►
From YouTube: Ceph Crimson 2021-03-31
Description
A
Last
week,
I've
been
occupied
by
the
deal
occupied
by
the
organic
to
focus
migration.
B
Sometimes
I
can
finish
the
testing
and
sometimes
they're
still
there
as
a
segmentation
fault,
but
I
can't
find
any
any
laws
1
and
will
trigger
the
segmentation
fault.
Sometimes
it
repeatedly
happened.
Sometimes
it
will
not
happen.
Even
I
run
the
test
over
one
hour,
so
I
still
try
to
figure
out
how
to
trigger
it.
In
what
condition
I
can
trigger
it
and
try
to
trace
it
and
another
one
is
read
the
sams
segmentation,
clear,
sigma,
clear
code
and
journal
code
and
a
little
issue
back
and
I
will
fix
it
tomorrow.
B
B
So
sometimes
it
can
finish,
maybe
until
repeatedly
many
times,
there's
no
error,
but
sometimes
it's
repeatedly
to
happen.
The
segmentation
fault
and
when
I
set
the
offsite
to
be
a
larger,
for
example,
about
the
50
gigabytes
fl
cannot
run
so
I'm
not
sure
if
it
is
fl
issue,
all
the
other.
B
C
C
I
mean
it's
pretty
clearly
a
timing
bug
like
it's
real,
I
just
don't
have
a
way
to
reproduce
it
right
now.
As
for
the
other
thing,
I
I'd
ask
you
to
identify
why
it's
happening.
B
Yes,
the
offsite,
I
created
many
jobs
and
each
one
set
the
different
offsite
and
if
I
started
the
outside,
a
bigger
size,
number
50
gigabits
and
so
the
fl
stopped
there.
But
no
ever
happened
in
the
in
the
nbd
side,
just
at
a
fl
stop
there.
So.
C
C
B
C
B
C
D
Hello
last
week
or
the
past
week,
I
picked
back
up
the
document
you
mentioned
last
last
week,
kefu
regarding
documenting
recovery.
I'm
still
working
on
that.
There
are
a
few
questions
that
sam
helped
me
answer,
so
hopefully
I'll
be
done
with
it
by
next.
My
next
meeting,
and
also
I
synced
up
with
the.
D
The
discussion
on
on
refactoring
with
either
unique
ptr
or
foreign
ptr
connection,
so
I
was
decided
for
now-
is
that
I'll
continue
doing
the
refactor
with
the
unique
btr,
but
keep
in
mind
that
we
might
need
to
change
that
to
foreign
ptr
later
so.
The
way
I'm
implementing
it
now
is
should
be
pretty
easy
to
to
transition
to
to
foreign
ptr,
maybe
with
like
some
sort
of
a
wrapper,
but
it
definitely
wouldn't
require
all
the
all
the
changes
that
I'm
doing
now
to
do
them
again.
F
Hi
everybody
I
was
working
on
introducing
to
classical
sd
the
changes
I
made
to
describe
a
state
machine
so
that
it
would
match
what
I
did
in
queensland.
F
C
Yep,
I'm
working
on
the
logic
for
extent
still.
I
came
up
with
a
scheme
for
embedding
all
of
the
relevant
allocation
information
into
the
lba
layer
and
avoiding
the
extent
map
entirely
so
I'm
working
on
implementing
that
it's
a
little
bit
more
complicated,
but
it
eliminates
an
entire
second
mapping.
So
it's
definitely
worth
it.
C
So
originally
the
idea
was
chad
may
created
a
an
extent
map
implementation
that
we
were
going
to
use
as
a
mapping
from
optic
addresses
to
logical
addresses
within.
C
B
C
At
least
that's
what
I'm
working
on
we'll
see
how
it
works
out.
I
may
still
need
it.
The
main
sticking
point
was
figuring
out
how
to
handle
clones,
but
I
think
I'm
going
to
add
what
is
essentially
an
indirection
into
the
lba
map
itself.
That
is
a
logical
address
inside
of
the
lba
b3
manager
itself.
The
entry,
instead
of
pointing
at
a
physical
address,
might
point
at
another
logical
address:
that's
eliminating
the
need
for
a
second
mapping.
C
G
Last
week
I
implemented
the
the
get
and
set
address
method
in
c
store
right
now,
I'm
trying
to
implement
the
read
and
write
meta
meta
methods
in
systore
and
I'm
also
looking
into
the
marx
problem.
That
might
be
my
call
might
be
caused
by
interoperable
future
and
the
sister
thread.
That's
all
for
me.
A
Just
a
little
more
background,
marco
is
obviously
reserving
some
some
memory
leak
and
he
tried
to
bisect
the
offending
a
commit
and
he
found
that
it's
a
it
was
the
entire
future
who
introduced
these
regressions.
That
is
looking
into
it
and.
G
E
E
For
oh
no
three
parts
last
week
I
fixed
minor
bug
and
to
and
to
implement
all
node
erase
feature.
I
think
I
have
figured
out
a
way
to
manage
tree
level.
Enviro
invariants
across
recursive,
merge
and
split.
A
Never
never
never
complete
unit
test
we
have
been
seeing
over
the
over
the
last
couple
months
is
that's
because
the
jenkins,
because
when
we
push
another
chain
to
appear,
they
can
get
about
the
running
task.
Sometimes
it
left
us
with
some
some
some
running
unit
test
and
they
did
not
get
chance
to
be
to
be
finished
by
by
the
parent
process.
So
I
added
a
step
to
abort
to
to
kill
them.
