►
From YouTube: 2019-07-18 :: Ceph Performance 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
All
right:
well,
maybe
we
should
just
get
started.
So,
let's
see
his
been
a
couple
weeks,
we
had
All
Hands
meeting
last
week
and
then
before
that's
worth
July,
so
a
little
more
to
go
through
the
normal,
but
not
not
too
bad.
Let's
see,
Adam
has
three
new
PRS
recently
for
his
charting
work.
So
that's
really
exciting.
A
I
have
not
gotten
the
chance
to
look
those
over
yet
but
I'm
hoping
to
at
some
point
here
since
for
me
to
integrate
my
m-my
double
cashing
PR
with
it
or
at
least
get
my
my
stuff
working
in
television.
Oh
wait,
hey
sage,
I
just
started
doing
over
the
PRS,
but
didn't
get
very
far
so
you
want
to
jump
in.
You
can.
B
B
B
Yeah
I
don't
know,
maybe
I
mean
this
is
a
Jenkins
build,
so
be
clean,
yeah
I,
imagine,
there's
a
practice
to
be
willing
order.
I,
don't
know
something.
There's
something
annoying
yeah
yeah
I
didn't
try
reproducing
it
look
away.
So
I
really
know.
Let's
see
this
one
from
Igor
is
a
good
step
in
the
right
direction.
I
think
we're
gonna
try
to
make
it
the
hints
more
explicit.
B
Yeah
but
we're
all
that's
quite
promising
and
then
I
haven't
liked
that
gear
last
one
here
about
the
apptracker
there.
A
A
B
B
I'm
saying
the
whole
point
is
that
we
don't
want
to
make
a
copy,
yes
and
whether
it's
constant
view
is
totally
irrelevant
because
they're
the
same
thing
so
they're
a
pointer
or
appointed
with
the
thighs,
but
by
the
way,
if
there's
no
copy,
though
mm-hmm
I,
don't
know
the
whole
discussion,
but
I
think
that
what
the
goal
should
be
is
to
not
make
a
copy.
Yes,
that's
unrelated
to
string
view
right,
I.
B
A
A
B
A
Always
had
a
second
string,
well,
I,
don't
know
if
we
always
did,
but
we
for
a
long
time,
we've
had
a
separate
string
version
that
we
had
separate
from
the
Casa
CARSTAR
version
right
and
then
that's
when
we
created
a
local
string
inside
event.
But
if
you
passed
it
in
yeah,
there
was
like
a
second
interface
that
we
only
use
sometimes,
and
we
didn't
use
it
in
like
a
robot
and
Patrick
wanted
to
combine
those
all
into
one
thing:
to
make
the
interface
really
clean.
B
D
D
A
The
other
thing
that
this
does
is
it:
it
uses
influence
like
sort
of
a
semi,
lock
list,
I'll,
say
data
structure
for
the
event
holding
the
events.
So,
if
is
luckless
up
until
you
have
a
certain
number
and
then
a
traverse
to
vector
behavior
with
locks
yeah,
it's
good
enough.
It
almost
never
actually
uses
the
vector.
B
C
A
B
A
B
B
A
A
A
B
B
B
B
B
It's
much
better
than
it
used
to
be.
Let's
see
all
right,
let's
see,
turning
out
the
off
track
or
by
default.
I.
Think
consensus
is
mostly
not
to
do
that,
although
users
always
do
it
if
they
want,
they
could.
B
Blue
store
cash
refactor.
Have
you
figured
out
what
was
causing
this
crashes
yet.
A
A
A
The
the
PR
before
I
made
the
changes
for
be
worse
and
listen
that
he
asked
for
didn't.
Do
it
when
we
ran
it
through
QA,
it
was
passing
done
so
I
think
it's
something
related
to
some
of
the
updates.
I
made
I'm,
just
not
I've.
Looked
it
over
I.
Don't
really
understand
why
any
of
those
would
cause
it.
But,
oh
okay,
it's
morning.
B
Cool
okay,
all
right
figures
inside
of
the
tuning
in
progress,
there's
the
sink
fairness.
B
All
right,
I,
don't
know
anything
else.
Here
is
of
note
yet.
