►
From YouTube: Ceph Crimson/SeaStore 2021-05-26
Description
B
B
Wait
what
no
these
three
pieces
of
code
were
from
the
x
out
of
change?
I
have
no
idea
why
it
has
an
effect.
Really.
C
A
B
Emailed
you
the
commit
that
reverts
them
before
that
just
comments.
Thank
you.
Thank
you.
There's
they
appear
to
be
perfectly
fine.
I
don't
really
understand
why
it's
causing
trouble
for
clan,
I'm
just
looking
into
it.
I
expect.
C
All
right
everyone,
let's
start
last
week,
I've
been
reading
materials
on
gaps,
collection
on
log
structure
file
system
and
to
be
specific
on
on
on
the
implementation
of
f2
fs,
and
also
try
try
to
hold
m
clock
to
assist
c
star,
because
I
believe
that
sister
would
also
benefit
from
from
qs
to
balance
the
to
have
a
better,
very
control
of
the
foreground
task
and
background
tasks
like
the
request
from
client
and
request
like
capture,
collection,
etc,
and
still
I'm
still
working
on
it.
C
Into
I'm
police
putting
it
into
two
sisters,
but
I
think
the
the
first
I
think
will
there
are
two
major
consumers
of
this
system:
implementation
of
the
name
clock
one
is
the
the
the
pg
and
osd,
namely
crimson,
and
the
other
part
is
the
systole.
I
think
both
of
them
will
be
using
m
clock.
C
Yeah
I
saw
that
cool,
okay.
The
reason
I
I
why
I
started
looking
into
it
is
that
sister
started
to
to
sorry
sister
started
to
complain
that
it's,
it
observes
some
some
store
store
threads
because
we
do
start
some
some
physics
right
in
m,
clock
and
use.
We
we
we're
actually
using
a
mutex
sys
physics,
mutex
writing
m
clock.
So
that's
why?
B
I
just
use
something
simple
or
we
can
port
it.
That
would
be
fine
too,
but.
C
B
A
C
C
You
mean
how
to
wire
to
a
threading
model
exactly.
C
B
C
Though
yeah
to
expose
this
kind
of
machinery
to
the
consumer
of
this
library,
yeah,
that
part
could
be
controversial,
and
I
I
really
need
to
to
talk
to
to
eric
on
this
to
get
more
inputs
from
from
inside
and
actually
anyway.
That's
me,
I
think
the
first
step
is
to
to
forget
about
the
posix
threading
and
to
to
to
implement
right
with
the
three
star
and
later
on
move
over
to
to
worry
about
how
our
posix,
how
we
can
adapt
it
to
posix.
That's
my
approach.
A
B
B
A
B
Yeah,
I
I
use
pull
specifically
because
it
doesn't
create
a
threat,
so
I
think
I
think
it's
using
standard.
I
think
it's
using
pthread
mutexes
internally
to
protect
data
structures
on
the
assumption
that
different
threads
will
be
accessing
it,
which
isn't
true
for
crimson.
So
we
may
only
just
need
to
if
death
out
the
mutaxes.
C
Not
that
simple,
I
think
some
the
threat
keep
pulling
the
the
the
request
from
from
from
pile
and
if
it
cannot,
it
will
just
wait
for
the
for
the
returned
time
time
period
and
wake
it
up,
return
and
pull
again.
It
again.
B
A
A
C
C
D
D
Another
one
is
a
debugging
on
the
system,
startup
crash
and
I
found
that
when
used,
we
started
start.
C
Probably
later
greg.
C
Okay,
ronan
hi
er.
C
B
I'm
still
working
on
converting
c
story's,
interruptible
future
quite
a
ways
to
go.
I
think
part
of
that
was
trying
to
get
clang
working
so
that
I
could
get
better
error
messages
on
the
amount
of
programming
meta
programming
I'm
doing
with
interruptible
future.
So
I've
got
that
at
least
close
enough
working
now
my
purposes.
That's
it.
D
I
found
the
after
the
first
startup
of
crimson.
I
see
to
do
the
maker
pass
and
after
that
the
ask
is
shut
down
and
start
again
after
the
occurrence
of
sd
start
again.
It
seems
that
rba
albutry
root
empathy.
