►
From YouTube: CDS Jewel -- Ceph 0 day for performance regression
B
B
B
Yeah
I
want
this
totally
about
performance,
things
about
the
whole
separate
and
maybe
a
little,
and
we
want
to
judge
this
probe
problem.
We
think
among
the
developer
staff
and
with
two
staff,
cluster
kabobs
and
I-
think
we
like
old,
necessary
way
to
avoid
the
most
improvement
preset.
For
example,
if
someone
give
up
fluence
the
four
month
except
we
can't
secure
a
monster
day,
we
kind
of
give
a
detail
not
number
about
how
much
the
this
prosthetic
gill
and
the
to
this
problem
I
said,
will
affect
other
companies.
For
example.
B
Maybe
we
consider
to
improve
the
rights
of
a
spider.
Actually,
we
had
secures
right
vamos.
Maybe
it
is
a
good
example
and
at
the
second
is
unknown:
performant
accreditation
for
angry,
aware
of
fish
pizzetta,
for
example,
from
dumping
and
to
hammer,
and
we
get
many
of
improvement
for
level
for
faster,
deeper
dispatch
and
the
PG
shutting
and
the
something
at
the
store.
But
actually
we
don't
get
enough
improvement.
B
Actually,
this
is
a
sim
problem.
We
have
some
formal
keola,
but
we
don't
know
after
the
past
much
the
papr
and
the
force
is
actually
it
is
for
users.
Others
have
some
customers
confused,
but
we're
he
upgrade.
He
will
find
a
lipoma
most
decorated,
because
different
warship
have
different.
Config
reveal
effective
and
effects,
and
there
may
be
a
felon
hosted
as
fast
or
sex
or
pcs.
We
need
to
improve
our
periods
where
we
afraid
that
to
Hama,
because
we
have
much
more
computer
time
and
we
need
two
more
objects.
B
B
So
Howie
about
this
problem,
I
think
at
first
the
way
into
how
advanced
test
class,
which
is
double,
and
we
need
to
do
delay
or
daily
or
/
PR
PR
promised
Tara,
and
then
we
could
like
mercury
level
for
harbor
PR.
We
needed
to
need
a
powerful,
a
level
4
or
performance
related
appear.
Maybe
it
is
not
a
performance
agenda,
but
we
needed
food
at
this
level
if
we
think
it
is
a
factor
of
most.
B
B
And
I
think
how
we
do
this
and
I
know
intel,
has
a
projector
called
a
seat
and
there
maybe
we
can
cover
a
disease,
but
before
this
I
I
don't
help
detail.
I
should
capture
about
how
we
do
della
de
Provence,
but
mark
I.
Don't
ask
you
or
some
ideas
about
it.
So
at
first
we
have
a
test
cluster.
Maybe
we
can
use
a
simple
stuff
to
play
or
using
a
QA
cast
for
put
framework
to
do
this
posted
okay
before
we
actually
do
this
later
and
as
yes,
we
have
woken
load
generation.
B
The
first
day
the
first
class
is
fio,
I
will
do
or
directly
with
mark
I've
got
a
little
bench
or
I'll
d
bench.
Maybe
we
can
pick
a
pic
some
of
these
and
then
we
could
run
some
critical
unique
in
that
hit.
You
need
to
test
program
such
as,
for
example,
CSE
or
module
internal
at
mark,
for
example,
d,
Bobby,
B,
marker
butter,
the
technical,
don't
touch
you
OSD,
so
it's
mean.
Maybe
we
could
vocals
at
the
body
and
we
can
enter
the
Bobby
Darin
tell
senator
cast
was
the
actually.
B
So
this
is
what
organization
and
the
we'd
get
hell
mattress,
for
example.
That,
basically,
is
on
him.
For
example,
we
how,
how
much
time
we're
on
a
test
and
the
week
needed
to
gather
this
on
him
and
at
the
time
that
we
needed
to
gather
of
self
of
damp
for
the
summer
some
numbers
and
it
depends
on
the
detailed
test
at
a
third
and
if
we
don't
a
file
or
I'll,
do
this
halibut
back.
B
We
need
to
collect
her
the
system
mattress,
for
example,
we
need
to
collector
scenes
from
our
state
and
Colonel
since
is
laboring
for
and
and
the
net
worth
or
CPU
range
memory
ology,
including
disco,
display,
village
and
disco
detail
about
it.
