►
From YouTube: 2019-11-26 :: Crimson SeaStor OSD Weekly 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
C
Updated
appear
so
I
I
think
when
the
reveal
most
time
I
built
to
the
right
base,
because
every
merger
PR
will
be
influenced
in
effect
that
the
files
become
manifest
is
touched.
So
after
the
most
comments,
it's
updated
and
I
will
to
the
repeat
and
the
hope
it
can
be
murder.
After
the
repeats
and
from
my
side,
the
manager,
as
could
ask
ask
of
me
to
if
the
chromaticity
can't
give
much
better
performance.
So
the
next
step.
C
I
will
do
the
performance
testing
to
give
some
multiple
testing
environment,
to
test
a
report
to
show
to
the
and
the
true
Shah
and
to
make
them
think
that
chrome
is
OST
is,
is
it's
better
than
say
Farsi,
then
they
will
think
it
is
necessary
to
continue
on
the
system.
So
before
we
started
on
the
C
style,
they
want
to
see
the
promised
asking
without.
B
But
I
I
just
want
to
remind
mind
you
that
the
C
story
I
think
it's
his
story
is
more
important
than
than
than
the
overall
austere,
refractory
all
right.
So
those
deep
because
I
think
it's
it's
more
important
because
it
can
can
can
bypass
a
lot
of
overhead
where
we
were
being
suffering
in
in
gusto
and
first
off.
So.
A
We
will
already
have
a
working
crimson
ysd
from
a
B
D
to
disk,
but
with
alien
blue
star.
So
to
count
is
it
possible
to
prove
that
alien
blue
star
has
significant
performance
issue
either
in
it's
ready
model
or
its
support
for
in
that,
in
appropriate
support
for
fast
flash
to
do
have
suggestions
to?
Because
we
need
to
convince
people
that
we
need
to
spend
a
long
time
to
implement,
implement
insist
or
you.
D
A
D
E
E
A
E
A
D
B
A
D
Right
so
that
there
wasn't
enough
to
yell
and
he
observed
a
decrease
in
write
amplification
of
a
factor
before,
but
that's
not
really
an
argument
against
blue
story.
You
can
do
that
with
the
existing
blue
store.
It
didn't
even
require
that
much
change.
No.
The
argument
for
why
we're
doing
C
store
is
because
blue
store
itself
and
rocks
DB
use
up
a
huge
amount
of
the
CPU
of
clásicos.
So
it's
not
unreasonable
to
guess
that
we
should
get
the
same.
I
TPP
up
improvement
on
the
back
end
as
we
got
on
the
front
end.
F
D
A
E
E
Also
in
for
T
start
this,
the
patches,
the
entire
change
that
extent
a
final
revision
and
it
could
accepted
by
rock
at
the
moment
it's
only
well
waiting
for
a
match
right
now,
I'm
working
on
the
watch
on
the
works
and
onwards
stop
operations
of
of
the
watch
out
of
OSD
I
implemented
them
on
the
high
level
right
now,
I
need
to
refactor
the
lab.
The
lower
layer
in
watch
dot
BC
was
the
worst
of
C
to
twerk
latest
work.
E
G
F
D
F
Can
read:
okay,
I,
don't
know
why
I
can't
but
I'm
trying
to
understand
why?
If
you
look
at
this
code,
I
have
a
function
here,
test
right
early,
which
is
music,
I,
think
and
these
using
the
fact
that
I
can
wait
and
then
I
think
in
the
thread,
net
sister
thread
and
then
I'm
hope
I
was
hoping
to.
If
you
look
here
at
straight
below,
I
was
expecting
a
foul
to
be
be
ready
available
after
a
day,
I
think
returns
now.
D
F
D
Happening
that
ASIC
call
passes
the
future
backup
in
the
reactor
or
something
and
that,
but
immediately
returns
a
future.
It
doesn't
necessarily
even
run
that
lambda
them
eventually,
when
that
lab
to
tell
us
get
run
and
that
gets
called
what's
gonna
happen
is
it's
gonna
stack
swap
back
up
to
the
reactor
and
it's
gonna
put
that
task
to
sleep
until
the
sleep
should
logically
have
woken
back
up,
then
it'll
unpassed,
which
back
in
or
stack
switch
back
em
up
here
to
return
from
that
call
and
then
continue
on
with
its
life.
