►
From YouTube: 2017-MAY-04 :: Ceph Performance Weekly
Description
Weekly collaboration call of all community members working on Ceph performance.
http://ceph.com/performance
For full notes and video recording archive visit:
http://pad.ceph.com/p/performance_weekly
A
A
C
D
Fair
mark
good
morning
mark
this
is
James
from
Alibaba,
so
yeah.
What
okay?
You
can
give
us
ten
days
to
present
the
one
we
already
achieved.
Where's
our
recovery
we
have
presented
will
to
their
team
of
their
performance,
get
Collins
we
already
achieved
and
also
the
patch
we
are
working
on.
So
we
want
looking
for
we're.
Looking
for
some
more,
you
know,
comments,
and
otherwise,
especially
from
Josh
judge.
Here,
will
come
their
recovery
right,
that's
right,
yeah!
D
So
we
want
to
share
with
what
we
already
did,
and
their
you
know
would
be
great.
Even
can
give
us
some
more
comments
and
their
because
right
now
we
already
put
them
into
our
production
and
I'm
gonna
fake
marker
box,
and
let's
see
whether
we
have
a
chance
to
review
it
and
you
possible,
like
you,
you
can't
give
us
more
comments
about
it.
Damn
degree
is
that
missing
here.
D
B
Right,
let's,
let's
go
through
the
poor
across
here
first
and
then:
well,
maybe
we
can
jump
into
other
stuff
that
you
guys
have
been
working
on
after
that
to
the
Sun
good
yeah.
C
B
So
there's
been
a
couple
of
new
pour
requests.
One
of
them
is
actually
for
me
here,
basically
when
doing
small,
random
right
or
clothes,
but
also,
probably
in
some
other
cases
as
well.
You're,
seeing
a
lot
of
Reed's
during
compaction
and
boost
or
vaak
siddhi
was
just
kind
of
going
up
and
it
turns
out
that
rocks
Phoebe
uses
a
random
access
file
breeder
for
doing
reeds
during
compaction,
but
in
reality
it's
almost
all
sequential,
even
though
it
seems
to
be
random.
It's
basically
just
doing
sequential
reason.
B
B
Luckily
it
actually
works
for
us
I.
Wasn't
your
would
just
given
that
we've
kind
of
got
our
own
blue
FS
back
and
ferocity,
but
it'll
appear
so
we're
fine.
So
that
word
course
is
just
a
new,
a
new
default
option
due
to
use
that
read
ahead
and
it
makes
compaction
the
compaction
thread
go
from
being
like
95%
busy,
mostly
doing
reads
to
about
like
18%
busy
and
spending
like
82%
of
the
time.
Just
waiting
around.
B
B
B
There's
a
PR
from
Sage
that
gets
rid
of
an
unnecessary
walk
in
the
shard
adopt
4q
I
actually
went
through
and
tested
that
and
the
the
Walcott
profile
in
the
area.
The
walk
profile
that
safe
spot.
It
would
reduce
time
they
didn't
actually
appear
to
do
that.
It
didn't
seem
like
they
actually
be
much
of
anything
removing
it,
but
still
it
was
basically
just
dead
code
that
didn't
need
to
exist,
so
even
if
it
doesn't
improve
anything,
it's
just
probably
good
to
get
rid
of
it.
B
What
else
there's
a
peer
from
Sage
about
putting
blue
of
us
in
the
middle
of
a
shared
device?
Does
the
idea
behind
this
is
that
if
you
are
putting
blue
store
entirely
on
a
hard
disk,
whether
or
not
you
should
have
it
start
at
the
beginning,
the
driver
in
the
middle
of
the
drive
and
can
be
idea
that
it's
in
the
middle
would
drive
the
C
key
either
to
either
end
of
the
platter
will
be
shorter.
B
B
But
yeah
I
don't
know
so
there's
that
one
there
was
some
stuff
that
closed.
B
B
Blocks
positions:
there's
this
file
store
one
here
that
I
haven't
looked
at
at
all,
but
muscle
pretty
minor
in
stages
when
you
merge
it
mark.
