►
From YouTube: 2020-08-07 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
A
C
I
was
saying
if
we
don't
have
an
agenda
by
the
way
this
is
xinjing
from
aws.
We
can
spend
some
minutes
to
review
the
x-ray
sampling
specification
and
see
what
we
can
get
from
there.
A
How
does
that
sound
yeah?
That
sounds
good?
In
addition,
I'm
gonna
I'd
like
to
reach
out
to
some
of
our
customers
and
get
some.
D
Guys
this
is
davi
from
aws,
I'm
looking
at
the
participants
here
like
this
is
pretty
much.
We
have
three
people
from
aws
and
we
all
work
for
x-ray,
plus
your
math,
which
I
apologize.
I
I
don't
think
we
have
met
before
here
so
shangxing.
I
don't
think
it
makes
sense
for
us
to
discuss
the
decree
sampling
as
requested
as
an
action
item
earlier
without
the
larger
audience,
but
does
that
make
sense
and
and
to
your
question
as
well
matt
like
it's
to
your
point
as
well.
D
A
Yeah
I
can
understand
that
and
what
which,
what
is
that?
Typically
people
just
might
be
on
vacation.
It's
august.
You
know
there
might
be
a
lot
of
people
just
out
on
vacation.
What's
what's
like
considered
a
quorum
on
this
group.
D
Yeah
like
so
I'm
even
surprised
that
ben
who
has
been
the
leader
in
this
group
is
not
attending.
Maybe
he
is
on
vacation.
I
didn't
get
any
mail
and
I'm
trying
to
see
if
maybe
they
have
changed
they
like
the
zoom
meeting.
D
A
Okay,
well
yeah.
I've
reviewed
your
documentation
on
x-ray
and
you
know
certainly
interested
in
in
the
spec,
but
also
don't
want
to
waste
your
time.
So
you
know
it
probably
does
make
sense
to
cover
a
bigger
audience.
So
you
just
don't
have
to
you
know,
repeat
yourself,
so
maybe
yeah.
Maybe
we
should
just
cancel
this
one
and
be
more
proactive
about
sending
out
the
agenda
in
advance.
You
know,
like
I
mean
I'm
happy
to
to
take
on
a
little
bit
of
that
responsibility.
D
A
Okay,
well,
hey
have
a
have
a
good
weekend.
I
guess
I
guess
we're
appreciated.
A
Good
to
meet
you
too,
I
mean,
I
think
you
know.
I
know
the
specifications
community
has
a
lot
on
their
plate
and
you
know
sampling
probably
isn't
top
of
mind
at
the
at
the
moment
for
them,
but
you
know
I
know
our
customers,
you
know
really
care
about
it
and
depend
on
it.
So.
D
Here's
the
deal
I
I
have
been
attending
this
meeting
since
I
think
the
first
or
second
one.
What
what
I
have
learned
in
particular
is,
I
think
every
vendor
has
their
own
perspective
on
sampling
and
and
one
of
the
agreements
that
we
have
had
on
this
group
at
least
informally,
is
on
days
back.
We
need
to
drive
on
the
lowest
common
denominator,
yes
right
and
it's
like
specifically,
because
certain
vendors
have
different
way
of
collecting
data
right
and
then
they
can
do
different
type
of
sampling.
D
So
we
cannot
approach
every
single
sampling
strategy
that
customers
would
love,
and
one
of
the
goal
is
like
make
sure
that,
through
the
collectors
and
and
the
extensibility
qualities
of
the
collector,
vendors
can
indeed
enhance
and
implement
their
own
sampling
that
they
currently
have
and
that
they
believe
their
customers
would
like
to
have
as
well.
A
Yeah,
that
makes
sense
I
mean
and
as
as
I've
been
digging
into
this,
you
know
it
kind
of
feels.
Like
you
know,
there's
I
mean
there's
there's
there's.
Basically
you
know
the
head-based
sampling
and
the
tail-based
sampling,
and
then
you
know
some
people,
you
know
try
to
sort
of
get
more
sophisticated
with.
You
know,
kind
of
complex
algorithms
or
no
sampling
right
and-
and
so
you
know,
like
you,
said,
yeah
lowest
common
denominator,
I
mean
it
seems
like
we
need
to
support.
