►
From YouTube: 2022-05-05 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
B
Wow
spencer,
if
you
have
something
prepared
for
today,
I
suggest
that
you
start
your
presentation,
it's
being
recorded,
so
hopefully,
josh
will
catch
up
with
that.
A
Yeah,
okay,
thanks
for
the
prompt,
so
yeah
what
I
have
is
still
so
recap.
Six
weeks
ago,
in
slack
in
the
sampling
channel,
I
sort
of
presented
this
like
my
conception
of
what
the
sort
of
prioritized
list
of
stuff
we
could
work
on
was,
and
that
pointed
to,
in
fact,
let
me
just
share
my
screen
actually.
A
So
that
pointed
to
this
message
here,
you
often
see
this
yeah
thanks,
okay.
So
this
message
here,
where
I
kind
of
conclude
that
the
thing
to
focus
on
would
be,
I
call
the
tier
sampler
configuration
and
what
I
meant
by
that
is
that
we've
been
discussing.
You
know
different
sampling,
algorithms.
We
could
do
in
different
positions,
either
within
sdks
or
in
a
collector
or
similarly
sort
of
either
agent
or
a
centralized
sort
of
position.
A
In
the
thing
we've
also
in
this
group
discussed,
like
multi-stage
sampling,
different
ideas
and
how
to
make
that
statistically
valid
on
the
user
side
of
it
or
like
more
vocally.
I
guess
I
should
say
I've
gotten
the
impression
that,
like
at
least
like
joshua
has
said
to
me,
as
well
as
in
github
issues,
that
people
are
pretty
keen
on
the
like
capability
to
for
either
for
sdks
to
retrieve
like
advisory
information
from
some
remote
place.
A
That's
one
consideration,
but
then
also
sort
of
subsumed
in
this
term:
well,
maybe
not,
but
so
jaeger
with
under
the
remote
sampling
umbrella,
bundles
up
some
ideas
about
like
adaptiveness
that
are
pretty
interesting
and
then
also
a
piece
of
software
called
refinery
from
the
vendor.
Honeycomb
also
has
this
notion
of
adaptiveness,
so
it's
not
quite
remote,
but
sometimes
they
intersect.
So
we'll
talk
about
that.
But
then
the
reasoning
went
that
if,
like
you
know,
sdks
and
a
collector
are
going
to
exchange
information.
A
What
should
the
format
for
that
information
look
like,
but
then
also
just
like
within
a
collector?
What
should
the
sort
of
specification
look
like
for
a
describing
some
sampling
to
be
done.
A
And
the
collector
has
processors
today
trace
processors
today
that
do
do
sampling,
and
so
part
of
this
is
like
looking
at
how
those
work
sort
of
their
expressive
power,
whether
they
could
be
expanded
to
include
some
of
the
like
more
like
sophisticated
capabilities
that
that
users
want
that
either
like
don't
exist
in
like
hotel
native
components
today,
but
you
know
maybe
exist
in
I
mean
there
are
a
couple
like
they're
relatively
popular,
it
seems
like
non-hotel
native
solutions
are
either
this
refinery
project
I
mentioned
earlier,
which
is
based
on
a
go
package
called
dyn
sampler,
which
demonstrates
some
more
flexible
sampling
methodologies
than
than
hotel
natively
supports.
A
A
And
also
aws
x-ray,
so
that
also
achieves
certain
objectives,
and
so
that's
kind
of
the
this
space
of
like
we've
got
some
hotel
native
stuff.
We've
got
some
like
kind
of
vendor
driven
stuff
that,
like
maybe
it
to
me,
it
seems
to
be
like
a
little
bit
ahead
in
solving
like
customer
needs.
A
Maybe,
unsurprisingly,
and
so
what
I
would
seek
to
do
is
incorporate
some
of
those
ideas
either
like
retrofit
them
into
like
collector
processors
or
like
sdk
collector
interfaces
like
communication
interfaces
or
or
like
create
a
new
thing
if
they
can't
be
easily
if
we
can't
easily
evolve
the
existing
hotel
native
components
to
accommodate
some
of
the
more
interesting
features.
A
So
that's
kind
of
the
origin
of
what
I
was
working
on
and
how
I
ended
up
interested
in
a
what
I
called
here
like
a
sampler
policy
data
model.
I
think
in
my
document
I'm
starting
to
call
it
just
like
a
sampler
data
model,
but
what
I
have
found-
and
this
is
what
I'm
eager
to
get
feedback
from
you
all
on
today
and
probably
in
the
future-
is
it's
difficult
to
it's?