E
Yeah
hi,
my
name
is
Andy
Solomon,
my
my
handle
on
the
on
the
meaning
is
kestrels
when
you
get
it
dance,
maybe
after
some
of
this
and
after
the
other
folks
have
gone,
there's
one
less
than
under
no
movement.
That
I
wanted
to
talk
about
today
related
to
the
fast
zero
algorithm
for
CRC.
If
there's
any
way
to
talk
about
that
at
the
end,
yes,.
B
Well,
let's
get
through
the
cold,
these
other
ones
here
and
then
and
then
yeah
like
fun.
He
wanted
you
take
it
away.
Okay!
So
let's
see
the
there
were
couple
others
here,
nothing
too
major!
Oh
there's!
This
set
the
HTV
throttle
cost,
so
that
was
actually
my
PR
and
we're
probably
gonna
change
it
again.
So
basically
we
we've
got
this
concept
now,
where
we're
throttling
and.
B
The
at
the
time
when
I
was
working
on
this
I
thought.
The
idea
was
basically
that
we
were
trying
to
kind
of
eliminate
longtail
latency
by
by
doing
these
throttles,
and
it
kind
of
hurts
like
I,
have
to
throughput,
if
you're
doing
say,
like
okay,
random
writes,
but
but
it
actually
does
eventually
kind
of
start
compressing
things
down
to
to
make
your
average
latency
go
up
a
little
bit,
but
your
long
tail
latency
go
down.
We
did
to
that
in
harvest
it
in
some
cases,
not
in
all
cases.
B
This
is
kind
of
complicated
but
the
gist
of
it
hardware's
each
other
around
1.5
million
mark
for
the
cost.
It
doesn't
necessarily
happen
that
that's
the
case
as
you
increase
the
concurrency.
Those
tests
were
done
with
an
I/o
depth
of
16,
but
when
we
bumped
the
Iowa
that's
way
out,
we
actually
saw
that
that
kind
of
benefit
went
away.
The
long
tail
region
see
benefit,
went
away,
reducing
that
and
but
then
we
also
saw
that
there
were
some
other
effects
that
we
wanted.
B
Basically,
we
want
to
do
see
the
throttle
coming
into
effect,
with
the
idea
that
that
might
actually
help
with
the
the
new
QoS
stuff.
So
we
redid
someone
else
with
a
very,
very
high
I/o
depth,
like
I,
think
it
was
8,000,
concurrent
ops
to
a
single
hard
drive
and
we
kind
of
saw
that
this
value
shifted
to
be
closer
to
around
600,000.
When
we
saw
the
throttle
just
barely
starting
to
kick
in
that's
kind
of
where
sage
I
think
wants
to
try
having
it
to
see.
B
If
it
helps
once
we
can
get
the
kilo
stuff
enabled,
but
yeah
I
don't
know,
I'm
not
I'm,
not
sure
how
how
much
this
will
actually
help
anything,
but
but
at
the
very
least
we
can
see
if
it
maybe
it
does
so
that's
going
to
work
out
with
them
on
that.
What
else
there
a
couple
things
are
updated,
not
opposed
the
only
real
big
one
here.
Is
this
right:
the
deferred
right
handling,
sh
sh,
did
I
think
that
was
failing
previously
and
that
updated
remark,
though,.
D
B
Was
not
testing
that
well,
this
throttling
work
is
basically
the
idea
here
is
that
were
for
hopefully
laying
some
the
groundwork
necessary
to
make
the
the
DM
Cox
of
work
better
or
work
well.
So
this
wasn't
actually
testing
game
clock.
This
is
just
testing
the
throttling
code
that
safe
merged
like
a
couple
weeks
ago,
okay.
B
A
Yeah
so
basically
there's
several
folks
working
on
the
implementation.
Still
right
now,
we've
got
cranked
up
trying
to
get
the
basic
version
of
the
unclocked
algorithm
emerge,
which
is
purely
local.