A
You
know
both
you
know
head
based
and,
like
you
said
in
the
collector
people,
could
you
know
layer
their
tail-based
sampling
on
top
of
the
head-based
sampling?
And
then
you
know,
I
think,
the
additional
requirement
that
we
have-
and
you
know
that
that
you
have
is
you
know
we
need.
We
need
some
way
for
our
customers
to
be
able
to
to
modify.
A
You
know
sampling
from
the
from
the
from
a
service.
You
know
from
the
cloud
store
settings
in
the
cloud
and
send
it
to
the
sdk,
and
I
I
think
I
I
again
I
haven't
you
know
completely
settled
on
this,
but
I
think
we
need
to
be
able
to
to
to
change.
You
know
the
customers
need
to
be
able
to
change
head-based
sampling
too.
So
I
think
the
sdk
you
know
would
also
need
to
be
able
to
receive
settings
from
the
cloud
you
know
not
not
just
the
collector.
You
know
I.
D
A
Okay,
so
I
think
that's
the
one
piece
that
I've
noticed-
that's
probably
missing
from
from
this
from
the
spec.
I
don't
see,
I
didn't
see
anywhere
well,
there
might
be
an
open
issue
that
I
haven't
read
through,
but
I
just
didn't
see
anywhere
where
the
the
sdk
was
going
to
be
able
to
receive.
A
You
know
settings
to
to
change.
You
know
to
override
head-based
sampling,
basically.
D
Okay,
one
of
the
things
so
so
just
for
your
information,
like
the
people
that
have
you
on
the
call,
I
manage
the
x-ray
engineering
team
for
the
back
ends
or
the
trace,
analytics
components
and
sandra
is
my
peer
manager
leading
the
sdk
p,
so
she's
super
interested
in
that
space
and
then
leading
that
particular
area
with
shenshin.
Who
is
one
of
our
lead
engineers
on
that
group?
D
One
of
the
things
that
I
myself
have
been
advocating
on
this
group
is
like
we
need
to
have
if
we
want
to
have
proper
sampling
support
at
the
launch
dates
or
later
this
year.
We
need
to
have
a
definition
soon
because,
as
you
said,
like
the
cloud
vendors,
they
take
some
time
to
implement
that
right.
D
So
if
they
want
to
have
support
for
like
head-based
sampling
at
alb
for
aws
or
api
gator
for
aws
or
the
load
balancer
offerings
or
api
endpoints
from
azure
from
google,
we
need
to
start
work
on
those
very
soon
like.
We
cannot
turn
around
that.
D
The
implementation
on
those
service,
like
as
we
do
for
the
client
side
on
the
on
github
right
like
it's,
not
that
quick.
We
go
through
large
deployment
cycles
and
so
so
on.
So
it's
it's
cross
organization
work
within
those
vendors.
So
it's
it's
super
important
that
this
group
focus
on
having
at
least
a
minimal
definition
that
we
can
start
as
cloud
vendors,
public
cloud
vendors.
We
can
start
work
on
it
for
our
with
our
partner
teams.
A
D
So
we
do,
since
we
do
pretty
much
had
base
sampling
those
kind
of
service,
api,
endpoints
alone
balancers.
They
they
can
make
sampling
decisions
right.
So
we
need
to
understand
what
is
the
minimal
definition
of
a
sampling
decision
and
what
should
it
be
carry
over?
So
we
can
start
talking
to
those
teams
that
that's
pretty
much.
One
of
our
all
minimal
asks.
D
The
discussion,
but
partially
for
aws,
like
we
have
that
interest
as
well.
We
have
the
interest
on
understanding
how
the
current
sampling
logic
and
features
for
x-ray
works
and
how
it
matches
on
the
on
this
pack
and
that's
what
sandra
is
driving
with
chiang
xin
sheng
shin
was
proposing
to
present
today.
Okay,
that's.
A
We're
also
gonna
need
lead
time,
because
this
is
gonna
affect
yeah,
our
back
end
service
and
you
know,
and
the
ux
and
it's
going
to
be
different.
You
know
from
the
from
the
way
we
do
sampling
today,
which
is
like
just
a
tail-based
sampling
which
is
not
like
you
know.
I
mean
it
served
us
okay
so
far,
but
it's
it's
not
it's
not
terrific
from
a
perfect
standpoint,
so
yeah
so
it's.
I
agree.
I
totally
agree
like
this.