A
really
big?
A
It's
like
a
really
rich
space
or,
like
very
highly
you
know,
high
dimensional,
and
what
I
mean
by
that
is
that
when
I
really
started
closely
looking
at
like
all
the
things
that
I
named
so
like
tail
sampling
processor
in
otel,
collector
x-ray
refinery
and
well,
those
are,
I
guess,
the
big
three.
But
there
are
a
lot
of
like
subtle,
oh
and
then
sorry,
jager,
jaeger's,
sort
of
remote
sampling
protocol
and
then
jager's
adapted
sampling
capability
and
like
comparing
and
contrasting
all
these.
A
There
are
many
sort
of
dimensions
or
characteristics
that,
like
these
are
differentiated
by,
and
that
is
part
of
the
difficulty
and
I've
like
arrived
at,
like
the
necessity
to
like
describe
this
like
sampler
data
model
thing,
and
I
don't
know
how
clear
it
is
or
I'm
finding
it
difficult
till.
I
draw
a
line
from
that
to
to
like
actually
solving
some
customer
problems,
even
though
I'm
like
convinced
that
this
is
the
first
step
to
do
so.
A
A
Okay,
so
I'm
gonna
jump
over
to
my
doc.
So
I,
based
on
this
group's
recommendation,
I
forked
the
totep
template
and
I
the
original
sort
of
heading
or
sorry
section
like
prompts-
I
think
some
of
them
remain
and
are
in
like
italics
or
maybe
not
all
italics,
but
so
maybe
I'll
give
you
a
moment
to
just
like
read
this
like
opening
section.
A
So
this
is
this
is
motivation.
Why
should
we
do
this?
Why
should
we
where
this
is
a
data
model
for
trace
samplers?
So
let
me
see
yeah
I'll,
just
put
this
much
on
the
screen.
So
I'll
give
you
a
couple
minutes
to
read
this
and
then
I'm
gonna
like
ask
people
how
people
reacted
to
it.
A
Okay,
so
again,
so
this
is
like
motivation
and
let
me
just
immediately
say
that,
like
there's
actually
a
lot
more
like
problem
sort
of
development
below
this,
but
then
last
night
I
was
like
revisiting
it,
and
I
thought
that
it
was
like
way
too
detached
from
like
sort
of
concrete.
A
A
lot
of
it
sort
of
makes
references
to
like
very
specific
shortcomings
of
other
systems,
and
so
I
don't
know
I
don't
know
one
question
I
had
was
like:
should
I
differ?
Should
I
just
like
assert
some
of
the
stuff,
but
then
the
nested
detail
like
defer
to
a
subsequent
section
or
or
what
so?
That
was
like
one
question,
but
I
don't
know
how
familiar
you
all
are
with
some
of
the
things
this
references,
but
did
anyone
have
like
a
a
reaction
to
any
of
the
statements
here.
C
I
don't
have
familiarity
with
much
of
this
outside
of
refinery,
but
I
didn't
find
that
this
was
too
much
detail
for
an
otep.
That
said,
I
will
caveat
that
with
having
not
read
a
lot
of
oteps,
but
I
feel
like
they
tend
to
be
fairly
in-depth
and
I
don't
find
this
going
too
far
but
yeah,
especially
if
there's
pieces
below
this
that
might
be
pushed
later.
I
think
this
kind
of
sums
things
up
well
enough
for
me.
B
A
Okay,
thank
you
both
for
the
feedback.
Ahmar,
do
you
have
any
thoughts
or
I
don't
actually
don't
know
what
your
focus
is
if
you're
like
building
stuff
with
open,
telemetry,
actively
or
yeah.
D
Not
really
we
I
mean
we
are
fine
now
with
the
specification
how
the
data
is
collected.
D
If
spans
are
sampled,
so
I
mean
the
configuration
is
of
course
also
a
topic
for
us,
but
you
know
it's
a
very
big
topic,
so
I
mean
it's
good
that
you're,
starting
with
that
yeah
and
I
will
of
course
yeah
you
follow
that
and
yeah,
but
maybe
just
I'm
almost
not
very
experienced
with
old
taps,
and
this
is,
I
think,
josh
can
help
much
more.
D
A
Okay,
thank
you
all
right,
so,
okay,
so
yeah.