F
D
F
F
D
F
D
F
D
F
F
Okay
yeah:
this
is
more
or
less
what
I
was
trying
to
do
in
factoring
the
code
which
had
returned
to
reality
micro
day
I
was
what
I
was
doing.
I
have
a
working
working
s.
Okay,
I
mean
socket
system
with
now
for
few
weeks
now,
but
in
the
wrote
testing
it
some
unit
tests,
scripts
and
other
heavy
load,
I
get
it.
I
still
see
failures
from
time
to
time
and
I'm
trying
to
root
out
those
failures.
This.
D
A
D
Can
probably
get
it
to
not
complain
since
it
is
test
code,
but
but
the
thing
is,
if
you're,
if
you're
looking
for
concurrency
bugs
with
different
futures
firing
at
different
times,
this
won't
do
that,
because
it's
gonna
suspend
the
whole
reactor
thread.
Scheduler
won't
run,
nothing
else
will
run
it'll
sit
there.
I.
D
One
suggestion
is
that
you
could
introduce
a
thing
where,
let's
say
one
of
the
operations
you
want
do
model
has
a
you
you
you
want
to
make
it
so
that
one
of
your
async
operations
has
a
more
variable
time
cost
than
it
really
does
right.
That's
the
whole
idea,
so
you
could
introduce
a
debug
variable
to
the
library
implementing
your
ASIC
you're,
a
admin
socket
that
actually
just
chains
a
sleep
continuation
on
to
the
end
of
each
one
of
them
with
a
random
wait.
So
the
next
layer
up
wouldn't
see
any
of
that.
B
D
Okay,
so
it
took
forever,
but
I
finally
figured
out
why
my
thing
had
introduced
performance
regression.
It
was
because
I
forgot
that
or
I
failed
to
notice
these
door.
Loggers
will
actually
evaluate
their
arguments
even
if
they're
not
gonna
run
so
I
was
admitting
the
PG
state
string,
whatever
okay,
so
that's
working
now,
so
it's
all
be
ready
for
review
and
probably
merge
I.
Don't
think
I'm
gonna
rerun
the
test
suite,
because
the
part
in
classic
OSD
didn't
change.
So
whenever
you
guys
have
time
to
look
at
it.
That's
fine
I'm.
D
I'm
moving
back
to
recovery,
backfill
one
thing
we
should
start
thinking
about
over
the
next
two
weeks
is:
we
should
come
up
with
a
couple
of
talks.
We
want
to
give
it's
a
block
up.
Cuz.
The
call
for
papers
deadline
is
like
first
week
second
week
of
December,
maybe
so
I'll
make
a
document
insuring
around.
If
anyone
wants
to
claim
a
topic.
A
D
This
is
a
good
type
of
training.
Another
thing
I
did
the
last
couple
of
days
was:
I
was
debugging
a
bug
with
the
messenger
to
implementation.
Generally
speaking,
dropped
connections
are
not
handled
correctly,
so.
A
D
So
continue
of
futures
can
outlive
the
messenger
which
tends
to
cause
seg
faults.
I
found
some
of
the
more
obvious
ways
that
that
can
happen,
but
the
other
thing
that
can
happen
is
that
when
a
connection
dies,
there
can
be
outstanding
futures
waiting
on
the
waiting
for
active
waiting
for
active
gate
or
waiting
for
ready,
okay,
I
forgot.
It
was
something
in
the
key
in
the
connection
protocol
that
will
fail
with
a
broken
promise
exception,
because
the
promise
went
away
right.
D
A
D
D
B
D
A
D
A
A
D
A
A
Characteristics
of
user
space
drivers
that
it
it
is
not
very
convenient
if
we
put
parties
in
separate
processes
like
for,
for
example,
the
the
DVD
K.
It
said
that
this
this
documentation
said
that
there
might
be
some
limitations
if
we
use
multiple
process
mode
and
even
we
might
use
the
we
might
need
to
use
RPC.
That
is
very
lower-level
implementations
that
is
needed
to
transfer
some
some
message
from
one
process
to
another.
A
So
I
think
if
we
continues
to
go
on
with
the
multiple
process
design,
we
might
need
to
have
some
more
complex
implementations
on
that.
So
so
my
question
is,
though,
is
it
better
to
to
go
our
design
waste
with
logical
as
these
or
multiple
processes?
If
we
really
want
to
use
the
user
space
drivers
like
deep,
deep
ASP
decay
or
pinky
man,
maybe.