That's
why
I'm
kind
of
started
joining
these
meetings.
I
think
yeah
we
needed.
A
D
Yeah,
it's
it's
it's.
The
main
point
is
like.
Yes,
all
the
client-side
work
can
be
done
on
the
github
repos,
but
like
us
as
cloud
vendors,
we
also
need
to
change
back-ends,
not
only
our
own
trace,
analytics
back-end,
but
like
all
the
managed
services
performing
sampling
decision
for
distributed
tracing,
and
that
is
a
way
longer
lead
time.
A
Yeah,
so
is
there
you
know,
is
there
a
way
I
could
start
kind
of
reviewing
your
spec,
you
know
in
advance,
or
would
you
would
you
rather
just
wait
and
I'm
I'm
trying
to
figure
out
what
our
spec
is
to
and
kind
of
working
on
a
vision
document?
So
you
know
any,
it
seems
like
you
know.
A
D
Can
you
add
it
to
the
work
group
note,
so
matt
has
access
to
it
and
can
and
everyone
else
can
have
access
to
it.
Since
the
goal
for
this
meeting
today
was
to
review
this
doc.
D
B
A
Silently
and
then
ask
your
questions
so
all
about
that:
okay
yeah!
I
have
some
really
messy
notes
here,
but
you
know
we.
I
can
clean
these
up
and
I
think
maybe
we
still
just
add
our
names
here
under
attendees
just
because
we
at
least
met
for,
like
you
know,
25
minutes
and
then
yeah.
Let's,
like
I
totally
agree,
let's
just
kind
of
push
out
that
action
item
to
the
rest
of
the
group
to
try
to
like
get
this
thing.
Moving
forward.
B
Yeah,
if
you
want
to
comment
on
our
spec
we'd,
be
happy
to
work
with
you
out
of
band
next
week
in
between
meetings
too,
but
it's
not
like.
We
have
to
wait
till
friday
to
discuss
this
so
yeah.
Are
you
active
in
the
glitter
group
as
well.
A
I
think
I
haven't
posted
yet
I
haven't
really
I've
just
been
looking
through
github,
so
nope,
I'm
not
active.
Yet.
Okay.
B
Well,
if
you
would
just
want
to
ping
us
on
there,
if
you,
when
you
update
or
have
any
questions,
that's
probably
the
fastest
way
to
get
in
contact
with
us.
A
Okay,
great,
do
you
have
anyone
from
your
side
who's
you
know
like
the
thing
is,
is
this
is
pretty
you
know
so
far?
This
is
pretty
engineering
driven
which
you
know,
which
is
which
is
okay,
but
do
you
have
anyone
who's
reaching
out
to
customers
at
all,
or
are
you
taking
any
like
customer
feedback
into
consideration
and
just
just
curious.
D
Yeah,
like
our
pm,
is
our
tyrell.
He
he
has
it.
D
We
have
heard
so
many
times,
customers
asking
for
tail-based
sampling
that
that
that's
the
best
definition
for
us
in
general,
like
customers,
would
like
to
to
collect
the
outliers
that
that's
the
most
common
ask
that
we
hear
from
customers
and
and
bizarre
is
the
person
like
talking
to
customers
and
hearing
that
feedback
consistently,
as
well
as
doing
reinvent
sessions
where
all
the
engineering
managers
attend
as
well
got
it.
D
B
I
think
they
do
because
other
tracing
providers
already
do
this
today.
I
think
honeycomb
was
talking
a
couple
of
months
ago
about
that.
D
Absolutely
like,
but
I
think
at
least
on
the
discussions
that
I
have
attended
sandra
like
that
there
is
a
sort
of
consensus
that
yes,
customers
may
have
a
hint
that
there
is
a
performance
concern,
but
they
don't
understand
the
cost
concern
for
now.
In
no
sense.
A
D
On
x-ray,
we
do
have
something
around
those
lines
like
we
do
offer
a
way
for
customers
to
tune
the
sampling
rate
through
the
console
and
distribute
those
the
sampling,
configuration
assignment
sampling
quotas
to
the
clients
in
a
distributed
manner,
and
we
collect
counters
from
each
client,
and
we
plot
like
a
time
series
showing
like
a
what
is
your
sampling
target?
How
much
you
are
sampling,
so
customers
can
see
some
sort
of
like
under
sampling
or
over
sampling
behavior.