A
This
is
indeed
a
sort
of
overview
with
an
intended
focus
on
like
acute
user
problems
but
like
like
then
noted
it
does
sort
of
rely
on
some
familiarity
or
like
it
just
throws
some
references
out
to
things
like
this
tail
sampling
composite
policy
that
you
know
might
not
be
familiar
and
so
yeah
on
the
subject
of
like
what
is
conventional
for
rotex
to
include
or
not
explicitly
how
I
actually
sort
of
started
this
and
now
some
content
that
I'm
second
guessing,
including,
is
what
follows
this.
A
A
Do
we
have
any
github
issues
on
you
know
stuff
that
we
might
work
on,
and
he
did
link
me
to
this
and
I
believe
william
has
like
attended
these
meetings
previously,
but
this
I
actually
was
kind
of
I
hadn't,
really
seen
this
until
this
morning,
and
I
was
a
little
bit
delighted
to
see
that
so
william
describes,
I
think,
he's
focused
on
sdks,
obtaining
like
the
remote
sampling,
sdks,
getting
information
from
somewhere
else,
and
he
cites
jaeger's
support
for
the
same,
and
then
he
discusses
you
know
aligning
on
a
sampling
rules,
format
and
publishing
that
api.
A
This
is
happily
intersects
quite
a
bit
with
the
broad
idea
of
this
otep,
so
that
was
neat.
There's
no
activity
in
this,
but
I
just
wanted
to
cite
that
and
I'm
unsure
if
there
are
other
sort
of
like
feature,
request,
nature
github
tissues,
but
I
would
like
to
reference
them
if
possible.
So
that's
that
one
okay.
So
as
far
as
this
is
where
I
start
to
like
tease
the
like
solution
that
this
will
propose
this
document
this
bottom
here.
A
A
And
then
propose
something
that
sort
of
subsumes
them
so
that
you
know
people
would
be
open
to
adopting
a
new
thing
because
anything
that
their
old
thing
did.
The
new
thing
also
supports.
So
if
you
know
the
term
domain
specific
language,
that's
kind
of
something
that
appears
later
in
this
paragraph.
So
it's
it's
more
akin
to
like
it's
difficult
to
describe
without
just
showing
it
so,
and
this
is
another
thing
I've
struggled
with.
A
So
these
are
also
rough,
but
the
concept-
and
I'm
like
jumping
way
ahead
into
so
pretend
the
problem
is
well
motivated
that
we
need
something
beyond
what
was
in
the
opening
couple
paragraphs.
One
thing:
I've
seen
that's
like
a
recurring
theme
is
that
a
lot
of
sorry
I've
named
now
like
jager,
x-ray
and
refinery.
So
those
were
my
sort
of
principal
influences,
slash
like
systems
whose
expressiveness
I'm
trying
to
meet
or
exceed,
and
how
many
of
them
work
is.
A
There's
often
like
a
sequence
of
like
an
ordered
collection
of
of
rules
of
some
kind,
and
each
of
those
can
like
match
the
trace
under
consideration
or
they
cannot
match
so
in
jaeger's
case,
for
example,
let
me
see
if
I
can
show
that
real
quick
actually
in
yeager's
case,
for
example,
this
is
a
json
representation
of
jaeger's
data
model
for
sampling
configurations
and
so
there's
like
a
certain.
A
Order
in
which
to
evaluate
this,
but
basically
every
every
trace.
So
this
is
this
is
this
is
essentially,
it
defines
a
a
function
over
root
spans
and
then
subsequent
services
are
expected
to
do
like
a
parent-based
sampling.
So
this
is
over
root
span.
So
I'll
say
that
this
is
like
defines
like
how
to
sample
a
given
trace
and
how
it
works.
Is
that
because
it's
advising
sdks
on
what
to
do
in
like
head
sampling,
it's
decisions
are
only
a
function
of
information
available
in
that
head
sampling
position,
specifically.
A
The
decision
is
based
on
service
name
and
what
they
call
operation
name.
I
believe
that
gets
translated
to
span
name
so
service
name,
which
is
a
hotel
resource,
attribute
service
name
and
spam
name
and
with
those
two
data
you
can
look
at
this
json
document
and
like
know
which
object
that
kind
of
looks
like
this,
so
it
says
type
and
maybe
param,
and
so
there
are
two
types
of
you
know.
There
are
two
leads
in
this
data
structure,
so
to
speak.
A
A
So
that's
how
jager
works
and
there's
a
certain
precedence
order
where,
like,
if
there's
a
rule
defined
for
both
a
service
name
and
a
span
name
like
the
most
specific
thing,
then
you
know
you
go
to
that.