It
doesn't
include
the
distributed
system
to
have
a
client
yet
and-
and
the
issue
that
we
were
running
into
like
mark
was
talking
about,
but
it
was
that
to
depth
below
us
was
getting
too
long
and
therefore
the
scheduling
at
the
OST
layer
was
not
having
as
much
effect
as
it
should've.
A
A
C
E
From
moran
--this,
so
I
I
don't
work
for
Moran
test
or
whatever,
but
I
had
an
interest
in
this
pull
request,
and
it
was
impacting
some
of
the
work
that
I
was
getting
ready
to
propose
and
so
I
wanted
to
test
the
performance
of
it.
I
put
so
basically
what
I
found
is.
This
looks
much
faster
than
anything
else.
That's
proposed
on
any
of
the
other
architectures.
So
when
you
pass
in
a
buffer,
that's
a
null
pointer.
E
It
assumes
that
the
length
is
is
entirely
zero
filled
and
on
Intel
it
calls
a
specific
function
to
calculate
V
CRC,
in
that
case,
on
power,
architectures
and
arm
architectures,
that
fast
zero
assembly
code
is
not
present
and
I
was
getting
ready
to
build
some
fat
zero
or
propose
some
fat
zero
code
for
PowerPC
architecture.
What
this
pull
request
does
is
implement
that
in
C++
and
it's
way
faster
than
the
Intel
assembly
version
and
the
code
tests
find
the
values,
look,
look
correct
across
different
architectures
and
different
algorithms.
B
E
B
Skirt
died,
I
suspect
sage
may
want
to
be
the
one
that
ends
up
proving
it.
Well.
Maybe
we
should
just
wait
for
a
me
Valentine
right
now,
if
you
coming
back
but
yeah
well,
how
I'll
see
if
I
can
bug
sage
this
week
and
see
if
we
can
get
his
eyes
on
it?
You.
B
Alright,
otherwise,
there's
a
couple
of
other
things:
the
adaptive
throttle
PR
here
that
hasn't
been
a
whole
lot
movement
recently
on
that.
B
Scale
it
oh,
this
separating
the
key
vsync
KB
sinks
right
into
two
parts
that
was
last
I
heard
that
was
dead.
Locking
that
would
be
a
nice
thing
to
get
in
if
we
can,
just
because
you
teasing
thread
is,
is
basically
like
busy.
B
And
have
a
question
of
Roxy
be
cash
versus
blue
effect.
Cash
is
another
interesting
one,
because
if
you
use
the
Xerox
BB
cash
aggressively,
then
you're
doing
all
the
ink
OD
code
before
it
hits
cash,
but
it's
smaller.
So
you
can
pick
more
stuff
in
cash
where's
on
the
blue
plus
side.
Oh
wait!
No
long
time,
video,
no
cash,
not
with
us
right,
yeah,
I'm,
actually
I
should
probably
find
out
why?
What
what
kind
of
Direction
alpha
professor
ed
Cash's
versus
using
the
block
cash
endoscopy
I'm,
actually
not
totally
sure.
C
Yeah,
okay,
so
I
see.
B
Yeah
so
KB
sync
thread
a
lot
of
time
in
their
computers,
for
a
small
random
writes,
appears
to
be
spent
doing
key
comparisons
in
raqqa
defects,
EB,
so
it'd
be
great
to
see
and
can't
dig
into
rocks
to
be
more
in
figure
out.
If
there's
anything
that
can
be
done
there
and
also,
probably
think
about
ordering
and
whether
or
not
we're
we're
kind
of
giving
it
a
data
that
that
is
a
easy
for
it
to
do.
Compare
whereas
kind
of
mentioned,
maybe
there's
something
to
do
there,
but
I
have
willfully.
F
Okay,
so
I
have
actually
pulled
out
the
pr
that
the
ads
I'm
missing,
CRC
cash
made
spur
counter.
You
think,
will
help
out
tracking
if
and
how
effective
in
the
CFC
cash
and
buffer.
If
what
I
was
expecting
to
see
that
the
is
that
the
the
CRC
cash
in
buffer,
this
is
pretty
much
useless
or
close
to
your
left,
and
what
I
found
out
is
that
it's
actually
quite
opposite.
