►
From YouTube: GMT 2018-09-18 API WG
A
A
B
B
B
C
Optimization
pads
for
the
sword
are
for
the
soldier
on
performance
improvement,
so
in
the
Psalter,
when
we
do
those
sorting
based
on
the
resource,
allocating
resources,
it's
very
efficient
because
it
has
a
lot
of
you
know
the
resource
metadata.
So
if
we
just
add
and
disorder,
sorting
only
cares
about
the
quantity,
so
I
think
been.
Added
are
pets
that
introduced
a
resource
quantity
class,
and
then
you
just
use
that
to
do
the
sorting,
though
I
think
the
result
was
like
improved
performance,
significant,
like
35
percent
in
some
setting.
C
So
while
looking
at
the
code
I
realize
that
there
other
places-
probably
we
need
this,
not
least
internally,
but
also
some
of
the
external
API,
is
like
quota
setting.
Well
pretty
much.
You
only
care
about
the
resource,
quantity
and
the
mean
allocatable
resources
as
well.
So
right
now,
for
example,
when,
when
you
set
up
the
quotas,
we
need
to
do
a
lot
of
validations
to
make
sure
that
there
isn't
any
metadata
there.
C
For
example,
when
you
the
specify
the
quota
we'll
make
sure
that
it's
not,
you
cannot
specify
a
quota
for
reservations
and
something
like
that
so
I,
so
that
I
were
looking
at
this
code.
I
feel.
Maybe
we
should
make
this
a
first
class
construct,
expose
it
externally
as
well
like
introduce
a
protobuf
construct
and
use
it
internally
for
where,
for
the
code,
we're
only
quantity
makes
sense,
and
these
were
also
improved
performance
as
well.
A
C
C
C
I
think
there
are
several
places
in
the
code
base
that,
like
only
care
about
quantities,
yeah
like
some
some
some
places
in
the
allocator,
where
we
calculate
like
the
quantities,
use
the
allocated,
quantities
to
or
cluster
capacity
to
calculate
what
we
should
allocate
and
also
in
this
order
to
calculate
shares
and
some
external
places
quota
when,
when
you
set
up
set
up
a
quota,
it's
just
a
quantity
guarantee
and
also
me
allocatable
resources
and
I.
Think
there's
some
like
Jay
mentioned
to
me.
There
in
some
of
the
resource
provider
say
out
there
also.
B
C
C
C
C
C
A
C
C
C
C
B
Yeah
I
mean
I
think
in
my
mind
it
probably
only
makes
sense
for
scalar
quantities
and,
in
any
case,
I
think
that's
the
current
application
that
we
have
a
guy
I,
don't
think
our
current
applications
require
such
a
construct
for
sets
or
ranges,
but
it
might
be.
Oh
yeah
and
the
name
of
this
class
is
already
scalar
research,
yeah.
C
B
C
C
C
So
if
you
so
rock
right
now,
there's
only
one
field
core
get
guarantee,
and
this
is
used
for
both
guarantee
as
well
as
limit.
So
if
we
make
the
change
we're
at
a
couple
that
so
for
other
versions,
you
can
only
see
guarantee
and
if
we
are
using,
if
a
newer
version
that
is
capable
of
doing
limits
inherent
this
guarantee
only
request,
then
it
well
I
think
followed
all
the
semantics
of
setting
the
guarantee,
as
well
as
a
limit
to
the
to
the
other
guarantee.
But
we
can't
do
that
in
Reverse.
C
C
B
C
B
C
C
I
think
we're
drifting,
but
mine
is
I.
Think
the
minimum
capability
check
well
yeah
I
need
to
recall
my
memories,
so
basically
we
would
persist
some
of
them
so
for
the
newer
master,
which
is
capable
of
doing
the
limits.
The
world
right,
for
example,
quota
limit
as
a
string
into
the
minimum
capability,
and
then,
when
the
old
master
get
up,
you
give
a
first
track
the
minimum
capability
field
and
thinks
that's
in
string
format.
It
was
see
that
okay,
this
replicated
log,
is
tagged
with
quota
limit.
C
C
C
C
B
D
Actually,
I
have
coaching
with
the
research
quantities.
I
was
wondering
if
that
is,
it
is
reasonable
to
like,
say,
model
poured
here
like
say
pours.
We
we
have
like
intervals
like
right,
but
the
quality
can
be
like
a
number
of
cores
inside
that
interval.
Does
that
make
sense,
so
we
can
enforce
some
sort
of
quota
on
even
non-scalar
resources.
D
C
B
D
It's
probably
like
it's
probably
not
well
I'm,
just
thinking
about
we
may
well
I,
don't
have
a
congruence
case
for
now,
but
it's
like.
We
can
also
import
quota
on,
like
I
10,
like
resource
with
ID,
essentially
but
yeah.