D
It's
worth
pointing
out
that
the
only
reason
why
it's
annoying
to
run
multiple
by
clásicos
these
in
the
same
process
is
because
there
are
a
bunch
of
process
global
like
variables
for
no
particularly
good
reason.
It
was
just
convenient
at
the
time
we,
as
far
as
they
know,
haven't
made
any
those
choices
yep.
D
Also
point
out
that
I
think
the
DPD
K
concern
is
that
if
you're
trying
to
run
more
than
one
OSD
on
the
same
device,
then
multiple
wait:
okay,
oh
yeah,
yeah
yeah
right
if
you're
ready,
multiple
O's
to
use
on
the
same
deck,
it's
a
problem
if
there
are
different
processes,
so
in
that
sense
yeah
it
might
be
helpful.
Actually.
A
D
D
D
A
I
I
also
looked
at
the
c-store
designed
so
so.
I
have
several
questions.
Maybe
I
will
add
it
later
to
that
document
that
that's
all
for
me,
so
so
I
already
plans
the
zero
success.
Alikom
topic
I,
maybe
have
one
about
the
messenger
progress
and
the
power
our
native
stack
evaluation
about
it
in
us
and
Pratap
ik.
Okay,
that's
all.
E
B
A
B
D
E
Okay,
one
good
question
related
to
watch
that
is
already
implanted.
The
implementation
in
the
current
elimination
in
classical
OSD
I
was
thinking
about
a
case
when
a
single
MOS
D
up
in
instance
carries
multiple
watched.
Stop
operations
like
if
I
even
do
do.
We
have
any
case
any
user,
any
really
user
of
the
radius
protocol
that
will
put
multiple
operations
related
to
work
in
the
device
in
a
single
MOS.
Do
you
could
easily
notify.
A
E
Know
I'm
asking
because
I'm
worried
that
the
in
current
in
classical
OSD
that
we
are
the
up
facts
of
watch
and
and
watch
out
of
order
in
different
order
that
we
got
those
those
messages
in
those
operations
inside
my
job
in
in
such
combination,
such
sequence.
Actually,
we
would
I
am
afraid
we
could
have
a
leakage
of
water
instance,
because
we
are
at
the
from
primary
lock.
Pg
duo
is
the
OP.
Do
I
do
op
effect
actually
disconnects
first
then
goes
to
the
list
of
items
to
connect.
D
E
A
D
We
never
we
don't
even
want
to
replicate
implementations
that
are
awkward
like
if
we
can
make
improvements
on
the
way
it's
laid
out.
We
should
we
should
do
that
just
to
make
it
easier
on
ourselves.
Watch
notify
is
on
its
third
version
of
the
protocol,
because
the
first
two
were
really
buggy.
We
missed
some
things
so
some
of
the
ways
it
works
right
now,
our
historical
artifacts
and
not
actually
good
ideas.
D
A
D
A
D
Almost
for
sure,
we're
gonna
have
okay,
so
just
a
blue
store
uses
rocks
TV
for
everything,
because
it's
there
not
because
it's
necessarily
a
good
idea
and
oh
no
data
and
OMAP
data
had
almost
nothing
to
do
with
each
other.
Oh
no
data
is
known
to
be
because
it's
keyed
by
the
object
cache
it's
known
to
be
fixed
size
and
uniformly
distributed
over
the
space,
because
it's
a
it's
a
hash
right.
By
contrast,
Oh
map
data
is
fully
user
controlled
and
needs
to
be
listed.
D
D
B
B
D
D
Moreover,
the
objects
projects
function
is
namespaces.
You
can
have
duplicate
keys
on
different
objects
right.
The
same
key
can
exist
on
many
objects.
There's
no
reason
to
why
not
blue
store
handles
this
by
prefixing
with
them.
What
the
other
came,
Oh
map
is
logic
Lincoln's
contained
within
an
object.
B
D
D
A
B
D
No
reason
they
would
share
an
interface,
they
aren't
even
keyed
by
this
by
the
same
thing
once
keyed
by
a
string,
the
others
key
by
an
H
object.
What
are
the
points
of
this
is
to
avoid
constantly
stratifying
a
ton
stringer
buying
things
which
rocks
deep,
which
blue
store
does
all
the
time,
because
it
has
to
constantly
turn
in
memory
structures
into
strings,
so
it
can
give
them
to
rocks
TV
and
then
interpret
them
back
out
all
that
stuff
called
cycles.
We
don't
want
to
do
that
if
we
don't
have
to.