►
From YouTube: 2023-01-12 meeting
Description
OpenTelemetry Prometheus WG
B
C
I
spent
the
New
Year
in
Thailand,
I'm,
still
I'm
still
out
in
Asia,
so
it
is
very
late
here.
If
you
see
me
nodding
off.
C
E
C
Yeah
all
right
well
welcome
back
everybody,
I
guess
this
is
our
yeah.
This
is
our
first
meeting
of
the
new
year
and
yeah
I
guess
hopefully
we'll
be
we'll
be
off
to
a
good
start
here.
C
It
actually
kind
of
works
out
nicely
that
we
will
be
going
over
the
presentation
that
Felix
started
because
I
guess
yeah
that
can
kind
of
catch
to
those
who've
been
out
for
vacation
or
for
holidays
or
whatever
everybody
can
kind
of
catch
up
a
little
bit
as
we
go
through
that
and
review
that,
but
I
guess
yeah,
just
as
where
we
kind
of
left
off
last
year
was
We.
Were
you
know
we
kind
of
made
a
bunch
of
progress
towards
okay?
C
What
will
it
actually
look
like
to
implement
profiling
in
otel,
particularly
as
it
relates
to
the
collector
and
with
that
trying
to
be
somewhat
consistent
with
the
way
other
signals
are
handled
in
The
Collector
and
that's
sort
of
the
point
where
we're
at
right
now
is
trying
to
figure
out
how
we
can
you
know,
work
together
or
I
guess
get
feedback
from
The
Collector
Sig
to
figure
out.
C
You
know
how
yeah
how
profiling
fits
in
how
profiles
fits
in
that
kind
of
stuff,
so
yeah,
that's
kind
of
where
we
left
off
and
the
I
guess
we'll
get
a
little
bit
more
into
it
more
in
the
meeting.
C
But
before
we
do,
is
there
anything
yeah,
I
guess
if
you
have
anything
else,
you
want
to
add
to
the
agenda,
feel
free
to
go
ahead
and
add
it,
but
yeah
I
think
that
will
probably
be
the
bulk
of
this
is
kind
of
you
can
just
go
through
some
people
left
some
comments
on
the
on
the
presentation.
I
guess.
Obviously
we
need
to
give
the
collector
Sig
enough
context
so
that
they
can.
You
know
at
least
give
us
somewhat.
C
You
know
meaningful
advice,
maybe
not
in
this
first
meeting,
but
at
least
you
know
further
on
so
I
think
that
will
probably
be
the
bulk
of
this,
and
then
there
were
some
questions
in
slack
both
about
from
Pete
about
Spence
versus
profiles
and
then
about
the
common
Trace
format
that
somebody
added
in
there
yeah
anything
else
that
I'm
missing
or
anything
else.
Anyone
wants
to
add.
C
All
right,
yeah,
I
kind
of
give
a
little
bit
of
an
intro
I
guess
yeah
Felix
you,
you
know,
obviously
spent
some
time
thanks
for
putting
together
the
the
presentation
that
is
or
I'll
paste
the
meeting
notes
in
the
chat
by
the
way.
For
those
who
don't
have
it,
you
want
to
kind
of
yeah
I,
don't
know,
maybe
yeah
it
might
be
helpful
just
to
kind
of
go
through
it.
C
You
know,
maybe
in
a
little
less
detail
than
you
might
for
the
collector's
sake,
since
obviously
we're
yeah
pretty
familiar
with
most
of
it,
but
yeah
I
guess
maybe
just
like
the
sort
of
where
your
your
mind
or
mindset
was,
as
you
were,
building
it
and
you
know
what
you'd
like
feedback
on.
B
Yeah,
let's
do
that
and
at
any
slide,
if
somebody
has
a
question
or
wants
to
discuss
it
just
interrupt
me
and-
and
we
can
discuss
a
slide
and
I
think
there's
at
least
one
slide,
where
I
have
some
thoughts
that
I
would
like
to
to
raise
and
make
sure
we're
aligned
but
yeah.
So
this
is
a
slide
deck.
Everybody
can
see
this
right,
yep
so
yeah.
B
Basically,
this
is
just
a
quick
summary
of
sort
of
our
group
and
when
we
started
our
meetings,
meeting
notes
the
vision
documents
that
we
published
in
September.
B
So
that's
just
a
quick
intro,
then
I
was
trying
to
summarize
from
the
profiling
Vision
Otep
kind
of
the
most
Salient
points.
If
somebody
feels
like
these
things
are
not
the
ones
and
we
should
do
others,
please
comment
but
yeah.
Basically,
we
want
to
do
stuff
in
production
with
profiling.
B
We
kind
of
are
interested
in
two
different
types
of
profilers.
One
category
is
kind
of
the
whole
host
profile
or
the
others
like
so
there's
some
more
runtime
based
profilers
that
get
data
from
runtimes.
B
We
do
care
about
I.
Think
most
people
here
care
mostly
about
CPU
profiles,
but
I
think
we
all
agree
that
hotel
should
probably
care
about
all
profile
types
in
the
long
run
minimal
overhead.
We
all
agree
that
the
less
CPU
memory,
Network
Etc
our
Solutions,
will
use
better.
B
We,
we
are
aligned
on
that.
We
should
have
a
good
answer
to
dealing
with
popular
Legacy
formats,
such
as
piproth
and
GFR,
and
that
basically
ignoring
them
is
not
an
option.