So
when
it
reads
the
root,
it
cannot
get
anything
and
so
that
the
real
extent
goes
to
a
get
a
dirt.
B
Quite
possibly
I
mean
there
are
a
ton
of
crashes
on
startup
right
now.
I
haven't
seen
that
specific
one.
If
you
want
to
track
it
down
and
create
a
patch,
though
that'd
be
cool.
D
C
Regarding
the
your
decision
on
keep
keeping
using
perf
counter
in
crimson
osd,
I'd
like
to
know
which
part
we
wish.
Oh,
we
should.
D
C
D
C
D
C
C
B
B
I
think
that
trying
to
if
deaf
out
the
perf
counters
and
the
classic
oc
components
would
be
a
ridiculous
amount
of
work.
That's
probably
not
a
good
idea,
however.
In
crimson
you
can
use
c
star
metrics
to
expose
the
perv
counter
state
as
a
c
star
metric
easy
right.
So
I
think
that's
what
you
should
do.
D
Yeah,
so
I
might
understand
right
so
in
the
chromosome
in
the
easy
style
we
can
have
the
programmatic
matrix.
D
D
D
Yes,
so
c
style
is
our
own
files,
no
classical
files.
So
right,
so
we
can
use
metrics
easily,
but
for
chromosomes
did
it
compare
the
coupled
with
the
classical.
D
You
you
just,
we
can
make
a
wrapper
and
the
classical
osd
file
called
the
proof
counter.
Some
function.
A
B
D
C
B
C
B
C
C
A
Last
week
I
was
trying
to
make
the
extent
placement
manager
work
and
right
now,
I'm
trying
to
implement
the
parallel
for
each
function
for
the
aerator,
which
will
be
used
in
extent,
placement
manager.
That's
all
for
me.
C
A
A
Yeah
currently,
the
ono
tree
should
be
able
to
work
with
you
again
now
and
the
api
is
merged.
A
A
A
B
B
A
Yeah
yeah
I
just
adjust
to
the
node
size
according
to
the
hard
limit.
B
A
You
mean
to
make
oh
no
size
flexible,
so
it's
starting
exactly.
B
B
A
E
No
worries,
okay,
the
production
backhand.
I
think
I've
implemented
the
missing
inject
data
and
m
data
errors.
E
E
Extended
the
coverage
for
the
unit
test
coverage
for
aerator.
Now
we
are
verifying
that
the
futurization
works
for
aerated
futures,
as
well
as
it
works
for
vanilla
ones.
Apart
from
that,
I
fixed
a
problem
with
obsequences.
E
There
was
assertion
because
there
was
a
failure
because
of
we
assigned
the
brief
of
id
in
some
very
rare
cases
twice
during
the
execution
of
of
of
client
request,
and
yesterday
I
sent
a
fix
to
x
in
the
lifetime
of
executor
to
to
match
to
free
the
object
after
completing
the
all
completed
sub
future
of
duo
is
the
ops.
That's
me.
E
That's
about
status,
but
we
discussed,
and
I
think
we
wanted
to
to
get
back
to
the
to
the
replication
of
tc
malloc
properties
and
get
hit
properly.
E
E
One
second,
let
me
find
the
name.
Okay,
we
wanted
to.
We
wanted
to
differentiate
the
qa
suit,
depending
on
variant
of
osd
that
is
being
tested
and
just
wanted
to
to
get
back
to
the
discussion
and
ask
what
would
the
status
have,
which
is
the
solution.
C
C
A
B
A
A
A
B
B
A
A
B
In
classic
and
later
on,
we
can
either
generalize
the
interface
to
work
for
whatever
memory
manager
we
have
in
crimson,
or
we
can
change
the
test
suite
to
I
forget
it:
that's
a
better
solution
than
removing
the
testing.
C
How
about
this
we
could.
We
could
create
a
command
which
can
only
be
served
by
by
chromosomes
d,
and
whenever
we
have
we
test
a
a
command,
which
can
only
be
several
of
us
served
by
classical
d.
We
can
do
this
like
something
like
this.
I'm
typing
the
chart
window
class.
