►
From YouTube: 2019-05-01 :: Ceph Developer 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
We
talked
about
the
testing
for
the
multi
data
center,
various
stuff,
anything
to
talk
about
there
do
we
know
what
the
next
step
is.
I.
B
B
If
we
want
to
talk
about
that,
don't
have
a
good
plan
on
is
how
to
actually
or
is
like
what
kind
of
memory
to
apply
and
a
memory
to
the
monitors
for
failures
or
not
failures,
but.
B
B
No
way
before
that,
just
for
the
election,
just
for
the
election
itself,
I
mean
it
turns
out
that
it's
act
like
we
actually
have
to
build
all
that
we
don't
when
there's
a
failure.
We
just
set
a
time
out.
We
have
don't
have
any
idea
which
my
thinking
no
building
in
those
interfaces
but
I'm
trying
to
decide
if
the
right
things
it's
like.
Do.
We
want
to
try
and
like
age
them
out
with
some
exponential
thing
or
just
like
have
a
linear
count
of
the
last
n
failures
are
these?
Are
the
faults
of
these
monitors?
B
A
A
C
A
Their
election
messages
they
can
just
say
this-
is
the
set
of
monitors
that
I
know.
I
can
connect
to
as
of
the
last
few
seconds.
So
when
they're
I
think
in
the
election
code,
then,
when
somebody's
gonna,
when
they
either
defer,
declare
victory
or
whatever
they
cannot
defer
to
somebody
who
has
like
a
smaller
quorum
than
a
set
of
kind
of
connected
peers
than
they
do.
B
B
So
I
think
we
want
to
remember
and
say:
okay
like
like,
if
you'd
like,
if
a
peon
times
out
the
monitor,
then
it
can
say
I'm
blaming
the
leader
because
he
did
not
renew
my
lease
quickly
enough
or
the
leader
can
say:
hey
I
didn't
accept
or
collect
that
quickly
enough
from
this
set
of
peons
and
incorporate
that
into
a
deferral
strategy.
Yeah.
A
A
C
C
An
interesting
matrix
property
that
shows
up
if
you
represent
the
sort
of
suspicion,
values
yeah,
mostly
because
I'm
wondering
what
happens
if
the
monitors
have
different
opinions
about
who
they
should
be
deferring
to
that'd
be
nice.
If
they
had
this
property,
where,
depending
on
what
partial
information
they
have,
they
tend
to
prefer
the
same
appear
if
the
same
period
has
connections
to
all
of
them.
That
would
be
a
nice
stability,
property.
A
B
I
think
what
I'm
gonna
want
it
to
be
deterministic
based
on
widely
shared
data
like
they
do,
the
all
all
probing
before
the
actual
lesson
across
the
start,
so
I
think
they
just
like
so
I
think
that
means
that
they
can
get
the
data
from
everyone.
They
talk
you
and
the
probing
and
share
it
with
everyone
they're
talking
during
an
election
for
the
process
thing,
so
they
can
actually
can
work
on
the
same
data
set
for.
A
A
Actually
don't
so,
we
could
actually
have
something
that
isn't
like
it
isn't
really
an
N
squared
problem.
It's
just
that
when
you're
deciding
who
to
defer
or
not
defer,
monitor,
you
can
say,
I
can
talk
to
these
three
peers.
Monitor
V,
says:
I
can
only
talk
to
two
and
wonder
one
says:
I
can
talk
to
three,
but
you
just
don't
defer
to
somebody
who
can
talk
to
fewer
people
than
you
do
yeah,
and
that
would
so
they
could
form
a
core
and
that
isn't
actually
fully
connected
and
I
would
actually
function.
Just
fine.
B
Yeah,
that's
plausible.
It's
also
not
necessarily
a
thing
that
we
want
to
do,
though,
because
well
I
mean
like
the
most
likely
outcome
of
that.
Is
that
all
the
monitors
is
that
if,
in
our
like
two
data
centers
in
a
the
third
tiebreaker
Monde
scenario,
then
the
tiebreaker
monitor
becomes
the
leader
and
they
probably
have
the
look
the
highest
latency
out
of
all
of
them,
which
is
undesirable
yeah,
but.