D
We
don't
alert
for
like
potential
cost
concerns
that
that
could
be
a
good
enhancement
for
us,
but
we
at
least
have
a
visual
way
to
communicate
to
customers
about
over
sampling
or
under
sampling
scenarios
and
how
much
they
are
sampling
compared
to
the
total
traffic
being
seen
on
the
customers
on
the
clients.
Sorry,.
A
Yeah,
that
makes
sense,
I
think
I
think
I
think
I
think
we
want
to
move
to
a
direction.
You
know
well,
at
least
with
with
small
volume
customers.
I
think
we
want
to.
You
know,
move
to
a
direction
where
most
customers
never
even
need
to
think
about
sampling.
You
know,
like
the
majority
of
our
customers,
have
send
a
loan
of
volume
that
you
know
like.
We
can
just
do
the
right
thing
for
them.
We,
you
know,
we
know
the
right
percentage
and
you
know
they
could
always
override
it
if
they
need
to.
A
But
it
seems
to
me
like
that
kind
of
alert
you
know,
should
probably
only
be
surfaced
to
like
you
know,
a
larger
customer
who's
using
you
know
who's
using
fixed-rate
sampling.
You
know
head
based,
you
know
for
a
smaller
customer,
it
seems,
like
you
know,
a
low
volume,
customer
tail-based
sampling
seems
fine
to
me.
You
know
because
they
don't
they're
not
really
constrained
by
perf
anyway.
So
I
just
don't
know
like
that's
kind
of
where
I'm
thinking
in
our
spec
like
at
some
point.
A
We
need
to
have
a
line
where
we,
you
know,
switch
customers.
We
kind
of
alert
them
like
say,
like
hey,
you've
reached
the
the
boundary
of
tail-based
sampling.
You
know
in
that
now
this
is
having
you
know,
a
perf
impact
on
your
application.
We
now
recommend
head-based
samples,
some
something
like
that.
You
know
that
that's
good,
that's
kind
of
what
I'm
thinking.
But
again
I
haven't
really
validated
this
with
customers,
yet
so.
B
Well,
isn't
the
entire
point
of
tail-based
sampling
to
capture
errors
and
exceptions
like
I
wouldn't
imagine
that
would
increase
their
traffic
volume
that
much
unless
they're
having
that
many
uncaught
errors
or
exceptions?
But
even
that
sounds
odd
to
me.
A
Well,
my
understanding
of
tail
based
sampling,
at
least
the
way
we
implement
it
is
like
all
the
all
the
traces
are
captured
and
stitched
together,
and
then
you
based
on
the
tail-based
sampling
settings.
You
know
we
we
we
drop
them,
we
drop
them,
but
after
we've
done
all
the
processing,
so
it's
so
it
is.
It
is
expensive,
even
if
we
only
decide
at
the
end
of
the
day
to
send
forward
the
exceptions
they
pay
a
price
for
it.
A
Now
there
could
there
be
some
more
efficient
implementations.
You
know
possibly.
D
D
There
are
vendors
like
x-ray
and
aws
that,
like
it's
just
client-side
collection
and
and
a
little
buffer,
like
short
buffer
on
an
out-of-process
component,
and
then
they
send
everything
towards
the
back
end
and
and
the
cost
will.
The
definition
of
cost
for
customers
and
performance
impact
will
change
for.
B
D
Right
like
for
us
what
we
have
to
discuss
about
customers,
not
fully
understanding
the
cost.
Concern
is
like.
Okay,
now,
I'm
buffering
in
process
for
30
seconds
latency
up
to
30
seconds
to
do
tail
sampling.
What
are
the
memory
constraints
that
we
can
impose
and
will
customers
need
to
upsize
their
instance
to
run
this
right?
D
So
so
those
are
kind
of
the
considerations
that
our
architecture
lead
to
when
you
guys
is
different
on
microsoft,
because
you
have
that
intermediate
collector,
which
customers
are
paying
for
that
collection,
price
point
right
and
even
though
you
drop
the
data,
so
they
get
no
value
from
the
data
that
was
collected
and
charge
it.
But
that's
what
you
need
to
do
to
do
tail
sampling
with
your
architecture
so
make
sense.
A
I
see
okay,
okay,
cool
well
thanks
thanks
for
for
posting,
the
link.