A
A
B
A
B
E
B
B
Have
to
I,
like
the
the
the
the
driver
for
this,
that's
what
I
was
going
for
is
as
soon
as
we
have
an
administrative
or
debug
or
introspection
interface
that
we
need
to
use
in
tests.
That's
crimson
only
that's
when
we're
going
to
want
to
fix
it
yeah,
so
that'll
that'll
be
the
point
at
which
to
do
something
about
it.
Until
then,
it's
not
very
important.
E
C
B
B
I
wanted
to
talk
to
you
a
little
bit
about
this
vm
clock
thing
if
you've
got
a
minute
sure,
okay,
well,
first,
before
I
monopolize,
does
anyone
else
have
no
discussions
going
once.
B
A
B
B
There's
no
extra
threat
here,
pull
request,
gets
pulled,
gets
called
directly
from
the
the
from
the
c
star
context.
B
B
One
of
them
is
used
by
rgw
and
it
does
create
a
work
queue
that
pulls
things
off
of
the
queue
and
actually
runs
them
in
line
in
its
own
thread
pool,
but
the
other
one
is
used
by
the
osd,
the
classic
osd
that
is,
and
also
crimson
osd,
and
it
does
not
create
its
own
thread.
Pool
the
osd's
own
work
queue,
calls
into
it
and
interprets
the
results.
It
gets
back,
crimson
without
being
private
right,
you're
waiting
for
it.
It
shouldn't
block,
except
that
in
classic.
Obviously
it
can
because
it's
called
by
many
threads.
B
It's
used
as
a
crossbar
between
the
messenger
async
threads
like
the
the
messenger,
the
asic
messenger's
worker
threads
and
the
osd's
worker
op
work.
You
threats,
you
may
recall
those
are
separate
pools,
so
the
messenger
inputs
things
into
this
queue
and
the
osd
takes
things
off
of
the
queue.
So
there's
a
mutex
protecting
those
data
structures
and
that
can
block
in
crimson.
We
use
the
mutex
because
I
didn't
there's
no,
if
defining
it
out,
but
it
should
never
block
because
there's
only
one
thread
accessing
it.
B
So
my
question
is:
what
did
you
see
blocking
because
I'm
only
aware
of
one
thing
that
tends
to
cause
c-star
to
complain,
which
is
that
when
c-store
starts
up
it
does
a
full
metadata
scan.
C
I
think
it's
more
related
to
the
related
to
the
scheduler
ocd
scheduler.
B
If
this
is
working
correctly,
then
this
data
structure
should
be
simply
a
data
structure.
It
shouldn't
even
be
starting
a
thread.
So
if
it.
A
B
C
Even
we
have
a
log
writing
the
implementation
within
game
clock
that
that
very
lock
should
not
be
noticed
by
the
installer
detector
right,
not.
B
Unless,
for
some
reason,
like
the
stall
detector
is
just
detecting
how
long
it
takes
the
task
to
check
back
in
with
the
reactor
yeah
right,
so
it's
not
going
to
detect
the
lock
it
might
detect
the
time
required
to
take
the
lock
if
the
top
lock
is
for
some
reason,
contended.
C
B
B
B
B
This
shouldn't
be
happening,
so
it
would
be
worth
getting
some
kind
of
dump
as
to
what
it's
doing
or
I
could
be
wrong
about
the
way
dm
clock
works,
which
is
possible,
but
as
far
as
I
understand,
pull
requests
request
or
the
pull
queue
is
specifically
not
supposed
to
create
any
kind
of
threading.
So
that
would
be
the
thing
to
work
to
look
at.
I
think,
just
to
make
sure
we
understand
what
it's
doing.
If
you
can
give
me
the
like
the
output
that
might
be
good.
B
Yeah
that
like
so,
if
you
can't
like
imagine
every
passing
that
comes,
is.
A
B
Like
if
there's
always
work
to
do,
but
it
never
has-
I
o
to
do
because
everything's
in
cash
or
whatever
or
because
blue
store
is
running
so
fast
that
it's
always
available
immediately
anyway.
That's
what
I
would
look
at,
I
would
look
at.
Basically,
I
would
try
to
first
figure
out
what
it's,
what
it's
doing.
C
Yeah
yeah
to
produce-
and
I
would
share
you
with
you-
the
directories
and
see
what's
going
on.