►
Description
CNCF Open Telemetry Community Meeting 2020-01-22 - part 2
A
B
A
A
A
A
C
D
A
A
I
think
we
can
get
one
more
on
that
pretty
soon.
Oh
I'll
advertise
this
one
after
the
meeting
here.
Okay,
let's
go
through
yours
next,
okay,
so
via
it
image
semantic
conventions
has
at
least
one
approval,
have
I
not
approved.
I
have
not
approved
yet
I
can't
remember
I
thought
I
did
oh
I
asked
the
question
about
host
image.
Yeah.
A
It
oh
I,
see
so
it
had
been
tagging
now
spent
version.
Okay.
Well,
that
I
have
looks
like
we
still
need
a
couple
of
provers
here,
but
it's
not
it's
not
still
at
this
point.
So
will,
after
this
meeting,
I
will
post
a
rally
to
get
some
of
these
things
and
you're
nearly
approved
done
common
event,
attribute
names,
I
feel
like
there's
been
some
lingering
confusion.
I
know,
there's
also
an
o-type
okay,
I'm
waiting
for
my
network
here,
there's
also
an
o-type
open
and
there's
been
some
shift
in
the
PR
overtime,
yeah.
D
A
C
C
A
F
What
chime
in
offense
I
know
on
a
PR
I
think
where
this
stuff
was
being
talked
about
and
I
think
kevin
has
been
leading
some
sort
of
charge
on
trying
to
define
some
semantic
conventions
around
events
that
represent
errors,
I
think
a
lot
of
that
lives
in
a
note
app,
but
yeah
I
guess
is
essential
for
a
1.0
and
so
I
did
kind
of
want
to
check
in
and
to
see.
If,
if
this
is
something
that
you're
that
you're
passionate
about
and
would
like
to
leave
the
child
I.
D
F
D
C
C
A
Work,
that's
sort
of
why
I
know
it's
a
swamp
I,
totally
support
getting
kind
of
finer
grained
information
about
errors
as
a
source
for
chase
sort
of
information,
but
yeah.
A
C
Don't
think
a
block?
Well
I,
don't
think
a
block
still
one
that,
oh
because
you
see
it's
really
not
changing
the
api's.
I,
don't
agree
it's
just.
This
is
just
another
kind
of
event,
so
here
you're
just
publishing
it.
Then
the
only
thing
that
it's
going
to
change
in
the
protocol
itself
is
is
now
you've
got
to
support
complex
objects
rather
than
just
simple
values.
C
F
Yeah,
my
point
of
this
being
a
something
we
need
from
1.0
is
right.
Now
most
the
SIG's
are
going
through
this
exercise
of
converting
over
data
dog,
instrumentation
and
there's.
You
know
numerous
points
in
the
instrumentation,
where
you
rescuing
exceptions-
and
you
know
are
aware
of
them-
would
like
to
be
able
to
kind
of
annotate
spans
with
this
data.
So
I
definitely
think
we
need
something.
F
I
know
it
can
be
a
a
swamp
like
I
would
settle
for
just
you
know
some
semantic
conventions
for
events
on
a
span
and
attributes,
even
that
might
not
be
the
best
fit
for
1.0
if
that
could
be
somehow
a
launching
pad
to
something
bigger,
but
I'm.
Also
if,
if
if
having
specific
type
events
for
these
is,
is
reasonable.
F
C
If
you
look
at
my
my
test
project,
I
got
it,
you
know,
I
got
a
the
data
types
defined
and
then
I'm
converting
in
four
different
back-end
system,
the
model
for
four
different
back-end
systems
and
there's
there's,
maybe
some
more
I'm,
not
100%,
satisfied
with
it.
Yet
there
needs
to
be
some
more
detail
to
it,
but
it
seems
you
know:
I
just
want
to
prove
that
it
works,
and
then
you
know,
as
far
as
the
actual
stacktrace
conversion
I
need
to
do
it
some
more
language.
Just
to
make
sure
my.
A
A
C
Copy
go
ahead,
yeah
the
only
cut
for
a
controversial
thing
or
thing.
That's
not
understood
is
this
message.
Id,
which
comes
from
T
RPC
is
is
where
that
comes
from,
because
on
a
GRC
P,
you
know
by
bi-directional.