In
case
of
five
store,
there
was
a
twenty
three
percent
efficiency
and
forty
four
percent
and
efficiency
on
blue
store.
F
F
Can
take
this
link
in
another
gear,
you
need
to
insert
the
person
to
20
between
education
and
to
guess
so.
This
will
actually
work
anyway.
I
haven't
performed
in
a
significant
testing
with
Dave,
but
these
initial
results
are
quite
interesting
and
I
hope
in
this
will
help
out,
seeing
whether
we
should
get
rid
of
this
certification
or
just
reduce
it.
F
Somehow,
because
who
another
thing
I
have
seen
is
that
in
most
cases
the
GRC
cash
was
just
collecting
of
one
element,
meaning
that
just
one
CRC
was
requested
and
cashed
and
nothing
else,
I
don't
have
a
right
now,
any
information
on
whether
and
this
was
followed
by
any
cash
gate
or
not
or
just
or
this
was
just
just
a
single
access,
but
still
this
is
a
something
we
can
also
think
you
can
test
out.
Further
goes
around.
D
F
I
have
posted
a
link
on
the
IRC
on
safe
devil
to
the
textile
that
pretty
much
displays
during
the
testing.
How
large
was
the
CRC
cash
ant?
In
most
cases
it
was
simply
empty
and
next
one
was
a
of
one
element.
So
in
most
cases
we
could
get
rid
of
this
CRC
map
and
use
some
kind
of
single
container
and
contain
just
two
last
GRC
calculated,
and
that
would
probably
reduce
a
memory
and
to
view
utilization,
especially
during
destruction
or
buffer
pointers.
But
this
is
right
now
just
what
I
guess.
F
C
D
D
D
F
No
because
I,
when
I
was
a
issuing
that
PRI
came
Academy
with
an
a
blue
cluster
for
my
axis.
So
disorder
is
somewhat
educated
grass,
because
if
you
are
copying
a
lot
of
data
back
and
forth,
then
you
are
pretty
much
it
crashing
the
processor
caches
and
you
are
going
to
type
it
out
it
for
that
and
it's
it's
always
a
hard
thing.
Also,
you
will
have
to
pay
penalty
for
extra
memory.
This
was
so
getting
great
deals.
Extra
going
to.
D
D
F
F
A
D
So
are
we
already
actually
in
their
provisional
performers,
are
meeting
we
already
booked,
then
why
we
also
care
about
their
performance,
especially
the
eternal
agency
and
energy
image
in
situ,
because
the
applications
right
now
we're
all
against
the
self
Kasper
is
so
demanding
in
terms
of
their
dependency
and
energy
agency.
So
we
can
air
a
corner
of
the
chance
and
shapes
of
improving
attendance,
the
engine
agency-
and
this
is
kind
of
bad
one
I
want
to
share
with
all
of
you
guys,
that's
a
lot
of
the
reason
why
I
asked
a
character
off
work.
D
The
PRP
is
working
on
in
terms
of
calmness
near
others.
We
care
about
their.
You
know
whether
there's
a
data
to
opportunities
claim
whether
they
can
improve
the
tendency
or
not.
So
we
have
several
folks
in
there
a
phone
call
as
well.
So
just
in
case,
if
I
cannot
explain
here
so
please
come
come
out
and
support
support
me
with
more
detail
explanation,
so
we
have
a
impression
for
the
staff
recovery.
We
we,
we
have
a
lot
of
times,
especially
a
phenol
injection
task
in
our
cell
culture.
D
Basically,
we
try
to
down
one
OST
to
know
why
you
were
iraq
to
see
how
the
subcontract
behaved
when
the
recovery-
you
know
guests
are
so
we
don't
want
to
weather
the
front
I,
along
with
the
application,
I
will
get
a
block
or
all
the
time,
because
some
of
the
our
application
is
the
e-commerce
application
and
there's
no
way
to
block
of
them,
even
for
example
down.
It
will
give
the
aisle
from
e-commerce
our
team
down,
which
was
meaning
either
one
payment
or
your
payment
or
either
one
real.
