►
From YouTube: Ceph Crimson/SeaStore 2021-06-30
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
So,
let's
start
last
week,
I've
been
reviewing
prs
and
doing
some
clean
cleanup
and
also
try
to
drop
the
the
wrap
ups.
My
pr
to
to
add
the
log
free
cue
to
the
shouted
queue
in
editing,
store
and
also
created
some
clean
up
here,
and
that's
me.
B
Yeah
I
was
working
on
the
sister
side
on
integrating
ioc,
tlf
control
interfaces
on
the
system
remote
set.
Actually
so
the
patch
is
submitted
and
has
some
reviews
submitted,
I
need
to
work
on
the
further
versions.
Yeah,
that's
from
my
side.
Okay,
sorry.
A
Yes,
I
think
iron
spi
is
under
our
view
on
the
sister
discussion
group
yeah
yeah.
Yes,.
D
Last
week
I
reviewed
some
peers
and
they
fixed
the
two
runtime
exception.
D
One
is
the
compare
x
attribute,
one
is
pg
in
rs,
and
so
currently
the
the
debug
builder
and
release
builder
can
also
be
tested
by
fio
and
reader's
bench
and
the
release
build
are
still
meet
the
ad
pin,
airbnc
and
ping
issue,
so
I'm
trying
to
debug
on
later.
Currently
that's
all.
D
The
back
output
and
it
quickly
meeted
the
either
ping
unable
to.
D
C
I
think
c
star
is
far
enough
long
now
that
I'd
like
bugs
filed,
do
you
want
to
file
a
bug
for
it,
along
with
any
relevant
information,
just
anything
for
reproducing
it
that
might
make
it
easier.
C
Cool
that
way,
I
expect
other
people
will
be
hitting
it
as
well.
It'll
make
it
easier
for
us
to
coordinate
efforts.
D
I
just
mentioned
that
when
we
debug
output
it
will
run
slowly
so
it
when
your
test
with
more
time
it
couldn't
meet
this
problem,
but.
B
E
Hello,
I'm
working
on
the
crimson
background.
This
week
I
sent
I
found
the
root
cause
and
sent
a
patch
picks
for
the
issues
in
trashing
test.
There
was
a
lot
of
a
lot
of
core
dumps.
Sorry,
a
lot
of
segments
around
the
pity
backend
load
metadata,
which
turned
out
to
be
caused
by
the
by
the
wrong
casting
in
alienster
here
is
the
pr.
E
Some
improvements
around
mostly
around
around
the
box,
like
dumping,
the
content
of
pratt's
self
info
in
the
support
handler
and
the
asset
movable
ping
that
I
hope
could
be
useful
to
tackle
some
weird
and
surprising
behavior
when
obviously
plus
password.
It
comes
to
moving
moving,
actually
a
const
qualified
variable,
which
might
run
a
bit
different
constructor.
That
somebody
likely
would
would
expect
more
info
in
the
in
the
comment
message
right
now.
I'm
taking
a
look
on
another
problem,
which
is
which
seems
to
be
very
tightly
cup,
which
seems
to
be
about
appearing.
E
One
of
acting
set
members
got
one
of
osd's
got
stuck
in
the
one
rpg's
got
star
on
one
of
the
osds
got
stuck
in
the
started,
slash
spray
state.
When
I
get
the
lease
message
which
goes,
which
is
causing
which
is
causing
transition,
the
crashed
state
and
transaction.
B
B
C
What
is
the
emily's
message?
Do
you
guys
go
on
go
figure
out
what
that
is.
E
And
some
tiny
prs
about,
like,
like
removal
of
some
ancestry,
copying
on
rmd,
get
object,
contact
stuff
that
is
actually
a
side
effect,
the
side
quest
of
the
of
the
load,
the
piggybacking
cloud
metadata
investigation.
That's
me.
A
Just
a
quick
note
on
the
pr142100,
I
encourage
all
the
team
to
take
a
look
at
it
because
sometimes,
if
we
forget
to
to
to
add
the
mutable
specifier
for
a
continuation,
the
the
the
application
still
compiles,
but
we
might
miss
the
opportunity
of
the
minimization,
because
when
we
copy
a
sorry
when
we
capture
a
variable
the
without
the
immutable
specifier,
we
end
up
with
a
immutable
variable
and
when
we
try
to
move
it
away
from
it
in
the
continuation
embedded
in
that
recon
continuation,
we
just
failed.
It
create
a
copy.
A
E
It's
funny
because
many
times
I
I
wanted
to
strip
the
mutable
keyword
down,
because
I
thought
it's
unnecessary,
even
if
the
original
outer
were
the
change
of
of
a
change
where
the
std
new
label
was
highly
recommended
highly
usable,
it's
still
quite
likely
that
somebody
in
the
future,
for
the
sake
of
cleaning
up,
will
take
the
mutable
specifier
out
of
the
of
the
lambda,
and
the
assertion
is
the
way
to
prohibit
it's
actually
a
better
way
than
commenting
to
to
in.
B
A
But
for
the
performance
case,
it's
it's
a
it's
about
to
have
the
mutable.
F
G
Hi
everyone
first
a
just
to
let
you
know
I
started
going
over
cvs
and
when
we
start
starting
interviewing
interns
for
the
team,
we
have
an
opening
and
started
getting
the
cvs.
G
That
is
waiting
both
give
us
your
your
tests,
separated
from.
G
Who
was
that
separated
from
egos
and
comments
or
approvals?
It's
it's
been
there
for
quite
a
lot
of
time,
and
I
think
it's
a
good
idea
to
to
get
it
merged.