B
B
A
lot
of
these
cases,
we've
seen
where
we
see
like
mammoth,
blue
SS,
reads
and
writes,
are
because
somebody
does
a
delete
on
something
that
has
a
bazillion
map
entries,
and
so
it
grates
this
mammoth
transaction
that,
like
deletes
individual
records,
that
notice
tee
times
out
and
well
if
that's
gets
all
pissed
off
and
whatever
I'm
really
a
range
of
avoids.
All
that
Bears
need
to
fix
that
build
the
build
thing.
We
see
those
I
guess,
yep.
A
A
B
A
A
B
A
B
A
B
Should
we
should
we
figure
out
what
version
we
want
to
put
a
Nautilus
and
put
that
in
master
so
that
that's
what
we
test,
if
I
wonder
if
we
want
fix
that
something
or
if
we
want
to
like
the
oldest
I've,
got
19
or
whatever
there's
like
if
30
major
changes,
then
you
might
want
to
be
a
little
bit
more
conservative
sure
or
not.
Maybe
we
just
want
the
best
I,
don't
know,
yeah.
A
B
Probably
triple
the
story:
oh
okay,
you
do
that
review
offline.
Probably
looking
at
this
performance
results,
spreadsheet
whoops
for
the
for
the
compaction
threads
yeah,
but
the
the
blue
and
red
chart
is
two
cores
versus
all
cores
and
that's
not.
Oh
sorry,
excellent,
doesn't
look
like
going
from
one
to
two
compaction.
Threads
changes
that
much
it.
A
Right,
yeah,
the
the
test
time
goes
down.
The
/
compaction
time
can
go
up,
because
now
you
can
have
overlapping
compactions,
so
each
individual
compaction
event
can
take
longer
is
overlapping
with
another
one,
but
the
actual
client
side
like
max
latency
goes
down
because
you're
not
causing
have
like
hierarchical
stalling
chains
as
much
okay
having
too
many
compaction
threads,
especially
with
you
know,
limited
numbers
of
course
is
not
good,
so
I
think
conservatively
increasing
it
to
two
is
probably
what
we
want
to
do.
B
A
Yeah
Adams,
not
here
I,
don't
think,
is
he
I,
don't
see
them
so
we're
we're
working
on
figuring
out
who
we're
gonna
layer,
my
blue
store
or
sorry,
the
the
the
oh,
no
double
cashing
fix
on
top
of
his
stuff
right
now
and
then
include
the
conversion
that
we'll
need
his
conversion
tools.
This
we
need
to
convert
first
sharding
anyway,
then
we
can
also
convert
into
the
separate
column
family
layout
that
I'm
introducing
there
bro
node
versus
others
have.
B
A
A
So,
but
in
the
sharted
world,
what
we'll
need
to
do
is
is
a
little
different
right,
because
we
don't
want
like
an
explosion
of
calm
families.
We
almost
certainly
don't
want
to
create
lots
of
per
shard,
calm
families,
so
my
guess
is
we'll
probably
just
want
to
have
own
own
separate
and
everything
else
be
in
the
cards
own
family
and
then
we'll
have
like
two
column.
Families
per
charge
is
I'm,
guessing
kind
of
what
we
will
probably
want,
and
we
we
have
hopefully
two
caches
I.
B
Mean
the
Khan
family
is
let
you
separate
out
cash
for
you
right,
and
so
the
ones
right
now
are
or
OMAP,
which
I
think
makes
sense.
That
should
be
at
a
different
service,
tier
or
whatever
than
the
internal
metadata
of
the
store.
The
other
one
is
the
PG
meta
map
of
the
PG
log
map,
she'd,
probably
be
treated
differently
than
the
regular
user
row
map
sure
and
the
other
one
her.
A
B
B
D
A
B
A
A
We
we
like
do
this
kind
of
crazy
thing
where
we
I
think
we
have
an
internal
and
external
queue
and
then
we
be
like
well
put
stuff
on
one
and
then
periodically
will
move
it
into
the
other.
To
avoid
like
having
work
in
the
internal
queue
block
blocking,
you
know
for
the
external
one,
but
we
still,
we
still
have
a
lot
of
work,
that's
being
done
there.