The
the
hotel
architecture
or
or
things
we
come
up
with,
should
address
what
we
do
with
these
formats
and
how
they
fit
in
and
I.
Think.
The
other
big
point
of
why
we're
even
doing
this
is.
B
We
do
want
to
eventually
correlate
profiling
data
with
other
signals,
such
maybe
tracing
being
the
most
important
one,
but
Hotel
generally
is
all
about
the
signal
correlation
and
then
the
other
more
details
here,
yeah.
If
somebody
has
thoughts
on
adding
removing
please
erase
them
otherwise,
okay,
it
can
also
do
that.
Async.
A
B
Yes,
and
for
the
sort
of
landscape
in
2022,
there's
like
a
lot
of
vendors,
doing
profiling
stuff
now
and
there's
a
lot
of
Open
Source
projects
and
I
apologize.
I,
probably
don't
have
all
the
vendors
and
definitely
not
all
the
open
source
projects.
But
if
somebody
feels
like
I
miss
an
important
name
here,
please
just
go
comment:
I'll
edit,
then
yeah,
maybe
the
two
most
popular
formats
in
Industry
right
now
for
profiling
or
pre-brough
and
GFR
I-
think
we
all
agree
on
that.
Maybe
there's
something
I
missed
but
I
think
there's.
B
These
are
the
major
ones,
but
there
is
a
sprawling
landscape
of
like
minor
formats
and
Mark
Henson
has
done
a
great
job
with
profilapedia
to
to
kind
of
survey.
What's
out
there
and
I
think
that's
important
to
mention,
and
we
obviously
all
like
really
are
fans
of
what
the
elastic
and
optimized
folks
have
done
with
their
profiler
wiring
format
that
that
could
be.
B
If
we
build
something
in
hotel
ourselves
as
a
format,
be
a
direction
you
want
to
take
and
yeah
it's
an
efficient
protocol
that
resonates
with
the
hotel
working
group
I
think
that's
fair
to
say
that
most
people
here
think
there's
a
lot
of
good
ideas
in
that
approach.
To
sending
profiling
data,
then
yeah
the
data
I'm,
going
to
cover
this
section
quickly.
Maybe
I,
don't
know
how
many
collector
six
though,
but
basically
profile,
is
just
a
stack
of
a
set
of
Stack
traces
with
Associated
counts
and
weights.
B
There
can
be
timestamps
on
it.
If
sector
is
just
like
a
list
of
program
counters
usually,
but
it
could
also
be
civilized
or
just
be
symbolized.
If
it's
from
a
dynamic
programming
language,
you
won't
have
programmed
condos
necessarily-
and
maybe
you
also
have
metadata
like
labels,
which
we
could
use
for,
spend
and
Trace
IDs
and
the
most
common
visualization
is
a
flame
graph,
and
this
is
just
like
a
quick
summary.
I
won't
go
into
detail
but
like
how
a
flame
graph
is
constructed
from
this
data.
B
Another
visualization,
that's
interesting
is
the
flame
score.
Visualization.
That
sort
of
makes
takes
advantage
of
having
timestamps
on
each
sample,
and
then
you
can
kind
of
see
spots
where
a
lot
of
CPU
usage
happens
and
drill
down
into
them
on
a
sub
second
range,
which
is
interesting.
B
I
do
mention
our
tracing,
Coalition
I,
know
I.
Think
pyroscope,
you
guys
have
one
too.
You
can
put
a
screenshot
of
your
thing
here
too,
and
we
can
put
some
side
by
side
if
you
want
I.
Yeah
just
wanted
to
highlight
that
tracing
correlation
is
already
being
done
with
profiling
data
and
that's
something
we
want
to
continue
doing.
If
we
do
hotel,
then
yeah
people
just
a
quick
introduction
to
the
format.
It's
a
an
open
source
project
under
the
Google
org.
B
B
The
whole
debate
we
had
about
that
and
yeah,
most
notably
the
go
runtime
produces
p
profs
and
it
also
bundles
the
people
of
fewer,
so
P
Prof
is
very
entrenched
in
the
go
ecosystem,
that's
fair
to
say,
and
then
there's
a
large
ecosystem
of
tools
and
converters
around
P
Prof,
which
is
also
of
interest
to
us.
Then
here's
just
a
quick
visualization
of
the
protocol
format
I'm
not
going
to
go
into
details
and
then
yeah
JFR
is
kind
of
similar
story,
but
for
Java
it's
part
of
the
jvm.
B
It's
a
generic
event,
recording
format,
that's
not
limited
to
profiling.
I
think
that's
important
to
say
so.
It's
more
bigger
scope,
I
would
say,
but
there's
no
official
specification,
there's
a
whole
new
community,
reverse
engineering
of
what's
inside
of
the
format,
but
it's
essentially
also
stack
traces
symbols,
I
think
they're
always
present
I,
don't
think
it's
optional
for
Java
as
it
just
for
go.
Somebody
could
correct
me
on
that.
B
Timestamps
there's
runtime
events
like
stuff
from
the
jvm
itself,
but
there's
also
user
defined
events.
So
it's
like
kind
of
an
event
Channel,
where
you
can
throw
anything
in
there
that
you
want
which
is
kind
of
interesting,
and
similarly
there
is
an
ecosystem
of
tools
and
converters.