When
the
when
the
checking
job
is
aborted,
it
seems
to
work.
A
A
D
Yeah
so
so
you
mentioned
we
have
so
we
have
the
the
log
of
the
primary
of
the
replicas
and
we
have
some
sort
of
an
authoritative
log
that.
C
Read
the
raido's
paper
and
we'll
discuss
it
next
next
week,
so
the
the
the
one
of
the
things
that
was
originally
interesting
about
stuff
is
the
process
by
which
we
obtain
an
authoritative
log.
It's
called
peering
and
that
state
machine
and
peering
state.h
is
the
thing.
But
the
best
explanation
of
it
honestly
is
the
original
paper.
C
Okay,
I've
got
a
link
for
you
here
we
or
go
don't
have
to
wait
till
next
week
either.
If
you
want
to
ping
me,
what
time
is
there?
Anyone
by
the
way,
I
actually
have
no
idea?
D
What
time
zone
I
want
I'm
on
israel,
so
it's
utc
plus
yeah,
so
it's
it's
8
15
a.m.
Right
now,
for
me.
D
Yeah
you're
deceiving
us
too
thanks
ronan
yeah,
so
the
second,
the
second
question.
C
D
C
D
The
the
second
question
was
so
in
in
pg
in
in
long
beach
recovery,
so
we're
going
over
the
missings
right
and
we
talked
about
how
well
in
our
urgent
recovery
we
we
it's
force,
we
cover
a
specific
specific
object
and
in
long
base
recovery.
My
question
is:
is
it
possible
that,
while
we're
recovering
from
the
missing
set,
there
will
be
some
sort
of
interruption
or
is
it
granted
that
we
complete
all
of
recovering
all
of
the
objects.
C
C
So
if
there's
something
in
the
unfound
set,
I
forget
what
the
code
does,
but
I
think
it
just
exits
out
of
the
log
base
recovery
process
and
puts
itself
in
you
know,
state
and
found
or
whatever.
A
D
C
C
I
can
quickly
sketch
the
peering
thing
just
to
give
you
some
structure
is
anyone
else
interested
in
yeah?
Well,
maybe
so
at
a
super
duper
high
level.
The
do
you
have
a
sense
of
what
I
mean
by
authoritative
log,
that's
sort
of
the.
D
C
Most
recent
updated
is
where
we
get
the
trouble,
so
as
an
illustrating
example.
Let's
say
they
call
there's
exactly
one
client
and
exactly
one
placement
group.
The
client
has
submitted
10
ios
on
the
same
object
because
of
the
way
liberators
works.
These
writes
that
the
rights
that
it's
submitted
will
all
occur
in
order
and
it
will
increment
that
object's
version
by
one
h.
C
What
are
the
valid
versions
of
the
authoritative
log
that
could
result
or
what
would
what
would
make
an
authoritative
logo
wrong?.
D
Well,
if
none
of
them
committed,
then,
if
we,
if
we,
if
we,
if
we
cancel
all
the
ios,
then
none
of
them,
oh.
C
The
the
osd
yes,
but
when
the
osd
comes
back
up
the
very
second
it
comes
up.
The
first
read
it
it
serves,
might
not
be
from
that
client.
It
might
be
from
a
different,
it
might
be
from
a
different
client.
So
what
possible
states
of
the
pg
are
valid
or
to
put
it
another
way?
What
do
we
need
to
do
to
not
break
the
rules?
D
Rules
we
can
serve
them
again
and
then
grants
that
you
know
we
will
make
sure
that
that
those
rights
are
committed
right.
C
But
we
don't
so
for
one
thing
they
may
not
have
committed.
We
turned
the
osd
off,
we
don't
know
so
from
our
point
of
from
our
point
of
view,
from
the
client's
point
of
view
or
setting
up
this
scenario,
it's
possible
that
all
10
of
the
ios
never
got
to
the
osd
they're
in
a
router
somewhere
or
a
switch
somewhere
in
transit
that
the
packets
themselves
literally
never
got
there
or
it's
possible
that
all
ten
committed
and
the
replies
didn't
get
back
to
the
client
or
any
intermediate
state
between
those
two.
C
C
So
the
answer
is
that,
from
as
long
as
the
client
never
saw
any
of
the
responses
it
is,
it
would
be
valid
for
the
log
to
contain
none
of
those
rights.
It
would
be
valid
for
it
to
gain
all
of
those
rights.
It
would
be
valid
for
it
to
contain
the
first
five,
the
first
six
or
the
first
seven,
but
it
would
not
be
valid
for
it
to
have
rights
two
and
six,
but
none
of
the
others,
because
they're
submitted
in
order
right
yep.
C
C
So
in
general,
the
way
we
do
this
is,
let's
make
a
few
simplifying
assumptions
about
the
way
osd
maps
work.
For
one
thing,
let's
assume
we
have
all
of
the
osd
maps
back
to
the
beginning
of
time.
C
So
when
let's
say
we
receive
a
new
osd
map
that
says
iosd10
and
now
the
primary
for
pg
1.12,
what
I'm
going
to
do
is
I'm
going
to
go
back
through
the
set
of
osd
maps
back
to
the
beginning
of
time
and
find
every
acting
set.
Osd
member
that
has
ever
been
in
this
set.
I
am
then
going
to
query
all
of
them
as
long
as
I
receive
a
log
back
from
at
least
one
acting
set
member
of
every
interval,
that
is
every
contiguous
sequence
of
epochs,
where
the
acting
set
was
this
the
same.