If
it's
just
for
your
service,
then
maybe
you
go
to
that
if
it's
just
for,
if
there's
a
rule,
that's
defined
for
the
span
name,
but
not
your
specific
service.
Such
as,
like
a
catch-all
rule
for
health
check
spans
with
name
slash,
health
is
what
this
would
match.
A
Then
you
would
go
here
and
then
there's
like
a
if
nothing
matched
at
all.
Then
you
would
go
to
this
default
strategy.
So
there's
a
sort
of
there's
like
a
preference
or
a
a
priority
between
like
for
any
given
route
span.
A
There's
a
subset
of
these
that
like
might
match
it
and
then
there's
a
ranking
among
those,
and
you
pick
like
the
highest
ranked
or
highest
priority.
So
that's
how
jager
works.
A
You
could
imagine
how
that
would
be
reflected
here
so
ignore
this
kind
thing
in
a
moment
for
for
the
moment,
but
so
there's
this
sequence
of
what
I
call
children
and-
and
you
could
imagine
that,
like
conditions,
you
know
match
span,
name
and
service
name
and
how
this
works.
Is
that
the
first
with
this
kind?
A
First,
the
first
child,
whose
condition
where
this
string
is
some
kind
of
predicate-
and
I
I
don't
yet
have
a
strong
take
on
what
the
the
language
of
this
string
ought
to
be.
There
are
a
couple
options,
but
I
largely
want
to
punt
that,
but
a
common
feature
of
at
least
refinery
implements
this
concept
of
life.
A
You
have
a
set
of
conditions
that,
like
a
given
rule,
they
call
it
can
match,
and
so
you
could
imagine
how
the
expressive
power
of
this
any
jaeger
policy
you
could
recast
in
these
terms.
A
While
I'm
okay,
I'll
pause,
does
anybody
have
any
feedback
about
that
sort
of
assertion?
I've
made
that,
like
this
encompasses
jaeger
stuff.
B
Well,
there
is,
there
is
definitely
a
number
of
questions
that
need
to
be
answered
here,
but
what
strikes
me
is
that
you
are
using
adjusted
count
rather
than
probability
expressed
by
by
the
probability
value
which
I
find
well
interesting.
I
I
think
it
it
will
work
and
it
might
be
even
better
for
the
customer
to
to
use
that
kind
of
thing.
D
B
This
is
a
good
observation
here,
but
also
what
what
will
happen
if
there
is
no
active
child
at
all
really.
A
Yeah
what
I
had
been,
what
isn't
present
in
this
block,
but
what
I
think
would,
owing
to
your
observation,
like
what,
if
nothing
activates,
I
think
there
would
have
to
be
a
default
sort
of
thing.
You
know
sampler
rule.
That
is
the
chosen
one
in
the
event
that
much
in
playing
the
same
role
as
this
bit
right
here
yeah.
I
think
that
would
be
necessary.
D
B
You
you
will
probably
want
to
have
a
catch-all
kind
of
an
exception
that
that
will
take
will
take
action
if
there
is
really
nothing
that
matched
whatsoever.
So
it's
not
a
question
of
of
this
first
sampler,
but
but
in
general
you,
you
need
some
kind
of
default
global
default.
A
Did
anyone
else
have
any
comments
or
questions
about
this,
and
I
should
say
that,
like
this,
is
I've
not
yet
pivoted
to
talking
about
like
this?
What
this
kind
first
thing
is
and
how
what
we're
looking
at
here
is
more
like
a.
A
A
This-
is
the
sort
of
extensibility
point.
I
guess
I'll
say
that
so
like,
whereas,
whereas
like
honeycomb
or
x-ray,
sort
of
imbue
into
the
structure
like
oh
this,
like
imagine
imagine
for
a
moment,
you
don't
have
the
words
kind
and
children
here.
You
just
have
this
like
list
in
yaml
and
what
these
other
systems
would
do.
Is
they
say
you
have
this
list
and
it's
like
evaluated
in
this
way
in
order
to
accommodate.
A
Other
like
yet
unforeseen
ways
of
composing
samplers
and
what
I
mean
by
that
is
in
in
the
in
the
experimental
spec
right
now
when
it
talks
about
composing,
and
it's
talking
about
composing,
consistent
samplers
in
the
in
the
experimental
spec
that
joshua
merged
there.
It
talks
about
taking
the
logical
or
of
multiple
samplers,
but
it
occurred
to
me
when
I
was
working
on
this.
That,
like
there
are
other
sort
of
compositions
that
you
could
do,
I
mean
the
boolean
ones
like
or
is
an
obvious
one.