It
might
be
smaller
than
the
people
of
ecosystem,
but
there's
a
bunch
of
things
out
there
so
yeah.
Yes,
the
picture
I'm
not
going
to
go
into
details
and
then
yeah
the
portfolar
wiring
format.
B
Maybe
the
elastic
folks
here
can
comment
if
I
captured
that
correctly
but
yeah,
it's
basically
the
format
developed
initially
optimized
in
our
elastic.
There
is
this
document
you
all
put
together,
which
goes
into
great
detail
and
that's
what
people
should
read
as
they
want
to
understand.
It
I
think
highlighting
that
it's
kind
of
a
grpc
protocol,
rather
than
a
file
format,
is,
is
very
important
here.
B
Stack
traces
are
hashed
and
transmitted
only
once
generally
future
account
increments
on
the
stack.
Trace
only
refer
to
the
stack
Trace
by
hash,
ID
oops,
wrong
buttons.
B
Symbols
are
typically
also
sent
only
once
like
during
Ci
or
something
similar,
so
you
don't
send
them
all
the
time
and
all
of
these
things
together
kind
of
result
in
what
we
believe
is
significantly
nor
networking
overhead
I
know
Sean.
You
were
at
some
point
exploring
how
much
that
difference
is
and
trying
to
quantify
it.
I,
don't
know
if
you
continued
on
that
thread,
but
I
I,
don't
think
anybody
needs
convincing
that
this
could
be
pretty
efficient
and
yeah.
B
This
could
serve
as
an
inspiration
for
the
new
hotel,
profiling
protocol
working
title
potentially
Oprah
and
I.
Think
we're
on
the
same
page
on
that.
B
This
to
all
of
us
seems
to
be
the
most
interesting
technical
direction
right
yeah,
then
the
architecture
goals
we
do
want
to
create
compatibility
between
client
exporters,
aka,
the
profilers
and
the
backend
receivers
exporter
and
receiver
Hotel
terminologies
I'm,
trying
to
to
use
that
a
little
bit
to
show
that
we've
looked
at
their
stuff
clients,
May
emit
Chief,
RP,
Prof
or
okraf,
which
doesn't
exist
yet
so
we'd
like
the
ability
for
clients
to
just
emit
that
directly,
and
we
would
also
like
the
ability
for
backends
to
only
support
a
subset
of
these
things.
B
B
Clients
should
not
be
forced
to
convert
between
formats
because
of
the
overhead
system
course
I
think
I
mentioned
this
before
for
go
I
feel
less
strongly
about
it.
I
think
it
can
be
done
relatively
efficient,
but
for
Java
I
think
it
would
be
deal
breaker
to
convert
three
of
ours
on
the
client
and
yeah.
The
question
is,
then,
how
could
we
make
it
work
and
so.
C
And
just
a
quick
question,
so
I
guess
that's
the
first
point
where
it's
kind
of
like
am
I
I,
guess,
right
and
understanding.
That's
the
first
point
where
it's
like
specifically
for
The
Collector
Sig,
like
how
can
we
make
that
work?
I,
guess
yeah
I'm
wondering
like
at
what
point
we're
like
okay,
where
we
shift
from
like
okay.
This
is
like
background
context
to
like
this
is
where
we'd
your
help,
yeah.
B
F
B
Can
make
this
more
clear
somehow,
when
I
present
yeah,
then
sort
of
like
the
ideas
we
had
is
like
one
idea
that
I
first
clicked
around
in
a
document
was
kind
of
The
Collector
conversion
idea
where
everybody
sends
whatever
they
want,
and
it
goes
to
the
collector
which
converts
it
if
needed,
and
then
it
ends
up
in
all
these
other
places.
B
That
could
be
done.
There
could
be
some
content
negotiation
in
here,
but
maybe
that's
not
needed,
but
one
thing
that
is
important
to
note.
If
we
go
for
such
an
architecture,
if
there
is
no
collector,
then
basically
the
compatibility
between
these
components
will
be
limited,
so
some
stuff
will
be
able
to
talk
to
some
back-ends,
but
some
communication
will
not
be
available
based
on
the
formats
being
sent
and
understood
by
the
receiving
end
and
yeah.
So
the
and
I'm
thinking
about
renaming
the
slide
actually
from
content
negotiation
to
one
signal.
B
So,
basically,
if
we
have
one
signal
that
needs
to
be
somehow
an
understanding
of
like
sort
of
the
sub
formats
that
are
being
transmitted
and
if
they're
accepted
or
not-
and
this
could
be
propagated
from
The
Collector
to
the
clients-
we
haven't
discussed
it,
but
I
guess
clients
could
also
just
send
something,
and
then
they
get
an
error
back.
If
something
in
the
chain
is
not
happy
with
converting
or
dealing
with
it,
maybe
that's
simpler,
but
yeah.
B
Basically,
the
idea
is
that
you
have
one
signal,
like
everything
goes
kind
of
through
the
same
channel
and
The
Collector
can
make
compatibility
work
if
there's
stuff
being
sent,
that's
not
understood
by
the
receiving
end,
and
this
will,
however,
create
situations
where
certain
clients
that
call
themselves
hotel
and
certain
back-ends
that
call
them
self
Hotel
would
not
be
compatible
with
out
a
converter
in
the
middle
so
that
yeah
go
ahead.
Okay,
you're,
muted,.
B
D
Just
one
question
actually
related
to
the
The
Collector
as
well,
very
cold.
