►
From YouTube: Ceph Crimson/SeaStor 2020-06-17
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
First,
one
was
about
non,
not
initializing,
the
memory
allocator
pools
and
in
our
in
our
threads
mr.
alligator
has
two-stage
initialization.
It
means
it
needs
some
assistance
from
the
thread
when
willing
to
use
to
use
it
to
get
to
overcome
the
32
max
in
limitation
level.
Second
thing
is
a
bit
more:
it's
much
more
problematic.
Actually,
it
turned
out
that
what
that
our
via
did
our
assumption
that
Bluestar
the
Troggs
DB,
has
manages
its
threads
in
static.
Our
way
using
thread
pools
is
violated.
It
has
some
consequences.
B
I'm
still
working
at
replay
I
noticed
a
better
way
of
structuring
the
way
the
root
node
works.
I'm
fixing
that
up
today
previously
I
had
a
magic
Delta
that
set
up
that
that
changed
the
location
of
the
root
node,
but
the
entire
root
node
is
all
of
16
bytes.
So
now
the
entire
root
node
is
in
the
magic
Delta,
but
I
am
going
through
and
cleaning
up
all
the
places
where
I
did
it
the
other
way.
This
will
be
cleaner
in
the
long
run
there
won't.
B
The
only
things
in
the
root
node
will
ever
be
like
three
physical
addresses,
there's
never
gonna
be
much
there,
that
is,
it'll,
have
a
pointer
to
the
lbh
reroute,
a
pointer
to
any
other
like
alligator
roots,
and
that's
a
it's
not
going
to
have
anything
in
it.
So
this
is
simpler
and
it
avoids
having
to
write
down
an
extra
pretty
much
empty
fork,
a
page
every
time
we
need
to
roll
over
it.
The
journal
so.
D
B
Time
we
roll
the
journal.
Well,
the
first
thing
we
write
down
will
be
the
or
before
we
roll
the
journal.
Rather,
the
first
thing
we'll
write
down
will
be
the
new
magic
Delta
that
contains
the
root
of
the
whole
thing.
If
the
Delta
always
contains
the
entire
node,
you
never
actually
have
to
read
right.
C
B
B
Put
the
whole
root
note
enough
in
the
magic
Nova,
which
means
which
means
the
magic
Delta
instead
of
containing
one
address
now
contains
like
two,
because
that's
all
the
route
that
had
in
the
first
place
the
real
root
node
is
the
LBA
root,
which
is
a
real
b-tree
node
that
really
exists
somewhere.
That's
something
else.
B
B
D
So
I
managed
to
clean
up
the
logic
club
to
a
better
structure,
to
be
reviewable
and
and
also
extended,
is
logic
to
to
attack
stuff
nodes
and
also
did
some
other
necessary
cleanups
to
proceed.
The
implementation
and
for
a
heartbeat
I
did
refactoring
to
to
reduce
proper
abstraction
for
full
review.
That's
me
thank.
C
A
Posted
a
class
English
comment:
I
posted
some
description
about
about
possibilities.
We
have
we
have
in
tackling
this
problem,
but
it's
pretty
it's
still
pretty
problematic
and
I
think
I'm,
not
sure
whether
whether
we
could
okay
some
background,
we've
lost
with
us.
In
my
understanding,
adapting
Veloster
I
have
a
link.
Builder
has
two
purposes:
two
guns:
first,
one,
the
most
fundamental
one
is
to
provide
the
stability
of
entire
crimson
technology.
A
That's
the
primary
goal,
but
there's
also
secondary
1/3,
the
eastern
from
hand
link
in
long
term,
in
other
words,
not
only
testability,
but
production
called,
but
production
as
well
and
I'm,
not
sure
whether
the
productization
is
possible
with
with
a
lien
start
using
balance,
trainer
I.
Think,
because,
if
we
switch,
if
we
switch
the
lip
see
a
locator,
it's
bad.
A
B
B
B
B
A
To
God
we
have
two
goals,
but
addressing
the
second
one.