He'll
be
okay,
chance
or
F
chance,
and
only
way
my
sins
and
a
whim
state
or
other
things,
I.
B
What
is
this
shadow
makeup
home
Istanbul
or
make
her
LPS
last
week
with
am
home
Fernando?
If
we
already
know
LPS
is
Istanbul,
it
is
maybe
makes
us
some
time
with
100
out
here.
Somehow
is
only
15
lbs,
so
we
need
to
make
this
match
available.
The
same
is
the
video
was
diverted
or
implement
afford
a
patch
that,
for
example,
already
know
we
can
get
100
LPS
and
later
we
can
get
stable,
300
iops
yeah
under
the
study
is
a
long
time.
I
think
it
is
simple
yeah.
B
So
I
think
this
is
some
guidance
about
how
to
set
up,
at
least
in
social
infrastructure,
and
we
need
to
some
hardware,
for
example.
Maybe
we
do
not
need
it
to
logic,
test
acosta,
but
away.
At
least
we
need
two
three
or
five
physical
machine
to
do
this
big
because
vm,
maybe
not
good
enough.
Wood
has
the
list
because
we
are
more
itself.
B
Hell
has
some
problem
for
force,
but
work,
and
maybe
it
can
give
some
Azure
without
about
for
us
and
the
another
language
too
long
test
caster
daily
and
you
are
reported
to
the
PR
and
the
on
my
list
or
on
harbor.
Then
we
at
first
we
need
to
focus
on
basically
I'll
pass
and,
for
example,
we
focused
on
fast
or
and
the
PG
sings
and
the
simple
mataji,
and
maybe
the
body
without
catcher,
or
enable
catcher
and
a
little
red
of
the
get
away
with
SS
three
and
the
staff
s.
B
C
So
I
did
wanna
point
out
that
there
are
a
couple
of
a
somewhat
more
convenient
workload
generation
tools,
there's
a
SEF
small,
a
bench
and
it
can
be
run
directly
against
the
oh.
So
it
was
here.
The
person
who
just
left,
I
think
was
the
person
you
want
to
talk
to
about
the
QA
anyway
was
sick
right.
Small
io
bench
has
a
version
that
runs
directly
against
rato
center.
C
D
Yeah
there's
also
a
while
back
Jason
made
any
liberators
for
running
rbd
tests
against,
and
your
zest
in
general
I
think
there
might
be
some
performance
tests
that
we
can
run
against
that
as
well.
Just
a
benchmark,
Liberty
itself
or
the
client
inside
in
general,
without
involving
full
stack
and
way.
E
C
Would
be
a
way
of
testing
the
testing
rgw
without
a
ratos
closer,
you
can
see
exclusively
the
CPU
overhead
of
registry,
w
yeah
yeah
yeahs,
yes,
okay.
What
was
the
other
thing?
Oh
I
would
like
to
point
out
since
mark
isn't
here
that
if
you
create
a
simple
Seth
deploy
script,
what
you
get
is
the
SEF
benchmark
tool.
Once
you
solve
all
the
problems
inherent
in
that
I
would
yeah
yeah
hailey
back
to
suggest
very
strongly
that
you
make
it
a
tooth
ology
tool
and
not
not
roll
a
third
cluster
setup
system,
yeah.
C
No
I
think
you
should
do
in
ratos.
Cece
BTW
is
cool,
and
that
might
be
an
appropriate
thing
to
do
later.
But
we'd
like
to
be
able
to
run
this
in
a
more
lightweight
fashion.
Cvt
is
more
for
setting
up
without
that
spark
for
details,
but
more
for
setting
up
larger,
more
complicated
test
where
you
want
to
set
up
run
a
bunch
of
different
parameter
combinations,
but.
A
C
G
B
D
C
B
D
B
G
C
C
H
So
you
say
the
issue
is
odd
for
coast
road
tests,
the
setup
is
very,
very
small,
I
think
the
URL
0
ma
three
or
four
osts,
where
this
be
a
issue.
Okay,.
C
So
there
there
are
several
different
the
length
and
breadth
of
ways
you
can
do.
Performance
testing
is
kind
of
astonishing,
though.
If
what
we're
worried
about
is
code
level
regressions
in
something
using
too
much
memory
or
taking
too
much
cpu
time,
then
we
actually
want
a
small
cluster.
Not
a
big
one.
Big
clusters
have
a
tendency
to
hide
problems
behind
big
walls
of
CPU.