A
You
could
also
do
like
the
logical
and
of
two
samplers,
another
sort
of
composition.
In
fact
like
this,
this
formulation
is
looking
at
this.
Like
rule
evaluation
thing
as
a
sort
of
composition,
much
like
logical
or
logical
and
or
like
taking.
The
first
argument
of
you
know
the
first
one
that
matches
that's
active
is
what
we
say
here:
the
first
active
child,
that's
a
sort
of
composition,
and
so
that's
kind
of
what
led
me
to.
A
I
don't
yet
have
like
the
the
citations
that
I
eventually
want,
but,
like
scattered
across
github
issue
comments,
I
have
seen
sentiment
that
that
people
like
want
more
like
flexibility
and
we
haven't
gotten
yet
into
like
the
existing
tail
sampler
collector
processor,
but
like
that
is
like
not
super
usable
in
certain
ways
or
like
difficult
to
use.
A
I
mean
not
ergonomic
and,
and
so
that
sort
of
informed
in
my
head,
a
desire
or
requirement
to
like
be
able
to
be
extensible
to
a
certain
degree,
and
so
that's
where
this
like
kind
concept
came
about,
whereas
the
systems
that
do
have
like
a
notion
of
like
you,
have
a
ordered
collection
of
rules
and,
like
you
pick
one
according
to
these
criteria,
I
could
I
kind
of
want
to
like
save
space
to
like
support
other
ways
of
combining
multiple
samplers,
and
so
that's
where
this,
like
abstraction
of
calling
it
kind
came
in
where
a
kind
is
sort
of
if
you're
familiar
with
they're
like
ayanna,
like
they
have
these.
A
This,
like
registry,
for
different
kinds
of
things.
It's
this,
like
internet
body,
that
you
register
like
media
types
for
things
like
image,
slash,
jpeg
or
image.
You
know
application,
slash
json
like
those
are
called
media
types
and
you
can
like
right
register
those-
and
I
would
imagine
like
this-
is
sort
of
this
kind.
Thing
is
akin
to
like
a
namespace
of
those
that
you
could
add
different
kinds
of
things
to
that
would
compose
children
and,
as
peter
noted
children
themselves
can
be.
A
I
haven't
defined
this
yet,
but
like
each
of
these
is
I
call
it
a
sampler,
because
each
of
these
sort
of
is
each
of
these
nodes
in
this
tree
is
responsible
for
producing
a
decision
or
is
capable
of
producing
a
decision.
If
it's
activated-
and
then
there
is
below
here-
is
like
a
algorithm
for
evaluating
this
tree
given
given
a
trace,
so
so
that's
kind
of
some
of
the
more
abstract
ideas.
A
So
we
started
with
this
concrete
example,
but
I
haven't
yet
like
sketched
out,
so
I
don't
have
like
a
sort
of
a
grammar
that
describes
this,
but
that's
ultimately
like
the
artifact
of
this
otec.
As
I
currently
see,
it
would
be
like
a
sort
of
grammar
for
this
and
probably
defining
you
know
some
like
some
concrete
types
of
like
nodes
that
could
appear
in
this
tree.
So,
like
we
see
a
few
here,
we
see
first,
we
see
probability
and
these
names
can
all
be.
A
You
know
quite
shed
upon,
but
like
first
probability
and
that's
it
and
there
are,
you
could
imagine
that,
like
other
kinds
would
be
like
some
kind
of
rate
limiting
thing,
but
so
like
it's
an
extensibility
point,
this
is
the
idea.
D
The
condition
is
a
string
right,
so
it's
it's
just
very
generic,
so
that
means
you
need
to
parse
it
and
also
compile
it.
Somehow
I
mean
is
this
something
I
mean
you
could
also
have.
A
Oh,
I'm
sorry,
yes,
yes,
jaeger!
Does
that
yeah
right,
I
was
gonna,
say
also
an
even
better
or
more
faithful.
I
should
say
implementation
of
what
you're
describing
let's
see
if
I
can
find
it.
A
Yeah
so
like
something
like
that
yeah,
I
am
like
not
married
to
a
string
representation.
In
fact,
some
of
the
the
reason
I'm
still
considering
it,
I
should
say
actually
is.
I
was
working
on
something
recently
and
I
became
aware
of
I
mean
I
think
this
has
come
up
once
before.
Go
learning.
A
A
Like
here's,
you
know
a
package
to
basically
like
create
little.