B
A
When
you
otherwise
wouldn't
right
and
I
mean
this,
is
this
is
some
sort
of
independent
of
whatever
happens
after
that,
like
they
form
a
quorum,
but
they
also
share
information
about
connectivity
and
they
realize
that
there's
a
huge
discontinuity
treatment,
two
days,
two
senators
if
they
can't
talk
to
each
other-
and
so
maybe
it
like,
doesn't
allow
something
else
to
happen
or
whatever,
like
all
that
stuff
yeah
the
next
step,
but
it
seems
like
all
things
being
equal,
it
might
be
better
to
have
a
quorum
than
not
have
a
quorum.
It's.
C
C
B
B
A
B
A
Like
whether
or
not
say
this
say
there,
you
have
that
split
and
the
people
in
V
aren't
allowed
to
connect
to
the
B
monitor
they
have
to
come.
Talk
to
the
C,
monitor,
I,
don't
know
if
that's
like
strictly
any
better
or
worse
than
what
you
would
have
if
you
actually
had
like
you
still
to
solve
all
the
other
problems.
A
B
A
C
A
I
think
I
think
I
think
we
were
over
simplifying
the
problem
before
I
think
you're
right,
but
but
given
that
I
have
this
feeling
that
we're
gonna
have
to
address
those
issues
anyway,.
C
A
A
Because
every
every
monitor
basically
has
a
score.
That's
how
many,
how
many
people
can
you
tell
many
other
monitors?
Can
you
talk
to
you
and
you
want
to
pick
the
highest
one
that
has
the
lowest
rank
or
whatever
that
whatever
the
criteria
is
the.
C
C
B
I'm,
actually
not
sure
I'm
comfortable
with
that,
though
yeah
well
I
mean
so.
Yes,
it's
nice
that
more
copy
to
the
monitor
data,
but
if
only
like
I
like
if
we
get
into
one
of
these
connection,
breaks
I'd
rather
have
a
local,
a
local
monitor
in
one
of
the
databases,
be
the
leader,
not
the
not
the
cloud
it'll
have
faster
commit.
Latency
is
a
lot
faster
access,
latency
for
the
clients
and
I
think
it'll
be
easier
to
communicate
that
the
state
of
the
system
to
administrators.
That
way.
B
Also,
that's
a
lot
easier
to
generalize
to
okay.
We
think
that
want
to
eventually
generalize
to
you.
Okay,
we
think
monitor,
C
is
or
not
you
know,
whatever
monitor,
database
or
datacenter
one
dot
B
month.
Monitor
is
like
unreliable,
so
we're
going
to
exclude
just
that
one,
because
it's
been
breaking
leases
or
something.
C
B
A
B
C
C
A
If
you
really
didn't
want
the
remote
just
like
you,
could
you
could
devise
a
simple
scoring
function,
though,
like
we
were
saying
that,
like
the
number
of
monitors
that
you
can
connect
to,
is
your
score
during
the
election
and
you
pick
the
highest
one,
but
you
can't
find
a
function
like
you
could
have
monitors
that
are
like
twice
as
preferable
as
the
other
ones,
and
so
you
just
like
scale
them
up
or
down
or
whatever
it
is,
and
each
monitor
could
just
like
say
my
score
is
7-9
and
you
yeah
it's
an
all
all
share
whatever
during
the
election,
and
so
you
just
defer
to
the
one
that
has
claims
a
higher
score
than
they
do,
and
so
you
could,
if
you
wanted
to,
you,
could
just
like
pretty
easily
say
you
prefer
not
and
depending
how
you
set
the
threshold.
A
You
could
make
it
so
that
yeah,
you
prefer
a
quorum
of
two
with
a
we're
in
a
local
data
center
versus
a
quorum
of
three
with
that
or
you
could
scale
them
based
on
like
how
way
key
their
links
are,
or
something
like
that
I'm
or
how
how
much
load
they
have
on
their
sure
I.
Don't
know
like
you
could
do
whatever
you
want.
There.
B
B
B
C
A
A
Think
that
the
first
step,
though,
is,
is
just
add
that
pathology
tasks,
the
enable/disable
links
blockages
or
whatever,
between
nodes
right
Sofia,
that
if
you
have
a
geology
test
with
three
nodes,
each
representing
the
data
center
man
that'll,
basically
shut
down
connections
between
B
and
C
or
whatever.