D
You
know
e-commerce
transaction
can
o'clock,
there's
no,
that
is
for
sure,
that
are
going
to
be
accept
acceptable
for
us.
So
recovery
is
aware
of
their
place.
We
pay
lots
of
attention
because
you
know
our
detail
is
normal
in
the
big
cluster.
We
are
running
right
now,
so
we
try
to
see
whether
there's
a
way
to
mitigate
their
io
impact
theory
recovery.
This
listen.
D
This
is
a
background
of
asset
recovery,
improvement
and
their
way
to
go
ahead.
There
are
any
questions
are
given
all
I'm
going
green
kibbutz,
so
we
do
from
research,
of
course,
in
the
surf
community
and
the
founder.
Several
for
request,
which
he
already
wrote
on
the
4th
of
our
action,
is
from
our
church
on
la
is
from
x.
Ky
is
a
he
has
a
patch.
He
has
a
pull
request
in
there.
D
It's
a
we
call
it
a
synchronous
recurring,
basically
one
so
there's
a
io
coming
during
the
curry
ii
and
the
nominal
I'll.
You
know,
ask
you
the
first,
then
the
correctly
we
did
a
a
synchronized
way
little.
This
is
the
part
about
the
as
militaria
that
otherwise,
the
partial
recovery
is
actually
from
young
he's
from
many
years.
Another
big
internet
company
in
China
he
try
cousin
only
their
theta.
D
The
mismatch
is
only
a
Porsche
Porsche
Porsche
of
their
white
object.
For
example,
you
have
a
full
mega
object,
but
they're
only
a
portion
of
their
object
is
missing
or
user
on
YouTube
Akari,
it's
just
like
it
was
snow,
educate,
there's
no
demonstrated
copy
of
their
for
Megan
for
their
data
recovery.
So
this
is
kind
of
part
recurring.
This
mean
of
our
recurring
improvement
is
on
these
two
pairs.
Of
course,
I'll
take
several
Ducks
are
related
to
the
chairs,
but
we
also
have
an
imitation
dinner
I'm.
D
Going
to
be
same
while
it's
true
PR
and
imitation
pass
so
far
is
address
our
problem.
We're
well
in
the
protocol
in
the
vodka
storage.
Long
story
is
in
there,
RVD
give
our
body
work,
but
therefore,
to
part
will
continue
so
much
out
there.
You
know
so
far
has
or
academia
support,
so
we
only
it
so
far
suppose
there
are
Biederman
IV
d.
D
Here's
our
touch,
environmentalism,
our
experiment,
our
experimental
environment,
but
they're
in
our
real
deployment.
We
have
way
bigger
faster
than
that,
so
we
we
just
today
well
share
with
the
OVA
guys
the
performers
8ax,
we
already
achieved
in
very
small
test
environment
all
times
in
one
have
a
three
servers
to
me
for
Ortiz
and
of
course,
we
use
SF
10.2
profile,
your
worship
trance
above
the
patch,
and
there
we
create
a
100
in
the
RVD
image
and
I
file
for
cable
right.
So
here
there's
a
data,
so
I
mean.
Are
you
on
the
line?
D
G
So
police
that
I
basically
I
run
five
minutes
head
to
just
even
if
women
were
both
attached
to
see
the
timeline
here
so
and
it's
a
ella
beginning,
I
run
fi
all
say
in
a
60
seconds,
60
seconds,
I
stopped
a
war
most
OST
and
in
in
a
two
minutes
timeline
I
will
start
the
the
one
I
just
stopped,
so
the
recovery
begins
I
capture.
All
the
bio-data
see
how
Allah
improvement
look
like
Tim.
Could
you
go
to
the
next
drive
sure.
G
Oh
yeah,
yes,
okay,
so
the
first
one
is
the
improvement
problem.
How
to
recover
APR
feel
the
green
part.
The
green
part
is
to
the
recovery
part
and
then
the
red,
the
red
line,
is
the
patch
the
version
of
recovery.