A
B
Yeah
yeah
and
I
mean
someplace
better
to
you.
If
we
can
like
avoid
some
of
the
weirdness
at
you,
I.
B
D
Got
make
sense,
yeah
I
think
once
we
have
liked
racing
fully
integrated
at
the
OSD
I
got
some
good
candidate
for
replacing
Becker.
B
The
nice
thing
about
Becker
is
that
it's
always
on,
and
it
gives
you
like
in
a
production
environment.
It
gives
you
diagnostic
information
like
when
the
problem
happens,
tells
you
where
exactly
where
and
how
it
was
blocked
and
not
I'm.
Not
sure
tracing
can
do
that
unless
you're
tracing
all
the
time
I.
B
D
A
So
so,
right
now,
I
think
ma,
Xian,
Tang,
I'm
sure
he'll
make
you
look
at
PR,
but
he
I
think
he
saw
something
will
get
25%
performance,
difference
or
or
roughly
something,
though
vicinity
as
we
keep
making
things
faster.
That's
just
gonna
get
worse
right
now,
like
we
can
make
the
OP
tracker
faster.
We
can
make
it
good,
but
it's
gonna.
D
Yeah
the
keepers
already
started
working
on.
That's
when
the
inspectors
would
they
just
the
each
at
I
was
black
on.
A
C
A
C
A
D
D
A
D
Not
this
week,
but
maybe
next
week,
if
we
can
get
Sam
here,
he's
been
heading
some
tracing
the
tableau
store,
we're
in
helping
investigate
that
us
stuff
and
spent
some
interesting
stuff
with
I'm
being
spent
more
in
the
allocator
and
more
variants
coming
from
that
than
from
Roxy.
B
is
not
what
we
expected
at
all,
but.
F
Cool,
oh
my
mark.
He
feels,
since
you
guys
are
both
here.
I
just
wanted
to
do
a
little
check.
Do
you
want
to
quickly
discuss
the
CBD
Jenkins
integration
stuff
that
we
talked
about
earlier
this
week
and
what
the
next
steps,
but
the
next
steps
are
for
that
when
you
do
scheduling
meeting,
hopefully
we
can
get.
We.
F
A
A
A
A
There
there
is
one
blocker
that
we
have
right
now
that
we
have,
we
should
maybe
talk.
We
could
talk
quickly
about
here.
Centos
eight
is
not
out
yet,
and
we
were
under
I
was
under
the
impression.
That's
what
we're
gonna
install
on
these
things
keep
watching
me,
but
what
does
cruising
need
something
more
easy.
C
Well,
we
I
think.
What
do
we
need
is
that
whatever
the
mainstream
in
the
next
year
or
1/2
years,
real,
it
is
the
future,
and
for
that
we
can
wait
like
half
have
a
month.
I
think
it
should
be
ready
by
the
end
of
this
month.
Right
if
everything
goes,
fine
I
will
chicken
gotten
what
pages
they
are
moving
very.
A
A
A
C
A
C
A
C
Another
interesting
thing
might
might
need
to
I
want
to
raise
these
that
I
have
a
pair
to
replace
in
mutex
when
the
safe
mutex,
because
the
difference
is
that
in
a
knock-back
mode
to
the
signal
tax
season,
it's
just
an
alias
of
mitigating
stand,
elaborate.
So
no!
No
over
habit
for
for
the
but
tracking,
the
Shalini
dependency
or
something
oh,
the
batteries.
D
A
Have
you
have
you
done
any
tests
to
see
if
that
changes
anything
I.
C
B
C
C
Ideally,
we
need
we
need
to
break
it
into
smaller
pieces
and
Yap
I
component.
But
the
pen
point
is
that
the
condition
the
Steve
condition
I
replace
the
parameter
type
from
mu
tech,
capital
minute
X
with
us.
It's
me
tax
and
we
use
safe
condition
everywhere.
So
if
we
want
to
replace
the
the
mutex
with
with
simply
text,
we
need
to
do
in
the
old
one,
because
all
those
place
we
use
with
the
connection
and
the
primary
yes
to
switch
at
a
time,
so
I
prefer
do
it.