It
is.
A
A
A
Actually,
no,
we
fixed
that,
though
it
used
to
happen,
is
that
there
was
a
defined
sort
order
in
the
monitor
map.
So
if
you
had
a
set
of
monitors,
they
would
always
sort
by
their
IP
address,
and
that
was
effectively
Randa
based
on
like
what
the
technology
machine
allocation
got
you,
and
so
you
know
which
machines
got
money
and
B
and
C
or
whatever
would
be
sort
of
affected
by
what
IP
address
was
on
that
machine.
But
that's
not
the
case
anymore.
You
can
adjust
the
rank
order
in
the
monthly
put
it.
A
A
A
If
we
have
ABC
whatever
we'll
probably
want
to
explicitly
like
specify
the
rink
and
probably
programmatically
vary
it,
because
sometimes
we
want
Monet
to
be
in
data
center,
a
and
Mon
P
to
be
in
the
cloud
a
month
C
to
be
in
datacenter
B,
but
sometimes
you
want
them
get
away
around
whatever,
or
sometimes
you
want
them
to
be
listed
first
or
second,
probably
I.
Don't
know,
maybe
that
well
to
pay
attention
to
that.
A
A
Cluster
and
block
communication
between
like
two
of
the
three
machines
yeah
and
just
have
the
leader
Mon,
be
on
in
different,
have
them
assigned
in
whatever
the
six
different
combination
or
three
combination,
orders
that
you
can
have
and
just
see
which
one's
work
and
which
ones
don't
because
I
suspect
in
some
of
the
cases
it'll
just
work,
because
the
leader
happens
to
be
the
one
that
can
talk
to
all
of
them
like
the
lowest-ranked
one,
I
didn't
the
other
two.
A
It
won't
figured
I've
like
at
a
facet
that
is
the
Mon
Mon
to
blockage
assignment
or
the
other
way
around,
for
you
block,
aviary,
block,
BC
or
block
AC,
and
then
just
see
what
effect
it
is
just
a
touch
the
baseline
and
then
we
can
start
experimenting
with
ways
to
address
it.
A
A
Or
maybe
not
you
know,
I
mean
it
might
be
that.
So
if
we
went
with
a
strategy
where
the
monitor
that
has
the
best
cannot
connectivity
with
the
most
peers
wins,
then
it
can.
We
can
just
basically
fabricate
a
function
where
you
know
you
get
one
point
for
every
one
that
you
can
talk
to
you,
but
if
it's
high
latency,
then
it's
a
little
less
than
one
or
whatever,
and
so
you
can
like.
We
can
scale
it.
B
B
A
B
B
A
A
B
Yeah
the
Brad's
asking,
if
we
need
a
real
monitor
for
this
or
not,
and
yes,
we
need
to
store
the
same
data
Brad
because
it
needs
to
count
as
one
of
the
nodes
that
commits
data.
If
we
have
five
monitors
and
it's
one
of
the
five
and
it's
like
doing
a
tiebreaker
than
we
need
to
maintain
our
or
more
than
a
half
patch
those
guarantees.
Yeah.
A
C
A
C
What
are
the
requirements
for
the
decision
about
who
defers
to
whom
for
it
to
be
stable?
Is
there
like
a
concise
way
to
state
that,
because,
as
I
understand
it
now,
there's
a
static
preference
or
right
yeah,
it
just
prefers
lower
ranked
monitors.
I
thought
it
is
so
that
has
the
benefit
of
being
widely
known
and
not
arbitrary.
In
that
any
two
monitors
deciding
between
the
two
sets
monitors
will
cover
the
same
decision.
What
are
the
consequences
of
changing
that
to
be
based
on
dynamic
information
that
might
be
different
on
different
monitors
and.
A
I'd
have
to
look
at
the
election
code.
I
mean
they
like.
The
elector
is
like
super
simple.
If
I
remember,
everybody
like
broadcasts
their
their
participation
and
then,
if
they
see
somebody
that's
better
than
them,
they
send
a
defer
message
to
them.
It
says
I
defer
to
you
and
if
somebody
gets
enough
to
first
and
they
spend
they
broadcast
a
victory
to
everyone
else
and
that's
basically
it
which.