The
blue
eyes
done
Hatcher
version,
so
we
see
from
here
we
get
a
lot
of
because
a
lot
of
improvements
from
the
partial
recovery
and
they
actually
recover
much
faster
than
the
unpatched
version
in
about
I
think
about,
let's
say:
30
30
seconds
yeah
about
30
seconds.
G
H
G
D
G
D
So
normally
we
will
rejoice,
we
we
feel
we
get
our
80%
almost
60
to
80%
of
performance
again
in
terms
of
Taylor
agency
and
origination,
see
a
curated
carry
stage
the
lineage
evening.
We
normally
see
the
out
go
to
zero
if
we
courage
happen
without
any
patch
with
the
partner
requires
patch.
Of
course,
we
fix
in
several
bugs
with
okay
meet
our
mission
about
it,
and
then
we
teaching
our
speaking.
We
almost
get
80%
I,
ask
improvement
and
all
didn't
see
intern
agency.
D
D
Here
is
what
we
think
are
four
to
Denise,
especially
we
are
looking
for.
You
know,
help
from
Josh
for
more
comments
and
their
more
input.
We
still
have
fun.
Some
I
will
drop
to
zero
German
German
states,
even
though
he's
aware
shot
in
like
one
second
gen
one
second
identify
again
with
we
try
to
avoid
any
of
chance.
She
to
that
I
drop
to
zero
from
the
kind
side
which
is
lovely
from
the
e-commerce
and
I
our
Mexico
applications.
This
is
the
white
right.
Now
we
try
to
address
thinkin.
D
One
is
of
course
right
now:
their
passion
only
works
or
multiple
tile
works
for
RVD.
We
don't
have
a
chance
to
check
for
our
G
Domino's
ffs,
so
this
one
is
a
bit.
If
it's
a
big
concern
over
here
and
there
we
right
now,
I
need
to
do
more
intensive
attack.
We
already
do
another
test,
I,
don't
fix
the
server
box,
and
this
over
box
need
to.
You
know
to
only
fix
accessing
a
three
or
four
bugs.
The
boxes
in
Italy
Auto
attribute
are
attributed
inconsistent
or
OMAP
into
incompetent
inconsistent.
D
D
You
know
that
gives
into
the
master
does
need
to
be
decided
by
side.
Yours
Josh,
and
there
occurs.
We
so
far.
Don't
have
a
way
to
do
forward
to
support
a
recovery
of
the
EC
and
we
so
we
made
our
cars
are
going
to
use
easy
as
well,
so
how
we
can
do
support
you
see
Jerry
Jerry,
recovering
Oh
without
a
proper
to
0.
This
is
or
another
concern.
We
have
right
now,
yeah
mark,
that's
pretty
much.
We
have
today.
So
we
we
share
with
you
know:
youth
advisor
John,
sighs,
prophetically
smile
here
so
Josh.
A
D
Yeah
I'm
going
to
send
you
their
links
if
you
like,
but
here
basically
Whitney,
is
a
to
pizza
girl,
PR
recovery,
PR,
we
are
working
on.
We
we
also
have
summer
bug
fix
towards
the
two
tiers
and
therefore
you
know.
Oh,
let
me
know
what
you
think
and
we
can
you
know
either
Walker
together
and
through
the
email
or
through
the
rc4.
A
Yeah,
so
in
terms
of
the
async
recovery
PR.
If
there
are
a
number
of
ways,
those
things
that
probably
aren't
quite
correct
for
in
terms
of
upgrades
and
like
you
mentioned
in
terms
of
just
handle
racial
coding
yet
so,
and
that
one
that
I
was
working
on
a
slightly
different
approach
today
and
hope
had
hoped
to
have
ready
for
luminous
and.
A
B
A
D
I
also
want,
but,
as
you
know,
we
do
have
a
real
woken
all
work
on.
You
know
roaming
every
day
on
with
these
two
cars
there
is
so
far
were
stable.
Now,
so
we
can
consider
about
it
as
well.
That
would
be
great
if
we
attempt
green
some
of
this
some
of
these
implementation
into
your
own
implementation.