When
we
were
discussing
this
previously,
we
had
kind
of
discussed
the
idea
that
for
a
stateful
protocol,
the
the
collector
would
have
to
be
pretty
complicated
to
be
to
deal
with
that
statehood.
Associated
pushing
a
lot
of
say
the
back
end
Machinery
into
the
collector,
and
that
was
kind
of
my
only
kind
of
comment
when
I
was
reading.
The
the
deck
is
around
this
section.
Where
we're
proposing
the
we've
mentioned
kind
of
Argus.
D
We
mentioned
the
pros,
obviously
of
having
the
the
converter,
but
we
don't
I
think
mention
that
particular
con
and
I,
to
my
mind,
I
think
that's
probably
the
most
severe
con
of
that
one
I'm,
not
sure
what
you
guys,
I.
B
Forgot
about
putting
it
in
there
to
be
honest,
it's
good
that
you
mentioned
it.
I
will
try
to
find
a
spot
to
put
it
in
I.
Don't
think
anybody
thinks
that's
controversial.
To
mention
that
cool
Alexa.
Do
you
still
have
your
hand
up
or
just
by
accident,
okay,
cool
yeah,
then
yeah
I'll,
probably
rework
this
a
little
bit
based
on
that
feedback
as
well
and
and
the
idea
that
maybe
the
content
negotiation
is
not
the
most
important
thing.
B
Similarly,
if
you
build
a
components
that
receives
profiles,
you
specify
which
of
CCU
support
and
then,
if
you
don't
have
a
collector
in
the
middle,
you
kind
of
as
a
user
intuitively
know
that
if
your
signals
are
not
overlapping
between
the
client
you're
trying
to
use
in
the
back
end,
then
there's
no
compatibility
and
that
creates
a
little
bit
less
confusion
about
what's
compatible
and
what
isn't
and
there's
a
more
explicit
way.
But
in
a
way
it's
also
kind
of
a
yeah
easy.
F
B
Of
the
problem
of
creating
compatibility
right
so
I
don't
know
if
it's
gonna
actually
be
the
way
to
go,
but
yeah.
That
being
said,
having
separate
signals
doesn't
mean
that
the
collector
cannot
help.
So
there's
still
in
that
idea,
you
can
still
have
to
collect
or
do
conversions
and
and
make
potentially
incompatible
Hotel
pieces
from
one
signal
to
the
outdoor
compatible
and
arguably
that's
something.
A
hotel
is
already
doing
right
because
they
have
exporters
or
at
least
receivers
that
can
take
stuff
like
the
Prometheus
data
for
metrics.
B
Even
though
that's
not
the
preferred
Hotel
format,
and
then
this
compatibility
gets
created
so
yeah
I,
guess
I'm
still
a
little
fussy
on
sort
of
when
something
is
considered
to
be
Hotel
compatible
like
if
it
has
to
be
end-to-end
compatible
or
through
the
collector,
compatible.
I.
Think
we've
heard
mixed
messages,
I
think
Josh
said.
If
you
need
a
collector,
then
the
two
things
like
if
two
things
can
talk
to
each
other
without
a
collector,
it's
not
a
hotel,
but
that's
one
of
the
main
questions
for
the
collector
group.
B
If
they
agree
on
that
statement
or
if
that's
more
fussy
enough
for
debate
but
yeah,
so
basically
the
recap
is
we're
trying
to
find
a
good
architecture
for
profiling
as
a
new
signal.
Given
the
challenges
that
JFR
and
people
are
too
relevant
to
ignore
as
formats
and
the
conversion
of
the
client
is
potentially
too
costly
to
consider
and
the
new
Opera
format
might
make
people
of
redundant
in
the
future,
maybe
but
defining
a
super
set
of
Chip
are
like
targeting
things
that
Oprah
should
do.
B
Everything
that
shave
R
does
right
now
would
maybe
be
too
complex
and
brittle
like
something
that
we
don't
want
to
do
like.
Even
if
we
were
willing
to
absorb
the
complexity
now,
shave
R
could
evolve
in
a
way
that
then
we
couldn't
fit
the
data
into
what
we
designed
anymore
and
I
think
that's
in
evolutionary
problem
that
that
I,
don't
think
there's
any
good
answer
to
like
Chef
horse
just
wants
to
be
its
own
thing
and
not
be
controlled
and
yeah.
B
G
Again,
or
is
it
yes,
yeah
it's
up
again,
yeah
on
on
GFR
I
GFR
is
generic
enough.
So
it's
it's
hard
to
imagine
that
it
would
evolve
on
the
data
representation
because
it
already
can
express
just
about
everything
but
I.
At
the
same
time,
I
agree
that
it
could
evolve.
For
example,
in
the
direction
of
streaming
like
JFR
could
become
more.
G
And
introduce
some
concepts
of
this
like
incremental
state
or
reuse
or
like
hashing
things,
and
things
like
that.
So
I
think
that's
a
good
point.
I
have
an
Hotel
architecture.
Question
is
collect
Hotel
collector.
Is
it
per
container
per
machine
or
per
something
larger
than
that?
Like
per
cluster
like
where
exactly
Hotel
collector
runs.
G
On
one
slide,
we
mentioned
that
clients
should
not
be
forced
to
convert
between
formats
because
of
overhead
and
I'm
curious,
like
if
the
conversion
happens
in
The
Collector
how
exactly
this
amortizes
the
overhead,
because
if
the
collector
is
per
machine
well,
the
conversion
still
happens
on
the
machine.
