's
going
to
come
up
with
sampling,
which
is
not
actually
connected
but
does
bring
us
together.
I
I
originally
had
this
interest
in
var
up
sampling,
which
we've
discussed
a
little
bit,
and
it
doesn't
turn
out
to
be
very
useful
as
a
head
sampler,
so
I
haven't
been
able
to
deploy
that
that
toy
of
mine,
but
in
the
metrics
spec
world,
we've
stabilized
everything,
except
this
exemplars
thing
and
the
exemplars
concept
is
an
opportunity
to
apply
sampling
in
the
metric
space.
A
And
here
then
there
is
an
opportunity
you
could
imagine
using
something
like
fire
up.
Of
course.
Maybe
that's
not
what
we
want,
but
I
am.
I
think
that
I'm
not
speaking
for
my
vendor
right
now,
but
I
know
the
hotel
metrics
group
stabilized
its
metric
spec,
which
but
everything
except
exemplars,
and
the
reason
is
that
not
enough
people
have
done
exemplars.
A
So
when
it
comes
to
the
exemplars
draft
spec,
it
includes
basic
reservoir
sampling,
meaning
that
you
can
configure
your
histogram
to
save
save.
I
guess,
but
my
coffee
kid.
A
You
configure
your
histogram
to
save
example,
span
ids
and
trace
ids
and
the
the
current
spec
says
just
use
a
unweighted,
unbiased,
unbiased
reservoir
algorithm,
which
is
really
very
simple.
You
know
it's
like
the
donald
knuth
algorithm.
It
gives
you
unbiasedness,
but
it
doesn't
give
you
anything
else.
I'm
in
a
meeting
right
now,
willa
I
may
have
to
go.
A
So
I'll
say
that
before
I
go
on,
but
you
could
imagine
asking
the
exemplar
sampler
to
spread
your
examples
across
an
attribute,
for
example,
or
across
a
latency
bucket,
or
something
like
that,
and
one
way
I
know
how
to
do
that
is
to
give
weights
to
those
items
that
are
inverse
their
probability
and
that
works
pretty
well
and
you
can
you
can
use
something
like
rap
for
that,
but
you
don't
necessarily
need
that
you
can
stratify
and
you
can
do
other
things.
A
C
So
well,
sampling,
I
think,
definitely
makes
sense.
So
if
you
have
either
you
you
have
attributes
or
if
you
have
a
multi-stage
sampling
architecture
right,
if
you're
sampling
at
the
source
and
then
you're,
maybe
at
the
collect
again,
and
then
you
want
to
balance
the
sampling
rates,
because
maybe
the
then
then
I
think
well.
This
definitely
makes
sense
because
if,
if
examples
already
passed
a
small
sampling
rate,
they
should
be
preferred
over.
I.
A
C
Pictures
exactly
yeah
whenever
you're
merging
samples,
I
think
we're
out
this
a
cool
algorithm
to
use
for.
A
Yeah,
so
that
seems
like
it
might
be
on
the
horizon.
I
know
that
once
the
metrics
sdk
spec
is
sort
of
finished
and
the
ga
is
there,
the
at
least
the
google
side
really
really
wants
those
exemplars
and
I'm
not
sure
what
other
vendors
are
going
to
be.
Jumping
on
that,
but
I
think
it'll
come
you
know.
The
prometheus
landscape
has
also
adopted
exemplars
and
that
there
are
vendors
who
are
predominantly
prometheus
providers
moving
into
trace
and
that's
one
of
their
selling
points
is.
You
can
find
a
trace
for
your
high
latency
metric,
for
example.
A
C
Have
maybe
one
thing
I
mean
it's
not
a
high
priority
for
us,
but
you
know
I
have
a
prototype
implementation
in
java
and
it's
stuck
in
a
pull
request
and
there's
no,
I
don't
know
no
progress
there.
So.
A
A
Finally,
I
put
a
lot
of
pressure
on
that
okay,
so
we
should
turn
our
pressure
to
that
I
can.
I
can
probe
some
people
who
have
java
approval
powers
for
that
ben.
If
you
have
someone
in
the
in
the
honeycomb
space
with
java,
prowess
or
peter,
if
you
have
someone
there
yeah,
that
would
be
useful,
but
maybe
if
you
I
guess
you
probably
put
this
in
which.
B
Was
yeah,
which
one
were
we
talking
about,
find.
A
A
Okay,
I
think
that
that
is
something
we
know
we
want
to
push
forward.
I'll,
take
a
look
through
it
myself,
because
I
promise
you
I
didn't
read
java.
A
Okay,
oh
yeah,
you've
got
a
lot
here.
This
is
okay.
I
will.
I
will
make
an
action
item
for
myself
to
take
a
look
at
this
things.
A
Yeah
you,
I
see
more
more
types
of
sampler
than
I've.