That
would
be
creative
because
and
I
said
we
want
to
have
extremely
good
our
low
turn
agency
and
their
performance.
Well,
what
are
we
dealin
care
of
us
very
prominent
next
next
week?
D
D
A
D
D
D
Little
full
marketer,
probably
next
week,
and
then
you
have
up
to
our
allocated
another
10
minute
sessions
and
we
gonna
share
another.
You
know
performance
improvement
we
made
and
is
it?
What
is
a
easel
is
like
you,
crazy
decreases
their
tendency
almost
four
times
with
a
center
work
node.
So
we
also
want
to
share
with
you
guys
and
looking
for
art
so
for
your
opinion.
Well
know.
B
D
H
Sure
sure
so
so
we
are
trying
to
optimize
the
performance
on
the
OST
side,
and
so
so
far
we
have
done
some
latency
breaks,
so
we
saw
some
unknown
agencies
spend.
On
the,
for
example,
a
few
transaction
function
called
when
the
key
key
worker
are
propelled,
okey
and
finally
caused
a
few
transaction
which
involved
teachers
or
actual
freedom.
We
are
using
the
booster
booster
again,
so
blue
stores
to
transaction
will
be
triggered
and
the
PG
worker
will
execute
that
function.
Action
needs
we
in
the
holding
repeating
log
we
are
thinking
about.
H
Can
we
make
these
functional,
as
are
to
be
executed
by
thread
for
inside
the
objects
or
back-end
not
executed
by
the
kikiwaka
thread?
So
PG
workers
registered
actually
to
the
put
this
transaction
requests
to
the
and
those
thread
for
inside
the
object
and
the
Bluestone
back
inside,
and
then
we
can
return
that
key
key
workers
with
and
release
the
auntie-ji
lock,
so
that
actual
thread
for
inside
the
blue
spot.
We
actually
excuse.
H
B
Depending
on
the
the
kind
of
workload
we
definitely
see
contention
on
the
the
the
PG
lock,
especially
with
something
like
4,
K
sequential
writes,
that's
a
early
famine,
one
layer
that
looks
like
at
least
in
blue
store
head
if
I
also
recently
Billy
simplest,
RPG,
Wolcott
profiling,
you'll
see
at
the
PG
office.
It's
very
much
contended.
D
Yeah,
so
we
do
some
of
that
experiment.
Well,
experiment
in
outside
we
lease
I
think
you
can
scream
like
a
50
microseconds
but
there,
but
we
right
now
100%
sure
whether
it's
a
safe
because
the
wake
and
like
a
make
it
awesome
her
eyes.
You
know
in
the
beginning
ways
have
a
similarity.
We
have
already
have
a
right
to
synchronize
but
they're.
D
All
the
metadata
annotation
is
steel,
synchronize
investing
right
away,
but
right
now
between
the
PG
and
objects,
though
we
can
mega
make
it
a
summarized
away
even
longer
so
which
can
release
the
P
log
in
sooner
than
later.
For
sure,
goodness
see
you,
there
littles
a
butter.
So
far
we
don't
know.
When
is
we
are
about
100
percent?
We
use
the
Wii
Remote,
apparently
research,
but
we
still
need
looking
for
sorry
Josh,
otherwise,
whether
it's
safe
or
not,.
A
Aside
to
tell
her
it
if
I
look
at
the
code
in
more
depth,
and
but
one
thing
that
springs
to
mind
is
that
it
might
make
more
sense
to,
in
the
long
term,
to
try
to
make
the
interface
and
instead
of
doing
single
callbacks
for
each
individual
completion,
have
a
single
callback
that
I
gets
called.
The
then
goes
and
inspects
all
the
I/os
that
have
been
completed
kind
of
like
an
eye
of
event
style.
A
D
Wait
I
think
you
know
we
we
going
to
implement
and
they're
also
awkward
we're
going
to
have
a
performance
data,
but
in
the
meantime
we
were
getting
in
touch
and
there
especially
you
and
there's
hatch,
and
it's
just
in
case
there,
whether
this
is
going
to
impact
the
suitability
Allah.
We
aren't
sure
what
is
safe.