Maybe
the
overhead
is
not
in
the
like
server
process,
but
it's
yeah.
B
So
I
think
there's
two
answers
to
that.
The
answer
one
is
that
the
ideal
state
for
like
somebody
deploying
profiling
is,
of
course,
that
they
have
a
channel
that
doesn't
involve
the
conversions.
B
The
conversion
is
kind
of
like
the
the
fallback,
if,
if
you
can't
create
compatibility
directly,
so
in
that
case,
I
think
the
overhead
is
a
little
bit
more
acceptable
because
it's
not
by
default
like
if
you
have
a
better
setup
with
a
exporter
and
receiver
that
understands
this
Improvement,
it
won't
happen,
and
the
second
thing
I
had
a
second
thought
there,
something
you
mentioned
actually
earlier.
G
Like
languages
like.
B
Yeah
having
to
implement
it
only
once
having
to
do
the
conversion
only
in
one
language,
I
think
that
was
the
second
Point
but
yeah
I
know
it's
a
fair,
fair
criticism.
G
Yeah,
maybe
maybe
it's
worth
expanding
on
what
the
overhead
is.
I
also
can
also
add
a
comments
so
that
you
don't
you
you
can
see
if,
if
there's
something
you
can
massage
in.
B
Yeah,
and
also
like
how
you
deploy
my
understanding
is
how
you
deploy
The
Collector,
like
if
you
really
don't
want
to
overhead
on
the
same
host,
you
can
obviously
put
it
on
a
separate
host.
That's
an
option
for
deployment
on
the
host
level.
You
also
get
resource.
B
Isolation
features
on
the
hosts
that
you
might
not
have
inside
your
application,
so
you
can
say,
like
hey,
I'm,
only
willing
to
give
this
scheduling
priority,
or
this
amount
of
CPU
or
memory
to
The
Collector,
whereas
a
new
application
like
go,
for
example,
does
not
allow
you
to
do
your
own
priorities
or
resource
isolation.
Things
inside
of
the
language.
B
We
should
ask
the
group
I
I,
I
I,
know
that,
like
data
docs
agent,
that
we
have
is
used
similarly
to
The
Collector
with
some
understanding
and
I
know
that
we
have
customers
who
do
all
kinds
of
setups
like
this.
So
I
would
expect
just
because
you
can
that
people
do
that
with
the
collect
on
the
hotel
as
well.
But
we
should
ask
the
six.
They
have
more
experience.
Yeah.
G
B
H
Yeah,
so
just
to
expand
a
bit
on
the
the
staple
versus
stateless
protocols,
I.
Think
more
generally,
in
terms
of
presenting
this
to
the
The
Collector
Sig,
making
explicit
some
of
the
implications
of
this
in
a
slide
might
be
useful.
So,
for
example,
the
the
statefulness
of
the
protocol
affects
the
the
ram
footprint
of
The
Collector,
we're
basically
saying
for
certain
deployments.
This
potentially
is
going
to
require
an
awful
lot
more
memory
than
you
typically
deploy
The
Collector
with,
and
you
know
how
does
the
Sig
feel
about
that?
H
That's
just
the
way
it
is
and
so
on.
Likewise,
converting
between
formats
might
be
CPU
intensive.
You
might
suddenly
need
to
deploy
the
collector
on
a
machine,
the
larger
CPU-
and
if
you
read
this
deck
you
can
you
can
infer
all
that,
but
I
think
making
it
explicit
in
one
slide.
So
there
can
be
talking
points.
H
So
that's
a
point
where
you
might
pause
in
the
presentation
say
you
know
what
what
do
you
feel
about
this?
How
is
this
going
to
sit
with
the
the
collective
people
and
yeah
that
might
be
really
useful,
I
think
the
other
one
that
belongs
on
that
side
is
probably
something
about
the
API,
because
one
of
the
things
The
Collector
does
with
its
current
protocols.
Is
it
exposes
the
data
flowing
through
it,
so
you
can
annotate
it
or
you
can
filter
it.
G
H
B
C
B
F
I
was
gonna,
say
you
know,
the
the
search
for
feedback
is
certainly
appreciated.
I'd
I'd
like
to
offer
just
a
thought,
perhaps
a
biased
opinion,
but
when,
when
you
ask
questions
like
that,
I'd
rather
or
maybe
you
can
see
if
this
fits.
But
maybe,
if
you
have
a
pocket
opinion
about
the
result,
you
want
steer
the
steer
it
a
little
bit
by
presenting
the
benefits
of
maybe
going
with
stateful
and
I
I
personally
I'm
a
little
skeptical
stateful
honestly,
but
I
don't
want
to
bring
that
into
it
like
rabbit
hole
on
that.
F
B
Yeah,
that's
great
yeah
in
terms
of
steering
it
I'm
curious
what
other
people
think
about
it.
I
I
kind
of
try
not
to
steer
too
much,
because
I
was
hoping
that
right
now
we
have
a
lot
of
things.
We're
not
sure
what
the
collectors
thinks
about
like,
especially
when
it
comes
to
like,
what's
Hotel
compatible
like
what
level
of
can
something
be
called
Hotel
code
compatible.