Oh
right,
you've
got
messages
going
back
and
forth,
but
it
would
also
apply
to
something
like
like
HTTP
server
sent
events
possibly
are.
A
C
A
A
F
Inquiry
on
that
that
error
discussion
like
what
do
you
think
about
us
at
least
kind
of
trying
to
get
some
air
time
for
that
at
the
spec
meeting,
just
kind
of
bring
it
to
people's
attention
again
and
let
people
know
that
this
is
still
out
there.
This
is
still
kind
of
a
thing
where
interesting
is
coming
so
sounds
good
to.
D
A
A
A
A
Kevin
Jimmy
thoughts
on
this
one
I
think
this
needs
to
get
merged.
I,
don't
think,
looks
like
it's
got
approvals
yeah.
A
A
A
A
G
By
the
way,
boys,
by
the
way
Joyce
you
need
to
actually
you
know,
I
can
obviously
see
good-looking
area.
A
A
C
C
E
E
C
A
A
A
A
H
A
B
H
A
C
Too
I
don't
think
that
should
be
an
API
I
mean
their
concern
was,
if
you
run
serverless
that
you
need
to
get
it
flushed
out,
you
need
a
specific
call
to
flush
at
the
at
the
end
of
your
execution
so
that
it
gets
pushed
out
because
on
a
lot
of
platforms,
that
instance
sticks
around,
even
though
your
your
execution
is
is
fixed,
but
you
know
the
other
way
to
do.
That
would
be
to
put
some
kind
of
a
venting
thing
on
in
the
SDK
or
you
could
listen
to.
A
A
Yeah
I
guess
my
attitude
is
that
this
interface
is
going
to
be
I,
don't
have
a
I,
don't
know
I
I
understand
both
sides
of
this
argument.
Making
a
flush
API
be
a
specific
implementations.
Exposed
functionality
seems
fine.
Making
it
part
of
the
broad
at
default.
Sdk
interface
does
seem
too
worried
people
village
minute
reasons.
I
had
a
question
about
on
shutdown
like
many
something
just
have
a
shutdown
hook,
and
some
I
think
that
the
cloud
providers
we're
doing
service
provides
folks
as
well.
A
A
E
F
A
A
It
started
out
as
a
Zipkin
specific
issue,
but
okay
I
personally
am
not
inclined
to
go
through
the
last
four
of
these
should
Tristan
were
here.
I
would
I
would
be
happy
to
talk
about
3:54
I
am
personally
responsible
for
347
and
I'm
trying
to
catch
up
with
it,
but
it's
not.
It
hasn't
moved.
Are
there
to
discuss
some
terminology
renames,
which
we've
kind
of
agreed
upon
in
the
metric
sig
meeting,
but
haven't
haven't
applied
to
this
document
yet
Chris.
Do
you
want
to
talk
about
probability
and
sampling?
H
H
A
H
H
A
A
A
D
A
A
A
B
A
A
A
A
Okay,
cool
I,
hereby
disagree:
sampling
decision
may
need
update
race
state,
not
only
attributes.
This
is
another
one
of
these
cases
where
I'm
feeling
a
bit
length
because
I
don't
see
how
you're
gonna
use
sampling
decisions
without
sampling
weights,
and
we
have
a
please
punted
on
weight.
So
like
it's,
not
really
sampling.
If
you
don't
have
weights
my
personal
opinion,
so
I.
H
H
H
Up
my
my
feeling
is
that
I
could
be
wrong
about
this,
but
I
think
by
removing
sampling
hint
were
actually
changing.
The
the
order
of
the
sampling
decision
so
before
we
make
the
sampling
decision
include
some
information
in
the
form
of
the
hint
and
now
it
seems
like
we
don't
need
to
like
pass
extra
information
from
the
sampler
along
with
the
the
tri-state.
If
that
makes
sense,
so
we
set
the
sampling
concision
separately.
H
A
A
E
A
A
A
A
A
Discuss
using
HTTP
request
an
HTTP
response,
so
it's
pretty
old
I.
C
There's
not
a
nurse
right,
there's,
not
any
semantics
conventions
for
headers
anyway,
right
now,
okay,
so
yeah.