Production
with
sister
alligator
is
pretty
problematic
because
we,
because
blue
star
actually
and
pain,
can
can
start
during
each
civilization,
but
still
it
can
spawn
us
money
threads
as
user
one
and
the
situation
at
worse
and
actually
got
unveiled
because
of
merging
the
blue
stirring
PR
on
my
2027
at
25th.
A
B
B
D
B
A
B
Not
absolutely
not
cool
yeah
now
right
now,
we're
sharing
all
the
config
stuff,
but
at
some
point
we're
gonna
have
to
figure
out
what
to
do
about
that
and
they're
almost
it
would
surprise.
I
would
be
if
there
were
even
that
many
config
values
that
we
want
to
default
to
this,
to
the
same
thing
like
I,
don't
think
this
is
going
to
be
an
exception.
B
A
Think
this
yeah
okay,
because
well
shrugging
in
Brewster,
drastically
increase
the
demand
for
Oliver
for
Fred's
created
an
installation
program,
let's
say
60
to
280
iPhone,
but
if
we
can,
if
we
can
enforce,
if
you
can
enforce,
users
do
not
use
their
sharding
or
any
other
thing
that
leads
the
huge
number
of
short
living
threats
created
at
utilization,
then
maybe
we
could
use.
Maybe
we
could
use
the
sister
elevator
even
in
production,
who
knows?
Definitely
it
can
be
used
for
testing
it's
even
now.
A
A
Scenarios
and
for
instance,
it
doesn't,
it
doesn't
truly
the
memory
to
underlying
up
to
the
operate
to
the
operating
system.
It's
perfectly
fine.
In
a
situation
when
you
have
static
out
threat
management
like
in
vanilla
sister
like
in
the
let's
say,
perfect
Rox
DB
that
one
that
would
use
only
tactical
thread
pools.
It
wouldn't
be
a
problem,
but
but
even
if
there
is
a
huge
well,
but
if
a
big
number
of
short
living
rats,
we
need
to
be
aware
that
the
memory
that
was
allocated
just
at
the
initialization
stage,
it
won't
be
reclaimed.
A
Actually,
it's
not
a
fundamental
limitation
of
the
allocator.
It's
I
think
it's.
The
situation
is
is
like
that
only
because
of
legacy,
because
because
doctors
future
was
never
and
was
never,
net
was
unnecessary
before
merging
the
alien
stuff.
Insist
on
the
actual
total
that
pretty
easily
and
I
think
the
pups
would
be
would
be
Murderball
with
upstream.
It's
actually
go
when
Winifred
the
dorsum
per
fret
local
structure
control
structures
in
sister
alligator.
A
Now,
when,
when
threat,
when
a
threat
exists
exits,
it
actually
does
nothing
except
setting
exiting
a
one
atomic
al
boolean.
Nothing
part.
There
is
no
call
to
M
advice.
Free
ink
tank
to
parallel
I,
don't
need
those
pages,
but
it
could.
Actually.
We
could
actually
that
feature.
The
question
is
actually
I.
Don't
work,
I,
don't
know
at
the
moment
whether
it's
worth
the
effort
or
not.
It
depends
on
how
much
of
the
memory
would
be
taken
at
the
insurer
is
a
property
I
have
no
idea
how
it
looks
in
in
the
real
scenarios.
A
C
C
We
should
have
a
lower
upper
bound,
then
if
the
sister
is
enabled
when
comparing
crimson
or
steel,
so
another
option
is
to
to
to
catch
the
sister
little
alligator
alone
it
to
reduce
the
the
struggle
structure,
which
is
initialized
when
we
as
a
new
strategy
located-
and
they
said
it
should
be
tax
free
because
in
sista
reactors
to
us
that
not
Exeter
right
practically.
So
it
should
be
no
big
deal
when
and
this
I
don't
think
assistance
an
ADB.
C
A
Danger,
the
initialization
in
sister
and
even
in
nicely
behaving
applications
post
applications
that
don't
manage
threads
dynamically
is
that
it's
actually
the
installation
can
be
a
kernel
itself.