That
would
be
easier
to
show
up
in
a
small
test.
C
However,
if
we're
looking
for
something
like
a
scaling
bug
where
we're
connect,
we're
opening
too
many
connections
per
OSD,
something
like
that
you're
right
that
wouldn't
show
up
in
a
30
SD
test,
but
on
the
other
hand
it
would
also
take
a
lot
more
resources,
so
we'll
want
both,
but
the
the
tooth
ology
one
is
still
valuable
and
bonus.
We
can
do
it
as
part
of
the
normal
rate.
F
But
the
problem
is
that
earlier,
a
lot
of
changes
that
only
only
benefit
you
know
small
setup
but
not
benefit.
You
know
largest.
It
happen,
for
example,
for
the
for
the
xio
messenger
async
method
doing
it.
You
know
test
result
and
also
helps
in
what
skill,
but
that's
a
helping
large
go.
So
if
you
don't
have
much
skill,
you'll
never
see
normal.
F
And
another
another
problem
is
that
into
the
totally
test
that
we
can
come
by
a
lot
of
hedge,
that
into
a
single
row,
but
in
the
performance
test
seemed
so
we
can't
do
that
so
and
that
the
performance
test
is
some
kind
of
sometimes
very
heavy.
For
example,
if
you
want
to
do
some
retest,
you
need
to
hold.
F
C
The
other
problem
of
the
machine,
I'm
necessarily
yeah,
yeah
uniform,
that's
all
true.
We
may
not
want
to
build
them
into
the
actual
sweet,
but
it's
still
valuable
to
structure
them
as
a
totality
task,
because
that
way,
you
don't
have
to
rewrite
the
part
where
you
pull
from
the
ammo
file,
the
particular
branch
you
want
to
run
find
it
built
in
the
repository
pull
down
the
packages
install
the
packages
set
up
the
cluster
in
the
mons
distribute
keys,
set
up
ssh,
it's
just
tedious
and
annoying
all
of
that
stuff
can
be
handled
by
tooth
ology.
C
F
C
C
Plus,
if
speaking
as
just
a
developer,
it's
really
easy
for
me
to
run
technology
test.
It's
really
hard
for
me
to
go
to
a
site
read
a
bunch
of
directions
on
first
I
need
to
pull
some
machines
out
of
a
thing
that
I
need
to
set
up
an
orchestration
framework
and
set
up
Steph
cluster.
Then
I
need
to
run
this
new
framework
against
it,
which
is
you
know
harder.
It
would
be
easier
if
we
could
just
use
the
existing
structures.
B
C
F
B
F
F
C
Can
do
it,
you
can
do
it
for
the
same
reason
that
you
can
do
it
in
QA.
If
you
find
an
issue,
then
you
have
to
pull
the
package
of
PRS
apart,
and
this
is
before
you
merge
to
master
by
the
way.
This
is
an
integration
branch,
so
you
start
testing
the
individual
ones
or
inspecting
them
and
removing
them
based
on
the
gas.
Until
you
find
a
clean,
subset,
merge
that
and
then
go
back
and
figure
out
which,
which
one
was
broken
you're
right.
It's
not
perfect,
but
I
mean
performance.
C
Testing
always
takes
a
lot
of
resources,
so
it'll
be
really
difficult
to
do
on
in
every
PR
basis.
Certainly
if
you
want
to
do
the
20,
LSD
or
40
OSD
cases,
which
we
probably
won't
be
able
to
do
on
even
that
basis
now,
but.
F
Us
go
ahead,
I
mean
you
know
that
we
understand
the
code.
So
if
you
see
the
failure
or
not
a
failure,
you
can
easily
tell
which
PR
caused
the
problem.
But
if
you
use
your
performance
decoration,
it
is
harder
for
you
to
find
which
change
cause.
The
performance
decoration-
and
this
is
yeah.
C
So
this
is
it,
but
it's
an
Aries
color.
It's
a
lot
of
subversive
pessimism
trade-off
right,
so
it
costs
you
more
time
in
the
case
that
you
thought
that
that
you
do
find
a
performance
progression,
but
it
costs
a
lot
less
time
in
the
case
that
you
don't
right,
because
you
can
always
fall
back
to
testing
every
PR
individually.
But
if
you
get
lucky
and
nothing
cause
performance
regression
then
yay
you
get
to
merge
it
after
one
run,
instead
of
yep.