You
know
mini
languages
for
like
business
rule
engine
that
sounds
cool
and
you
can
extend
this
with
like
custom
functions
or
whatever
I
think
len
is
built
in,
but
so
like
I'm
aware
of
this,
and
then
I'm
also
aware
of
there's,
like,
I
think,
a
very
similar
package
which
is
like
also
coincidentally
in
go,
but
it
is
also
do
they
have
any
examples
here.
Yeah,
it's
like
the
same
thing.
A
So
these
there
exists
a
couple
packages
that
like
could
be
leveraged
to
to
support
this,
maybe
but
like
I'm,
actually
not
a
go
programmer,
so
I've
not
deeply
evaluated
them,
but
yes,
this
could
definitely
evolve
to
be
like
an
actual
structured
thing.
A
Eventually,
I
would
yeah
I
would
prefer
to
for
that
choice
like
consider
yeah,
like
the
technical
feasibility,
slash
and
then
also
like
the
you
know,
ease
of
use.
I
guess
from
for
for
actually
authoring
these.
These
documents.
D
Did
you
have
maybe
one
one
thing
which
which
could
be
relevant
for
us
is
yeah,
I
mean
if
you
want
to
define,
for
example,
sampler,
and
you
define
a
key
like,
for
example,
this
the
span
name
or
something
like
that
some
attribute
of
the
span,
which
is
the
key
and
you
wanna
you
know,
sample
those
keys
with
a
high
frequency
with
a
smaller
sampling
rate
in
order
to
not
enter
a
rare
ones.
You
want
to
sample
with
100
probability
or
something
like
that
or
so.
A
D
And-
and
this
is
maybe
what
needs
its
own
configuration-
maybe
I
don't
know
yeah
I
mean
I'm
just
giving
you
some
input
there.
I
don't
know
how
to
solve
that.
Yes,
in
this
configuration
right.
A
But
what
you
described
actually
reminded
me
a
lot
of
so
earlier.
I
mentioned
this
go
package
called
dine
sampler,
which
is
an
ingredient
in
in
the
refinery
thing,
but
I
think
we
can
look
at
this
for
just
a
second.
A
So
how
one
of
the
I
don't
know
the
term
class
struct
one
of
the
things
exported
by
this
package
is
there
are
a
couple
different
objectives
that,
like
these
different
sampler
objects,
can
can
try
to
achieve,
but
the
one
you're
describing
is
pretty
similar
to
this
one-
and
I
think,
maybe
another
one
in
here
or
yeah.
This
is
sort
of
the
analogy
in
my
head
that
I
actually
like
is
like
dynamic
range
compression
from
like
audio
stuff.
A
If
you
are
familiar
with
that,
but
it's
like
the
more
like
really
high
frequency
things.
You
like
pull
down,
how
much
of
those
you
sample,
and
so
you
you
make
the
you
make
the
rates
of
the
more
common
things
and
the
less
common
things
more
similar
to
one
another,
and
so
I
could
imagine
that
being
supported
in
this
framework
in
the
form
of
a
different
kind.
Where
you
know,
maybe
you
send?
A
Maybe
you
don't
have
a
tree
and
you
just
have
a
single
node
with
no
condition
and
its
kind
is
a
sort
of
ema
like
a
one
of
one
of
these
things,
where
sort
of
internal
to
it.
A
It's
it's
doing
its
logic
to
like
inside
these
things.
There
is
some
state
to
track,
like
the
you
know,
observed
frequencies
of
different
cohorts
of
spans
where,
like
a
cohort
or
it
like
partitions,
the
rate,
the
the
population
of
like
incoming
traces,
it
partitions
those
by
computing.
A
This
thing
it
calls
a
key
and
a
key
is
like
the
value
for
a
set
of
like
span
attributes
basically,
and
so
that's
like
the
partition
key
of
a
given
trace
and
then
it
yeah
like
makes
the
more
common
things
less
common
like
it
makes
everything
relatively
more
similar
to
each
other
in
terms
of
volume.
D
We
also
have
something
similar
in
place.
I
put
the
link
in
the
chat
so
where
what
we
call
adaptive
traffic
management-
and
here
we
have
basically
also-
I
mean
it's
not
really-
to
open
telemetry
but
of
course,
yeah
it'd
be
nice
to
have
something
similar
in
open
telemetry
as
well.
A
Yeah,
thank
you
for
sharing
this.
I
was
not
aware
of
this,
so
I
will
study
this
and
see
sort
of
if
it
contains
some
like
features
that
are
potentially
you
know
not
represented
in
in
the
other
things
and,
if
so
I'll,
be
sure
to
try
to
work
that
in.