I'm
working
mostly
on
debugging
some
issues
with
the
help
of
nia
and
sam.
A
G
C
The
interruptible
future
thing
went
in
and
appears
to
have
only
caused
some
trouble,
so
I
fixed
the
obvious
stuff.
Last
week
I
put
out
a
pr
that
ought
to
fix
the
bug.
Zwehan
was
talking
about
where
the
osd
maps
were
showing
up
and
crashing,
or
the
osd
map
wasn't
present
on
the
osteen
was
crashing
up.
It
was
a
generic.
It
was
in
a
specific
case
of
the
generic
bug
of
c-store,
wasn't
respecting
transaction
ordering.
Yet
I
even
left
myself
a
comment
to
that
effect.
C
I
just
forgot
about
it,
so
I've
fixed
that
now
once
that
pr
merges
each
d
store
collection
has
a
mutex
associated
with
it.
So
any
particular
placement
group
will
have
to
finish
the
prepare
phase
before
it
can
submit
the
next
one,
which
is
fine.
We
can
do
a
very
tiny
bit
better,
but
it's
probably
not
going
to
be
worth
the
complexity,
because
this
is
only
for
pg.
So
if
you
have
100
pgs,
you
can
still
have
100
things,
repairing
at
once.
C
H
Last
week
I
modified
the
exchange
placement
manager
to
address
yinshin
and
sam's
concerns
right
now
they
exploit
the
extent
placement
manager
determines
the
extend
physical
address
at
the
time
of
actually
writing
it
and
the
rewritten
extents
are
persisted
to
the
underlying
disc.
H
Before
the
update
of
the
lba
tree
happens
last
week
I
also
rebased
the
extent
placement
manager
code
against
the
master
and
right
now
there
seems
to
be
multiple
interruption,
condition
leaks
within
the
the
post,
the
post
rebase
version
of
my
code
right
now,
I'm
I'm
trying
to
fix
it,
and
after
that
I
think
I
can
push
the
pr.
F
A
F
F
I
will
I
haven't,
sent
like
five
more
commits
about
the
fix,
and
I
I
think
I
will
send
out
today
and-
and
I
will
work
on
the
following
pr
tool
about
the
extent
placeholder
feature.
I
think
it's
also
related
to
the
leak
of
transaction
validate
and
last
week.
I
I
also
started
to
work
on
improve
improving
matrix
to
systole.
F
C
Yeah
I'm
out
next
week
at
ronin.
That
means
I
won't
be
doing
the
rados
qa.
E
Yes,
I
am,
I
posted,
I
posted
a
link
to
the
gist
with
some
locks.
C
The
idea
is,
the
purpose
of
this
message
is
that
do
you
know
about
the
the
read
hole.
C
E
B
C
Default
there
are
some
mechanisms
to
cancel
it
early
in
conservative
cases
where
we
know
it's
safe,
but
in
general
that
means
that
before
we
can
accept
regional
rights,
we
have
the
primary
has
to
maintain
this
lease
lease
ack
message:
sequence
with
its
replicas
on
a
10
second
basis.
Forget
the
exact
frequency
so
what's
going
on.
Here
is
almost
certainly
the
case
that
either
you're
getting
an
m
lease
message
from
a
prior
interval
or
the
activation
message
and
the
m
lease
message
are
changing
order
and
you're
processing
the
before
the
activation
has
committed.
C
C
Lots
of
messages
have
this
dependency,
so
in
general,
you're
not
allowed
to
change
the
order
of
peering
messages
sent
for
the
primary
to
the
replica
they
have
to
be
processed
in
order
and
all
the
way
from
the
message
handling
stack
through
to
putting
through
the
pg
through
to
putting
it
into
the
event
queue
for
the
peering
state
machine.
All
of
that
has
to
be
preserved.
E
I
see
I
already
I
already
added
the
debug
since
the
just
spawned
new
annual
the
background
on
technology.
E
Yeah
it's
now
it's
the
most
common
commonly
appearing
problem.
I'm
I'm
seeing
in
the
in
the
backhand
trans
cool.
B
E
A
A
C
Guess
it
doesn't
work
as
well
anyway,
for
whatever
reason:
that's
what
it
ended
up
being
called:
it's
not
an
uncommon
property
of
bird
systems.
To
be
honest,
even
in
seth,
it's
pretty
difficult
to
trigger
it
in
real
life.
You'd
have
to
get
mildly
creative,
but
in
very
small
systems.
It's
a
little
bit
possible.
So.
C
C
C
C
C
What
the
actual
requirement
is
that
all
of
the
osds
that
the
client
is
talking
to
are
all
in
the
same
network
bubble
and
none
of
them
can
talk
to
the
outside
world
and
they
all
agree
on
the
osd
map
and
they
all
got
partitioned
at
the
same
time,
because
otherwise
the
client
will
get
newer
osd
maps
from
whichever
the
newest
one
is
and
they'll
start
trying
to
talk
to
the
monitor
which
won't
work.
So
they'll
realize
they're
they're
down
all
right.
C
C
B
C
And
in
a
way
that
doesn't
trigger
any
of
the
other
behaviors,
but
to
be
clear,
this
is
actually
fixed
in
current
stuff
via
the
m
lee's
emily's
act
messages
when
a
prime.
What
a
pg
when
osd
becomes
primary
for
pg
it
obtains
a
promise
that
its
replicas
will
not
let
another
osd
become
well.
What
actually
happens
is,
I
think
they
put
something
into
the
osd
info
that
gets
propagated
to
the
new
primary.
That
says,
don't
accept,
reads
before
acts
unless
you
hear
from
the
original
primary.
C
A
Thank
you,
sam
anything
else.