This
is
not
the
you're
like
high
priority
thing
for
now:
I
guess,
because
I
don't
have
that
explicit
X
to
the
use
case,
yeah.
C
B
D
C
D
C
C
A
C
And
yeah
we
are
adding
updated
quota,
so
up
Dakota
will
be
again
direct
you
to
the
quota
request,
so
quota
request
is,
is
mostly
the
same
except
we're
adding
a
limit
so
I.
Imagine
if
we
want
to
do
this
quantity
stuff,
we
can,
you
know,
add
a
new
field
here
or
introduce
another
new.
Could
I
request
or
something.
B
B
C
Uh-Huh
I
think
so
it's
a
difference
here,
I
think
is
one
is
repeated.
Another
is
required
so
so
these
features
okay.
So
one
of
the
rationale
I
think
is
to
just
it
is
all
nothing
here,
so
you
can
reduce
Waldo's
quota
or
increase
well.
At
the
same
time
you
increased
the
other
and
and
in
the
past
you
can't
do
this
like
atomically.
C
So
if
you
reduce
a
rose
quota
and
that
quota
at
at
some
point,
it
might
be
like
unclaimed,
that
is,
it's
not
a
part
of
the
core
of
any
row
and
then
dedicated
my
allocate
that
part
of
resource
to
other
rows.
And
then
you
ask
between
to
increase
another
rose
quota,
and
at
that
time
we
don't
have
enough
like
heroine
to
satisfy
that
corner.
I.
Think
that's
one
of
rationale
to
reduce
this
call.
Yeah.
C
B
B
B
B
D
B
A
D
C
B
B
C
B
C
B
B
B
Benchmark
call
ingestion
path
on
the
mises
monster.
I
think
that
yeah
api
benchmarking
in
general
is
something
we
could
do.
Sephora
I
know
that
Ben
Maller
has
done
a
bit
of
this.
To
compare.
If
you
want
to
be
zero
performance
specifically,
is
the
call
and
Justin.
We
got
a
report
significant
regression
of
performance.
Okay,
I
wonder
this.
A
B
A
A
B
A
B
B
B
B
B
B
A
A
A
A
B
B
B
D
B
D
A
D
B
D
Reason
is
like
we
I,
think
what
and
then
initially
had
worked
on
some
HTTP
API
improvements
which
the
response
body
into
a
pipe
so
and
we
directly
like
process
the
outcome
like
read
the
pipe
out
in
body,
so
there's
no
way
we
can
like.
We
can
monitor
what
are
in
a
pipe
and
we'll
need
the
body
to
do
the
matching
like
a
future.
D
He
recalls
when
you
match
that
is
actually
a
subscribe
code
right
or
actually,
whatever
call
it
used
to
be
reversing
a
long
while
ago,
because
back
then
we,
the
the
like
we
process,
we
do
scan
through
the
the
request
body
and
you
and
use
the
term
through
the
matching
to
do
type
matching.
But
now
the
body
sing
the
pipe
so
that
the
Chris
body
or
response
bodies,
basically
empty
or
acquittal
Bridgestone
body,
are
basically
empty.
D
B
A
B
B
D
A
A
A
B
A
B
A
A
It's
not
bug
per
se,
one
of
the
increments
that
it's
one
of
the
bugs
that
we
have
recently
found
internally
was
with.
When
you
have
intermediaries
in
play.
Sometimes
Messrs
does
not
detect
that
a
streaming
client
has
disappeared
and
we
were
thinking.
Maybe
one
of
solutions
would
be
to
actually
add
a
heartbeat
from
the
client
to
the
server
right.
Now
we
only
heartbeat
from
the
server
to
the
client
one
way
for
streaming
api's,
and
we
were
thinking.
Maybe
a
more
robust
mechanism
would
be
to
also
have
clients
send
send
a
heartbeat.
A
Yeah
I'd
bet
call
just
like
how
we
have
heartbeat
events
today
and
that
way
server
could
also
detect
more
than
instead
of
just
detecting
based
on.
When
the
connection
gets
closed,
it
could
also
detect
based
on
like
a
time
not
on
the
heartbeat.
Sometimes
there
are
cases
where
connection
is
not
getting
closed
for
whatever
reason,
but
the
client
has
gone
away,
especially
their
intermediaries
in
play.
Click
proxies,
that's
one
improvement,
you're
thinking
about
recently
that
comes
to
mind.
So
the.
A
B
A
Am
NOT,
I
think
I
would
say,
look
at
the
Twitter
streaming
API
to
see
if
they
had
anything,
that's
one
of
their
inspirations
and
how
we
designed
our
API
so
see
if
they
have
any
mechanism
they're
the
expert
heartbeats
from
the
clients
or
something
and
if
they
do
maybe
should
also
start
thinking
about
that
as
well.
I
mean
that's
not
too
hard
to
implement
against
just
in
figure
out
the
right
time.
What
is
and.