Thank
you.
A
A
I
have
a
good
like
I'm,
making
a
compelling
case
for
this,
because
there
already
is,
like
you
know,
jaeger
has
a
language,
and
why
can't
we
just
tweak
this
and
I
feel
like
I
have
to
back
up
a
long
way
and
say
like
what
broadly
like,
do
people
care
about
when
they
do
sampling
and
that's
like
you
know,
taking
100
steps
back
it
feels
like,
but
I
also
feel
that
this
stuff
is
pretty
important.
A
So
this
section
identifies
what
I
think
this
is
like
my
like
grand
unified
theory
of
like
why
people
are
ever
interested
in
sampling
and
then
so
I'll
give
like.
You
know,
30
seconds,
to
read
this
section
here.
C
A
So
one
of
the
something
I
want
to
know
is
like
has
anyone
ever
heard
of
a
like
use
case
or
concern
with
sampling
that
isn't
spoken
to
by
by
these
three
sort
of
categories.
A
And
the
reason
I'm
interested
in
that
is
because
this
is
a
pretty
sort
of
like
the
claim
that
this
is
like
a
spanning
description
or
like
a
sufficient
set
of
a
sufficient
description
for
like
why
people
are
ever
interested
in
sampling
is
used
to
like
underpin
some
later
claims
of.
Like
you
know,
if,
if
we
you
know,
this
is
all
that
there
is
to
solve.
Basically.
A
And
so,
like
don't
have
to
answer
right
now,
I'm
going
to
share
this
later.
I'm
probably
gonna
preserve
this
in
some
form,
but
the
whole,
like
the
original
formulation
of
this,
like
motivation.
A
Section
was
like
starting
from
this,
like
very,
like
user-centric,
like
requirements,
oriented
focus
and
then
making
a
like
a
really
detailed
argument
that,
like
the
stuff
that
users
have
access
to
today,
doesn't
balance
these
goals,
and
I
I
hope
that
that's
like
a
fair
way
to
to
frame
it
that
they're
not
being
balanced
because
as
it
turns
out
a
lot
of
the
tools
we
have
like
are
good
at
reducing
the
amount
of
data.
Like
you
know,
anyone
can
put
the
like.
There's
a
like
tails.
A
A
You
know
only
keep
five
percent
or
one
percent
of
things,
and
that's
extremely
good
at
solving
the
first
two
objectives,
or
at
least
it
reduces
it
in
a
sort
of
indiscriminate
way,
but
it
utterly
like
tosses
out
this,
like
sampling
error,
notion,
because
it's
going
to
weigh
oversample
some
things
and
weigh
under
sample
some
things,
and
that
is
sort
of
a
pattern
with
all
the
current
offerings
in
that
like
they
are
more
capable
with
some
of
these
one
or
two
of
these
things,
but
like
suffer
with
respect
to
the
third
thing
and
obviously
like
what
I
seek
is
a
system
that
can
like
simultaneously
sort
of
honor
each
of
these
requirements,
but
at
a
really
high
level
and
we're
not
going
to
get
into
this.
A
I
don't
know,
I
don't
know
if
I'll
keep
it
but
like
how
how
it
went
was
you
know,
there's
stuff
built
into
the
sdks,
but
those
do
head
sampling
which
have
this.
You
know
limitation
on.
They
put
a
low
ceiling
on
how
clever
you
can
be
and
like
in
pursuing
those
above
goals
trace
id
ratio
based
is
like
a
simple
hash
of
trace
id.
So
it's
totally
ineffective
at
like
minimizing
sampling
error
for
different
sub-populations
of
traces,
jaeger,
remote
sampler.
A
A
This
is
traces
per
second,
which,
like
is
not
ideal,
a
lot
of
trace
stores
work
in
like
spans
per
second,
and
so
it's
not
great
that,
like
there's
a
sort
of
impedance
mismatch
in
the
you
know,
units
in
which
the
limit
is
expressed.
That's
like
gonna,
be
a
recurring
thing:
jager
jaeger
rate
limiting
is
not
the
only
thing
that
suffers
from
that
so
yeah.
I
think
this
this
verbalizes.
What
I
just
said,
which
is
that.
A
No
I'm
gonna
skip
that
yeah.
The
collector
stuff
is
like
it
was
complicated
in
part,
because
these
components
actually
like
aren't
super
well
documented
in
terms
of
their
precise,
like
decision
making.