B
I
think
means
that
you
can
you
basically
defer
to
everyone.
Who's
got
a
higher
score
than
you
or
look,
or
in
our
case
the
score
is
rank
so
like.
If
you're
ranked
3,
you
will
defer
the
ranks
to
1
and
0
and
they
all
get
those
messages
yeah.
So
they
actually
do
need
to
agree.
You
won't
be
able
to
commit
anything
successfully
that
gets
lost,
but
if
they
disagree,
you'll
just
continuously
run
elections
because,
like
rank
3
will
say
no
I
was
supposed
to
be
leader.
C
B
A
B
A
B
A
B
They
probe
to
everyone
who
they
think
is
up
or
or
who
was
in
the
previous
round,
I'm
sorry
when
a
time
on
happens,
then
they
then
all
the
modern,
then
the
monitor,
who
called
the
timeout,
which
will
generally
mean
either
the
leader
or
else
everyone,
because
the
leader
timed
out
will
do
bootstrap
or
will
go
into
bootstrap
and
will
probe
and
that
I
think
triggers
probing
triggers.
Although
all
probing,
although
only
to
check
that
yeah.
B
Went
through
this
today
and
I,
don't
think
we
do
that
at
all.
In
words,
it's
always
the
whole
process
which
was
confusing
to
me
because
I
thought
we
just
like
we're.
Gonna
call
start
election
and
I
didn't
see
it
anywhere
yeah
or
no.
There
are
some
there's
some
CLI
commands.
They
will
call
trigger
it.
Don't
we'll
have
a
call,
an
option.
A
B
A
B
B
C
A
B
C
C
B
A
B
A
It
removes
the
requirement
that
everybody
have
the
same
shared
information.
That
is
exactly
the
same,
so
they
have
a
shared
understanding
of
what
scores
are
that's
hard.
But
if
everyone,
if
I
participate
in
the
question
says
this
is
my
score,
then
we
just
have
to
get
enough
information
to
them
ahead
of
time
so
that
they
can
choose
a
good
score
and
since
that
ahead
of
time
includes
round-trip
Kings
with
everyone,
then
that
should
be
an
opportunity
for
everyone
to
tell
anybody
else,
any
information
that
they
already
have
themselves
all.
B
A
B
C
I
get
up
some
saying.
The
right
answer
is
to
solve
that
problem.
Now
to
overload
the
monitor
election
process.
Remember
the
miners
are
just
the
auditors.
It's
just
those
notes,
that's
not
every
like
the
data
center.
It
doesn't
tell
you
if
a
rack
is
out
or
something
that
it
cuts
off.
The
auditor,
for
instance,
I.
C
A
Well,
okay,
so
that
well
either
way
that
the
first
step
is
to
to
build
the
technology.
Hooks
actually
do
this
and
we
can
build
those
and
we
can
build
the
test
and
show
that
two
out
of
three
combinations
fail
or
whatever
it
is
before
we
write,
make
any
changes
to
the
model
action
code
at
all
right.
We
can
do
that
and
in
the
meantime,
we
can
map
out
a
specific
change
to
the
election,
and
we
can
argue
about
that
over
the
next
week
or
whatever.
A
B
A
B
A
A
B
A
Mean
it's
famous
they're,
basically,
two
knobs
that
we
can
adjust.
There's
the
rocks,
TV
cache
size
and
there's
those
do
map
capsized
cache
size.
Those
are
the
two
and
we're
gonna
use
the
whole
priority.
Cache
frame
right
that
marks
it
up
to
like
prioritized,
rocks,
TV
index
caches
and
then
balance
the
OSD
map,
cache
that
read
the
rest
of
the
rocks
to
be
memory.
A
All
right,
oh
I'm,
Brad
I,
did
have
one
thing:
I
wanted
to
check
in
with
you
about
I,
don't
know
if
your
mics
working
or
not,
but
way
back
after
you
did
the
change
to
the
Oh
map
approximation.
We
talked
about
adjusting
it,
though,
that
it
could
hopefully
trend
towards
accurate,
even
after
a
scrub.
So,
for
example,
when
you
insert
a
key,
it
would
add
that
many
bytes
to
the
total
and
when
you
deleted
a
key,
it
would
subtract
its
estimate
of
the
average
key
size.
Something
like
that.