You
don't
need
to
worry,
but
actually
rocks
DB
as
we
as
we
know
from
the
from
the
debugging.
It's
not
entirely,
not
a
nice
to
be
hating,
POSIX
application.
A
A
There
are
also
other
configurable.
We
could
tweak
like
the
number
of
thread
pool
for
the
a
IO.
By
default,
it's
16
we
could
minimize
to
1.
There
is
also
Roxy
beep
configurable
for
stirring
the
number
of
the
dynamic
of
the
dynamically
create
that
threads
for
reading
mem
tables
by
default
it
sixteen.
We
can
my
name
pretty
easily
minimize
it
to
one.
C
C
A
B
How
do
you
date
well?
We're
gonna
have
to
do
both,
but
at
the
first
step
we
definitely
have
to
validate,
because
if
someone
does
something
wrong,
it
shouldn't
crash,
they
should
get
an
error
message
so
that
part's
mandatory.
Think
of
that
later
on,
as
we
get
closer
to
production,
will
change
default.
A
B
C
C
We
come
up
with
a
furry
angel
up
or
upper
bound
of
the,
for
example,
the
threat
number.
We
can
add
that
stating
to
to
to
option
Tata
CC,
because
we
have
that
machinery
in
place
already,
so
we
can
have
a
different
pound
for
T,
star
crimson
and
for
for
classical
D
that'll
be
a
nicer
way
to
to
stove.
It
I
think.
A
A
B
B
So
the
real
question
is:
how
work
would
it
be
to
refactor
blue
store
so
that
we
can
extract
the
threading
and
concurrency
primitives
and
I?
Don't
know
I,
don't
know
what
the
answer
to
that
is
I
feel
like
the
answer.
Can't
be
crazy,
like
it
just
on
some
level
can't
be
thought.
I
could
be
wrong,
particularly
given
that
we're
totally
fine
with
eating
the
whatever
you
call
it
the
thread
API
overhead.
That's
totally
fine.
A
B
A
B
A
B
See
what
you
mean
yes,
except
that
it
only
kind
of
isn't
right,
like
we
have
a
long
way
before
we'd,
be
able
to
productize
it
a
lot
of
auditing
and
testing
so
that
part's
true
either
way.
The
only
advantage
we
have
here
is
that
code
exists,
and
it's
not
clear
to
me
that
that's
really
most
of
the
work.
A
B
B
Right,
like
it
just
knew
so,
let's
not
take
a
really
long
time
there.
It's
the
transactions,
that's
the
problem,
like
writing.
A
transactionally,
correct
object,
store
is
a
lot
of
work
even
doing
it
badly,
like
performance,
isn't
even
the
issue.
It's
just
that
all
of
the
work
required
to
make
sure
your
failure.
Atomic
I
mean
yeah.
It's
it's
work
there
no
two
ways
about
it.
B
Then
we
need
it
to
be,
because
that's
how
technology
works
right,
that's
a
Belen!
What
we
can't
do
away
with,
moreover,
that
testing
wouldn't
actually
test
again,
like
one
of
the
biggest
reasons
with
all
she
does
fill.
Your
testing
is
because
it's
doing
it
on
and
on
disk
format,
customers
will
actually
use
or
users
will
actually
use.
So,
if
we're
doing
that,
testing
on
and
on
disk
format,
that
will
never
see
the
light
of
day
that
really
reduces
the
value
level
down
to
zero.
B
It's
still
valuable,
but
given
how
much
work
that
store
would
take
to
write
I,
don't
know,
that's
not
lobbyists
when
to
me
there
is
a
fourth
option
that
is
worth
considering.
We
can
create
a
Riedel
process
with
a
simple
inner,
IPC,
API
and
I
think
we
should
consider
doing
that.
That's
what
we're
doing
here,
except
for
the
minor
detail
that
it
will
actually
work,
and
it
will
be
simple
about.