So
I
had
to
read
a
lot
of
source
code
to
understand
how
so
so
tail
sampling.
A
It
takes
this
like
big,
you
know,
list
or
array
structure,
and
then
you
know
much
like
actually,
where
I
have
kind
it
has
type,
and
then
it
has
certain
semantics
of
how
this
array
is
evaluated
and
one
of
the
elements
of
this
array
this
example
one
is
this
type
composite
and
each
each
type
each
these
are
called
policies
these
objects.
Each
of
these
has
a
certain
you
know,
based
on
the
type
its
own
evaluation
semantics.
A
Much
like
how
I
was
talking
about
with
my
thing,
but
the
most
powerful
of
them
is
this
composite
thing,
but
the
way
it
works
is
also
kind
of
weird,
and
I
I
like
eventually
worked
out
how
it
works,
and
this
isn't
really
documented
in
this
level
of
detail,
the
like
actual
decision
algorithm,
but
one
of
the
things
that
I
found
weird
about
it,
and
this
is
what
got
me
back
to
like
the
importance
of
the
like
those
three
like
those
three
goals
appear.
A
Is
that
the
way
that
composites
configure
you
actually
say
like
I
want
to
reserve?
You
say:
let
me
see
if
I
oh
shoot,
let
me
see
if
I
can
get
back
to
it.
You
say
I
want
some
number
at
most
of
spans
per
second,
which
is
good,
that's
spans
per
second,
that's.
A
I
think
what
most
people
would
prefer
compared
to
traces
per
second,
so
that's
good,
but
then
how
it
works
is
it
asks
you
to
these,
what
they
call
sub
policies
so
for
each
of
these
elements
you
sort
of
allocate
there's
this
matching
array
where
you
sort
of
allocate.
A
This
is
saying
like
to
the
policy
named
test
composite
policy,
one
I
want
to
reserve
50
of
my
global
span
throughput
for
this
one,
and
I
want
to
reserve
20
of
my
global
span
throughput
for
this
for
this
one
and
what
I
fundamentally
don't
get-
and
we
don't
have
time
to
like
look
at
the
prs
and
like
issue
discussion
that
that
brought
this
into
to
life
this
feature.
But
what
I'm
not
yet
clear
on
is
like
why
this
is
a
intuitive
or
sensible
thing
to
ask
users
to
do.
A
A
Certain,
like
subpopulations
of
of
traces,
I
mean
I,
I
guess
it's
a
valid
way
to
think
about
it.
Maybe
I
am
like
too
biased
slash
familiar
with
like
atmar.
What
you
were
referring
to
and
what
I
was
showing
like
honeycomb
and
it
looks
like
dynatrace
may
have
support
for
is
like
this.
Like.
A
Like
pulling
down
the
volume
of
high
volume
subpopulations
to
make
it
more
similar
to
just
like
decreasing
the
difference
between
like
certain
sub-populations,
but
I
guess
this
is
another
valid
sort
of
way
to
frame
like
traffic
shaping
or
it's
a
way,
but,
like
I
don't
know
if
it's
a
good
one.
So
that's
kind
of
my
like
unpolished
thoughts
on
this,
which
is
manifested
and
like
I
don't
know,
I
don't
know
that
anyone
actually
thinks
about
or
life
has
their
original
desire.
A
In
these
terms,
like,
I
suspect
that
people's,
like
you
know,
root
most
desire
is
like
not
in
these
terms
and
as
a
result
like
this
is
not
a
great,
not
a
great
interface,
but
I
don't
know.
I
think
I
like
need
to
speak
to
the
people
that
worked
on
that
to
understand
what
they
were
thinking
about
when
they
did
it.
That
way,.
A
A
A
It's
like
a
work
in
progress.
It's
yeah
it's
difficult
because
the
space
is
really
nuanced
or,
like
some
things,
do
some
technologies
their
products.
Do
like
one
thing
well
and
some
do
it
slightly
a
different
thing?
A
Well,
and
there
are
many
different
things
to
be
done
well
or
poorly,
so
it's
difficult
to
synthesize
in
like
a
concise
way,
I'm
finding,
but
that's
what
I
have
so
far
and
I'm
gonna
keep
working
on
it
slash,
I
hope,
to
share
like
in
slack
for
more
targeted
discussion
like
smaller
pieces
of
what
I
just
did,
so
we
can
focus
a
little
bit
more
on
parts
of
that,
but
I
know
that
was
long.
Thank
you
for
your
attention
and
your
time.
Everyone
and
I
hope,
to
share
more
with
you
later.