B
If
the
collectors
required
to
make
things
work
right
and
I
mostly
want
to
get
answers
to
these,
because
I
think
that
would
restrict
the
design
space
for
us
a
little
bit
more
and
then
we
could
like
regroup
and
be
like
okay.
Given
that
information,
it's
easier
to
now
see
what
what
we
should
actually
be
doing.
C
I
would
say
also
yeah
we
haven't
got
to
the
very
last
slide
with
the
specific
questions,
but
yeah
I
mean
I.
Think
it's
also
good.
You
know
obviously
there's
a
whole
lot
of
stuff
we
could
throw
at
them
and
what
will
likely
be
probably
even
shorter
time
than
you
went
through
this
in,
and
so
you
know,
I
I
think
it's
also
good
to
keep
in
mind
just
like
what
you
know
being
as
targeted
as
we
can
for
like
the
time
that
we
have,
and
you
know
getting
the
most
valuable
feedback
that
we
can.
C
That
will
help
us.
You
know
kind
of
move
forward.
You
know
even
just
a
little
bit
at
a
time
and
then
maybe
add
in
other
layers
over
time.
I
think
it
would
be
nice
if
we
even
could
get
someone
from
the
collector's
sick
to
maybe
come
to
these
meetings
occasionally,
especially
as
we're
like
you
know,
trying
to
make
progress
on
this,
but
I
think
yeah.
G
B
B
Okay,
yeah
I'll
clean
that
up
later
and
yeah.
Maybe
a
quick
look
at
what
could
be
on
that
slide?
Maybe
some
of
this,
maybe
not
all
of
it.
Maybe
this
question
is
not
the
worst
best
one.
We
can
discuss
the
order
this.
B
This
might
be
already
too
deep
in
the
weeds,
but
I
think
this
question
spelled
out
a
little
bit
more
clearly
like
if
it's
acceptable
for
The
Collector
to
be
required
for
hotel
compatibility,
I
think
is
one
of
the
key
questions
we're
trying
to
answer
so
I
will
try
to
rephrase
it
a
little
bit
to
be
clear.
B
We
want
to
get
people's
thoughts
on.
If
the
answer
to
that
is
yes,
like
the
collectors
required,
then
the
question
is
like
how
do
people
feel
about
sub
signals
and.
B
So
the
sub
signal
idea
was
this
thing
on
the
slide
here,
where
we
basically
say.
Instead
of
having
one
signal
for
profiling,
we
have
a
profiling,
CFR
profiling,
slash
P,
Prof
profiling,.
G
Signal
signal
is
what
else
speak:
okay
got
it.
B
Oh
tell
speaker
I'm
trying
to
get
into
that
happen.
It's
interesting
terminology,
then
yeah
comment,
support
architecturally
you
configure
which
formats
they
support
or.
B
I,
don't
quite
know
what
was
meant
by
that
I
think
I
copied
this
from
somewhere.
Let
me
come
back
to
that.
The
the
one
I
understand
is
the
this:
profiling
require
a
p
data
representation
that
also
goes
to
Jonathan's
point
about
maybe
not
having
or
having
the
ability
to
filter
data
in
The
Collector
before
sending
it
on
so
in
in
The
Collector
there's
this
idea
that
data
that
comes
in
generally
gets
converted
into
P
data,
which
is
like
an
in-memory
representation.
B
I
think
in
for
all
the
existing
signals
say
just
reuse,
the
wire
format
representations.
So
whatever
the
protocol
buffer
sync
generates
I
think
that's
what
they
call
their
P
data,
but
for
profiling.
We
should
explore
like
if
P
data
needs
to
be
like
all
the
data
taken
apart
or
if
for
profiling,
the
P
data
could
just
be
containing
plots
with
JFR
and
P
Prof
in
it
and
yeah
I.
Guess.
B
I
will
make
a
dedicated
slide
about
this,
but
yeah
how
how
buried
our
people
about
potentially
doing
stateful
receiver
things
where
the
receiver
has
to
buffer
up
a
lot
of
data
before
it
makes
sense
to
send
that
data
on
and
yeah.
That
I
think
is
one
of
the
key
questions.
G
B
E
C
B
Yeah
yeah
we've
already
rescheduled
a
little
bit
to
make
sure
on
the
collector
side.
There's
some
important
people
present.
C
Yeah,
if
I
remember
I,
don't
remember
exactly
mine
for
that.
Fourth
point
that
you
said
you
don't
remember
what
it
was.
Wasn't
that
something
along
the
lines
of.
C
C
I
thought
when
I
was
reading
it
I
was
thinking.
It
was
something
along
those
lines.
I,
don't
know
if
that
jogs
your
memory
or
maybe
not,
basically,
what
you're,
showing
in
slide
18
about
you
know
being
able
like
back
ends
being
able
to
sort
of
like
choose
which
formats
they
support.
I
thought
when
you
say
how
would
the
collector
want
format
support
to
show
up
architecturally,
like
you
know,
how
would
you
say
how
would
the
backend
say
what
formats
they
support
like
that's
what
I
thought
that
was,
but
maybe
not.
B
C
If,
if
you
don't
know,
then
maybe
it's
I
would
say,
there's
possibly
too
many
questions
on
here,
anyways,
so
yeah.
C
Yeah
yeah,
so
maybe
maybe
remove
it
for
now
and
focus
on
some
of
the
other
ones.
I
mean
I.
Think
to
me
when
I'm.