Put
into
my
mind
that
I
can
see
what
they
are.
I
can
see
what
you're
doing
here
and
or
you're
doing
a
lot.
I
guess
my
the
I
will
look
at
this.
The
question
I'm
going
to
have-
and
I
imagine
that
others
have
is-
how
will
I
get
to
use
this?
What
I
mean?
A
The
reason
why
I
mentioned
that
lightstep
hasn't
really
figured
itself
out
here
is
that
users
come
in
often
an
individual
user
and
they
have
this
sort
of
perspective
of
I'm
running
my
one
service.
I
just
want
to
control
my
sampling,
but
when
you
come
in
with
a
pretty
large
sort
of
system
and
a
big
company,
that's
got
1000
engineers
or
you
know
a
million
lines
of
code.
That
approach,
maybe
doesn't
work
very
well,
and
I
think
you
know
you
don't
just
say:
here's
a
library
go
ahead:
figure
sampling
for
yourself.
A
There
sort
of
needs
to
be
a
big
picture
and
I
and
while
I'm
I'm,
not
sure
what
that
big
picture
is,
I
I
don't
know
if
I
could.
Even
I
basically,
I
don't
know
what
the
big
picture
is,
but
I
know
that
it's
not
that
what
we've
provided
and
there's
like
my
little
sampler
pr
and
go
is
just
like
two
classes
that
you
can
use.
It's
not
answering
the
big
picture,
that's
what
I
know
and
I'm
so
I
see
this
too.
I
see
a
bunch
of
sampler
classes
and
I
can
see
I
technically.
A
C
Yeah
I
mean
those
implementations
should
be
seen
as
some
kind
of
reference
implementations
right.
So,
of
course
I
mean.
D
C
A
lot
of
things
missing,
which
you
would
need
in
practice
because
yeah
we
talked
about
configuration,
for
example,
last
meeting,
and
but
I
mean
for
us
at
the
moment.
The
most
important
thing
is
that
you
know
if
there
is
sampling
that
we
get
the
values,
what
we
need
for
estimation-
and
this
is
specified
so
and
yeah.
The
implementation
of
the
samples,
of
course,
is
that's.
A
That
they
would
be
well,
they
haven't
done
it
yet,
but
but
that
they
would
be
welcome
to
just
go
ahead
and
do
whatever
they
want
for
sampling
and
they
can
use
powers2,
and
they
can
send
us
these.
These
three
states
to
help
us
count
them
no
one's
taking
that
up.
Yet,
as
far
as
I
know,
but
that's
definitely
something.
I
said
that
they
could
do.
A
D
C
D
A
That's
very
helpful,
then
thank
you
peter
yeah.
It
sounds
like
we
have
spoken.
We
will
get
this
we'll.
Do
this
I'll
get
someone
else
to
light
up
with
java
powers
to
also
chime
in
for
sure.
B
I'll
definitely
share
this
with
the
team.
We
have
our
own
kind
of
wrappers
of
the
sdks
that
we're
slowly
building
out
java
and
net
being
the
first
ones,
and
we've
had
to
do
our
own
kind
of
sampler
inside
of
our
wrapper
version
so
that
it
passes
the
right
information
along,
so
that
honeycomb
can
do
the
right
math
on
our
end,
and
I
would
much
prefer
us
not
having
to
have
our
own
special
ones
and
instead
being
able
to
use
ones
built
in
so
there's.
There's
some
nice
synergy
there
of
having.
B
A
A
Okay,
well,
we
have
one
new
guest
here
that
I
would
like
to
offer
an
opportunity
to
speak
wrong.
I
see
you
on
the
screen.
Would
you
care
to
introduce
yourself.
A
Don't
have
to
it
could
be
a
ron
g.
Sorry!
Well
then,
thank
you.
All
the
tasks
are
to
review
the
java
and
try
and
get
big
pictures
from
the
companies
we
work
for
and
to
bet
it
sounds
like
you
guys
have
one
of
the
more
concrete
needs
in
that
space.
So
that's
good.
That's
encouraging
and
maybe
spencer
will
come
back
with
some
some
thinking
that
he's
done.
If
he's
still
at
it
and
we'll
see
you
all
in
two
weeks.
D
A
Yeah,
I
I
I
will
he
he
was
dming
me,
but
that's
you're
right,
that's
not
appropriate.
I
and
I'm
I'm
not.
C
A
Limiting
factor
I
will
I'll
go
back
to
him
and
ask
if
he
would
mind
having
the
same
discussion
that
he
started
with
me,
especially
because
I
wasn't
answering
him
until
days
after
in
the
hotel
sampling
slack
channel.
That
would
be
a
start
we
could
also.
I
I
think
the
time
is
to
create
github
issues,
but
I'm
I'm
just
the
priorities
on
the
tricks
right
now
so,
but
I
will
thank
you
peter.
I
will
relay
that
to
him.