So
you
would
have
to
add
that
first,
but,
like
this
says,
span
span
kind,
although
that's
not
true.
There's
the
span
kindness,
whether
it's
come
from
the
client
or
the
or
thee
I
mean
this
list
seems
pretty
more.
The
server.
H
E
E
G
A
G
A
A
H
A
H
H
C
C
H
H
H
C
H
A
Yeah
I
feel
like
sampling,
is
something
that
should
not
work
in
a
coordinated
way
across
vendors,
it's
too
complicated.
So
that
doesn't
bother
me
what
just
was
said:
I
but
I'm
more
familiar
with
people
wanting
to
generate
their
own
random
number.
Is
them
random
numbers
just
for
performance
reasons
like
you
have
a
thread:
local
generator
that
you
can
access
without
a
mutex?
That's
the
one,
you're
gonna
use
to
allocate
your
spend
ideas
so
that
you
don't
have
contention
just
kidding
if
access
just
getting
access
to
a
number
generator,
but
it's
not
a
requirement.
A
A
D
C
E
F
One
one
thing
that
I've
had
experience
with
and
trying
to
create,
like
a
taxonomy
of
stands,
is
that
you
have
three
areas.
One
of
them
that
often
comes
up
is
you'll,
have
like
database
clients
that
operate
over
a
kind
of
like
actually
like
this
HTTP
REST
API,
but
they're
also
kind
of
a
database
and
like
yeah.
C
A
Personally
feel
like
the
right
way
to
do
this
is
to
have
schemas
and
the
thing
can
have
multiple
can
satisfy
multiple
schemas.
Therefore,
it's
multiple
types
of
thing,
so
you
know
the
schema
saying
to
be
a
error
message.
You
must
have
a
stack
trace
and
a
type
like
I,
don't
know
like
you
just
give
a
description
of
what
types.
If
these
fields
are
present,
then
you
are
a
this
type
of
thing
and
then
you
can
make
it
Francis.
C
C
A
Yes,
you
can
define
a
type
as
a
union
type,
so
it's
it's
I.
This
feels
like
a
very
great
area
that
I
don't
care
about
enough
to
get
hung
up
on.
So
I
would
be
in
favor
of
striking
this
myself,
but
I
don't
want
to
speak
here,
I'm
not
being
with
this
right
now.
What's
your
life,
so
Oh
we'll
leave
it
in
a
milestone
and
move
on.
C
A
A
E
B
A
I
have
somewhere
I,
don't
remember
where
I
recommend
it,
but
you
know
context
of
HF
66
that
you
have
this
current
current
spam
sampling
API,
which
has
like
six
arguments
in
them
and
all
of
them
come
from
the
context.
So
you
could
just
have
a
sample
with
a
context,
argument
and
I
know,
there's
a
performance
argument
to
say
well,
you're
an
SDK
and
you're
in
the
middle
of
this
delicate
operation.
To
start
a
new
span,
you've
got
six
fields
lying
around.
A
You
should
just
call
the
sample
with
your
six
fields,
but
you
can
also
simplify
the
API
call
what,
with
one
context
and
now
in
context
of
otech
66,
we
have
a
upstream
context.
We
have
a
current
span
and
has
any
number
of
attributes
that
are
in
so
distribute
context
or
the
correlation
context.
So
you
should
not
use
all
of
those
and
I
don't
feel
I,
but
I.
Don't
also
don't
feel
like
I
care
enough.
I
keep
saying
that
about
smoothly,
which
is
really
contradictory,
but.
F
F
A
D
A
A
A
A
Some
of
them
are
tricky.
These
two
are
tricky
I
think
we
can
close.
One
of
them
unspecified
values
is
still
a
topic
that
keeps
coming
up
so
I'm
gonna
leave
that
one
for
sure
I'm
working
on
the
metrics
API
specs
this
week,
so
I
will
get
this
one
done
something
easier
for
me.
I
can
do
this
one
too,
so
we're
down
all
the
way
to
invar
for
where
to
send
fans
which
came
up
earlier,
so
the
context
for
TLS,
oh
geez
resources
just
an
attribute,
sir
okay.
This
has
got
to
end
all
right.