Looking
at
these
questions,
the
you
know,
sort
of
back
and
forth
negotiations
seemed
like
something
when
I
believe
it
was
Josh
at
that
one
meeting
was
talking
about
it.
C
That
might
be
somewhat
contentious.
You
know
that
I
guess
there
is
like
a
little
bit
of
Precedence,
but
like
I
guess,
I,
don't
know
how
much
is
it
acceptable,
yeah,
I,
guess,
yeah,
anyways,
I
think
there's
probably
more
important
questions
in
that
one,
so
I
would
be
in
support
of
maybe
removing
that
fourth
question:
yeah.
B
And
also
we
Auto
and
rewrite
these
a
little
bit.
I
think
this
is
probably
one
of
the
more
most
important
questions.
B
One
is
pretty
important,
so
yeah
I'll
rearrange
this
a
little
bit
and
so
yeah
once
I
get
a
chance
to
do
cleanup,
I'll
ping,
the
the
slack
channel,
to
give
everybody
a
chance
to
comment
again.
A
B
C
I
mean
I,
think,
that's
you
know,
yeah,
that's
pretty
good
I
added
some
of
the
feedback.
I
will
color
it
slightly
differently.
B
Yeah
and
everybody
who
had
like
there
were
some
great
comments
where
I
actually
I
should
have
probably
created
random,
slides
right
away,
but
for
those
who
had
feedback
make
sure
it's
either
in
the
meeting
notes
or
make
sure
it's
in
the
slides.
Those
are
going
to
be
the
two
places.
I'll
look
for
it
when
I
revisit
some
of
these.
C
C
So
I
wasn't
sure
I
think
the
collector
Sig
channel
is
just
Hotel
collector.
Does
anybody
know
for
sure?
Maybe
that's
a
question
we
can
ask
in
slack
to
Josh
and
tigron.
C
C
Show
up
if
we
give
some
heads
up,
yeah
yeah
I
was
just
making
sure
I
had
the
right
channel.
Okay,
so
yeah
we
can
I,
guess
offline
draft
a
message
and
send
soon
so
that
they
have
a
chance
to
look
at
it.
I
don't
know.
Do
we
want
to
if
we're
going
to
be
working
on
the
dock,
anyways
I
suppose
we
could
still
send
it.
C
I
mean
yeah,
I,
guess
I,
don't
want
to
wait
too
long
to
send
them
the
dock
ahead
of
time.
It's
kind
of
a
working
dock,
anyways
I,
don't
know
how
you
all
feel
about
that.
Sorry.
B
C
You
cool
I,
guess,
basically,
are
you
cool
with
us
sending
a
link
before
it's,
like
you
know,
in
its
final
state
that
you're
going
to
present
to
them
cool
all
right,
pretty
straightforward
for
that.
C
I
guess
I
will
go
ahead
and
do
that
they
meet
I,
guess
yeah,
let's
also
figure
out,
it
will
be
next.
C
E
B
E
C
Anyway,
yeah
we'll
figure
out
what
time
they
meet
and
post
it
in
there
by
the
end
of
this
meeting,
I
guess,
while
I'm
looking
for
that,
we
could
probably
move
on
meaning
I'll
figure
that
out
Pete.
You
had
a
question
in
slack
about
spam,
spans
versus
profiles.
F
Is
he
yeah,
yeah,
sorry
I'm?
Turning
on
the
video
there,
I
I
was
thinking,
you
know
how
to
bootstrap
an
explanation
of
this
and
and
explain
I
guess.
The
Genesis
of
the
question
was
just
like
explain
yourself
and
I'm
like
I.
Looked
at
I
stepped
out
of,
like
my
sort
of
local
Optima
and
I'm
I
was
looking
at
the
definition
of
things
and
I'm
like
don't
the
the
thought
that
occurred
to
me.
That
was
just
a
little
bit
out
of
the
blue.
F
B
The
first
thing
that
comes
to
mind
for
me
is
that
works
well
for
CPU
profiles,
where
you're
dealing
in
the
same
data
domain
being
time
of
some
sorts.
But
if
you
have
a
memory
profile
where
you're
talking
about
bytes
allocated,
then
you
quickly
find
that
the
metadata
around
the
span
is
not
all
that
relevant
at
least
yeah.
You
might
have
a
time
frame
over
which,
like
allocations
happen,
do
you
usually
don't
collect
each
one
of
them
that
would
be
probably
prohibitive
and
so
yeah?
B
G
G
Yeah
you
would,
you
would
need
to
weigh
them
like
you
would
need
to
Define
how
you
weigh
them,
like
you
would,
basically,
you
would
get.
You
would
have
to
get
deeply
into
okay
like
it
gives
some
it
gives
us
some
attributes,
but
what
kind
of
like
I
think
Hotel,
probably
has
no
notion
of
annotations,
like
spends,
can
have
can
have
expense
fans
have
like
key
value,
attributes
of
some
sort.
F
G
Some
of
the
attributes
need
to
be
like
some
of
the
attributes
that
are
stack
rather
than
per
stack
frame,
so
we
need
to
Define
okay,
like
how
do
we?
How
do
we
attach
attributes?
What
are
the
attributes?
How
do
we
attach
attributes
to
the
whole
stack
rather
like?
Should
it
be
attached
to
the
leaf,
and
then
it
means
that
it's
I
think
like
maybe
it's
worth
writing
down
like
pros
and
cons
of
the
approach.
A
G
F
G
F
G
Yeah,
because
it's
weighted
by
a
different
thing,
it's
not
like
the
races
and
spans
are
always
waited
by
time.
Essentially.
But
when
you
look
at
the
flameware
hey,
we
got
a
new
team
member.
G
But
with
yeah,
but
with
profiles,
things
awaited
by
by
whatever
like
it's
even
CPU
time
like
CPU
time
is
not
exactly,
is
not
exactly
the
kind
of
time
that
traces
and
spans
deal
with,
because
traces
and
spans
always
deal
with
wall
time.
Yeah
yeah
also
known
as
elapsed
time
yeah.
G
C
All
right,
we
do
have
one.
B
Quick
follow-ups
on
this
so
I
think
the
general
idea
of
taking
it
one
level
higher
of
like
hey.
Is
there
an
existing
Hotel
signal
that
profiling
could
go
into
I?
Think
at
some
point
we
talked
about
logs
potentially
being
a
channel
because
that's
more
of
a
like
hey,
take
it
back
of
data
and
put
it
into
a
hotel
I,
don't
think
we've
seen
I
have
not
seriously
explored
if
that's
feasible
in
any
way
or
form
like
if
binary
attachments
or
something
are
supported
there,
but
maybe
that's
something
to
also
ask
the
collector
sick.
B
It's
probably
none
of
our
preferred
way
of
doing
this,
but
we
should
at
least
have
thought
about
it
before
we
come
up
with
some
crazy
new
way
of
doing
things.
Yeah.
F
A
C
Yeah
I
mean
we
can
definitely
think
about
it
before
then,
yeah
I'll
be
interested
to
see.
If
the
collector
Sig
has
the
same,
you
know,
thought
you
know,
thought
process
or
line
of
questioning
there
yeah.
My
time
zone
was
on
the
wrong.
D
C
Know
wrong
time,
which
is
why
I
couldn't
find
it,
but
yeah
so
it'll
be
January
18th
Wednesday
at
9
00
a.m
to
10
a.m.
Pacific
time
I
believe.
If
I
set
my
time
right,
January
18th,
Wednesday,
9
A.M
to
10
a.m.
I,
don't
think
I
tried
to
put
a
link.
I'm,
not
sure
I
can
actually
put
a
link
to
it,
but
it
is
on
the
hotel
calendar
there.
C
If
you
have
the
hotel
calendar,
downloaded
yeah
I
would
say
you
know
obviously
I
assume
Felix
you're
planning
on
going
and
then
yeah
I
mean
as
many
people
as
can
I.
Think
that
would
you
know,
help
with
the
discussion.
If
there's
questions,
thoughts
and
also
just
I
mean
I'm
sure
they
record
the
meetings
as
well.
So
we
can
also
afterwards,
you
know,
share
the
recording
to
it,
but
you
know
particularly
yeah.
C
Maybe
if
someone
from
elastic
is
able
to
come
as
well
in
case
they
have
questions
about
the
format
or
something
that
would
be
nice,
but
you
know,
of
course,
obviously
it's
extra
time
so
up
to
you
all
and
then
yeah.
We
have
one
more
question
on
here
about
the
common
Trace
format.
C
B
I
think
it's
at
least
worth
discussing
if
something
else
is
out
there
that
potentially
has
a
lot
of
interesting
characteristics
when
it
comes
to
deduplicating
stack,
traces
and
symbols,
and
all
these
things
that
we
are
interested
in
in
I,
don't
know
if
this
common
Trace
format
is
false
in
that
category,
but
Thomas
mentioned
it
looks
interesting,
so
I
guess
people
who
want
to
take
a
look
could
try
to
do
that,
and
maybe
we
can
talk
about
next
time
once
we
have
somebody
here
who
has
looked
at
it.
C
G
Could
could
someone
could
someone
add
me
my
personal
email
to
that
meeting
to
collectors
check
because
I
might
have
removed
from
my
calendar
and
I
don't
see
it?
We.
A
C
C
Yeah
I
think
we
can
add
people
I
guess
maybe
paste
your
okay
actually
right
there.
Let
me
see
if
this
works,
email,
yeah,
I,
guess
paste
your
email
in
the
chat.
If
you
want
to
get
a
specific
invite,
I.
C
C
I
don't
know
if
anybody
can
just
edit
that
it
might
have
to
be
like
a
you
know,
some
of
those
special
permissions
or
something
but
I'll.
Try.
If
not,
then
we
can
ask
Morgan
or
tigon
or
somebody
to
try
and
do
it
other
than
that.
I
think
we're
good
I
will
try
and
like
write
up
just
like
a
short
message,
letting
them
know
giving
them.
The
link
to
the
presentation
should
I
also
link
the
architecture
proposal,
the
doc
as
well,
or
should
I
just
link
the
presentation.
B
C
Enough,
okay
sounds
good.
I
will
maybe
wait
a
little
bit
and
send
it
out
in
the
next.
You
know
like
24
hours.
E
Good,
all
right,
cool
I'll
take
a
picture
emails
all.
C
Right
well,
I
think
that's
everything
yeah
welcome
back
everyone!
Hopefully
that
meeting
goes
well
and
we
can
get
some
good
momentum
going
for
the
beginning
of
this
year
and
yeah
I,
don't
know,
get
everything
figured
out
and
moving
forward
relatively
quickly.