►
From YouTube: Ceph Developer Summit Quincy: Orchestrator
Description
00:00 - Dashboard integration: Facilitate initial deployment tasks and Day 2 ops.
28:38 - Cephadm resource scheduling: autotuning osd, mon, mds memory
30:23 - cephadm host draining
43:07 - cephadm-exporter
46:34 - cephadm HA
Full agenda: https://pad.ceph.com/p/cds-quincy
B
C
B
B
B
Like
you
know,
removing
this
rgw,
like
removing
other
w
damage,
can
cause
a
loose
connectivity
and
there
is
like
alert
like
not
enough
remaining
and
nerf
statements,
so
it
is
not
safe
to
stop
more
at
this
time,
because
there
won't
be
enough
mons
to
be
in
quorum
and
stuff
like
that.
So
this
won't
be
possible
and
we
cannot
put
this
hosted
remainder
unless
we
clear
these
errors.
B
B
Create
it
will
create
another
host,
sf03
and,
and
another
thing
right
now,
I'm
working
on
is
like
the
ability
to
create
a
host
in
maintenance
mode
from
the
dashboard
side.
So
if
I
create
a
host
like
this
and
check
this
box
and
create
this
host,
so
that
host
will
be
in
maintenance
mode
when
it
is
created
so
that
I
know
none
of
the
services
will
be
added
to
it.
While
the
time
of
creation
and
also
we
can
delete
this
host
like.
B
B
Yes-
and
this
is
all
from
the
host
side-
the
features
that
are
overlapping
between
sephardium
and
the
dashboard,
so
I
think
one
of
the
one
thing
we
can
add
over
here
is
like
when
we
are
creating
a
host,
we
can
include
like
adding
the
labels
when
we
are
creating
the
host
itself.
So
that's
some
of
the
things.
Maybe
we
need
to
create
over
here
too.
B
B
I
think
alpha
alphonse
yesterday
went
over
the
section
briefly,
so
we
can
see
all
the
services
running
over
here
and
service
and
disadvantage
the
service
running
on
this
host
same
with
this
nfs
and
like
that,
we
can
see
it
from
over.
Here
we
can
also,
if
you
want,
we
can
create
a
services
and
we
can
delete.
B
D
F
Of
that
just
to
comment
that
the
services
dialog
box,
I
think
that
is
too
much
generic
okay,
when
you,
for
example.
Yes,
okay,
that's
that's
the
the
dialog
box.
Okay,
we
have
something
that
is
common
to
every
kind
of
service,
but
I
think
that
we
need
a
special
section:
okay,
in
order
to
specify
configuration
parameters
that
now
is
possible
and
also
specific
attributes
for
for
each
of
the
of
the
services.
Okay.
F
So
I
think
that
in
order
to
do
that,
probably
we
need
to
provide
to
the
dashboard
some
kind
of
schema
okay
in
order
to
describe
what
what
are
the
attributes
that
the
user
can
can
use
in
its
kind
of
service?
Okay,
because
in
this
moment
I
think
that
is
too
much
generic
okay.
So,
in
most
of
the
cases
it's
going
to
to
deploy
something
that
is
by
default
and
probably
not
well,
we
need
to
use
expect
files.
A
D
Yeah,
that's
one
of
the
one
of
the
barriers
of
this
approach.
Right
now,
every
service
has
a
crafted
html,
please
to
say
so.
If
we
have
this
exchange
of
schemas,
we
can
avoid
this
kind
of
out
of
things
situations
where
some
changes
happen
in
sfadm,
but
we
notice
that
in
in
dashboard
and
so
probably,
if
we
we
can
agree
on
on
that
kind
of
schema
for
the
respect,
we
could
use
that
definition.
Instead
of
of
hard
coding,
the
different
inputs.
F
Even
from
the
part
of
the
command
line
interface,
I
I
think
that
is
going
to
be
also
useful.
Okay,
because
in
the
documentation
not
for
all
the
services,
we
have
all
the
possibilities
that
you
have
to
do
that
you
have
in
the
in
the
expect
file,
okay,
so
to
have
any
kind
of
command
in
order
to
provide
at
least
any
schema
of
what
are
the?
What
is
the
description,
the
possibilities,
the
attributes
of
each
service
is
something
that
is
going
to
be
useful,
also
for
the
command
line,
interface.
E
D
E
Yeah,
maybe
if
there's
like
just
a
like
a
an
example,
if
you
can
point
us
at
how
to
do
it,
that
would
be,
I
think,
one
other
question
I
noticed
here
like.
If
you
do
services
type,
the
types
include
like
mom
and
manager.
Those
are
singleton
services,
they
don't
even
have
an
id
and
you
so
you
can
only
create
one
of
them.
I
wonder
if
they
should
be
like
grayed
out,
if
you
already,
if
they
already
exist,.
E
Unless
this
is
a
like
a
conversion
from
a
like
an
adoption
conversion
from
a
non-step
idiom
or
whatever,
then
you
would
create
it
even
then
like
it
does
you
have
to
do
this
to
the
dashboard?
You
could
just
do
that
through
the
cli.
So
could
we
could
just
leave
those
out
entirely
for
the
create
path,
vermont
manager.
G
E
E
C
E
E
B
Yeah,
so,
okay,
okay,
I'll
move
on
so
ost
yesterday,
iphone
so
short
over
here
and
I
have
configured
a
nfs
over
here,
and
so
we
can
create
a
nfs
export
with
this
phone.
So
currently
we
don't
support
enough
v3
or
it's
only
v4
yeah
and
I
have
also
configured
the
file
systems,
so
it
can
be
seen
over
here
with
the
teachers
can
be
expanded
and
yeah.
The
object
could
be
the
other
thing.
C
B
G
E
Thinking
out
loud
here
that
when
you
do
the
create
service,
there's
a
section
of
it,
that's
related
to
the
placement.
I
wonder
if
it
makes
sense
to
have
like
a
little
a
widget
or
a
pattern.
That's
always
used
for
defining
the
placement
that
we
use
both
for
this
create
flow
and
also
for
the
edit
flow,
because
just
that,
the
way
that
you
describe
the
placement
is
a
the
somewhat
complicated
thing.
The
placement
spec
like
either
you
specify
explicit
posts
or
you
can
do
labels
or
you
can
do.
E
E
C
C
E
A
D
D
H
Yeah,
no,
my
only
small
topic
was
the
we
had
a
preview
discussion
sebastian
and
me
about
the
the
development
environment,
for
example.
This
one
that
is
using
a
nissan
in
order
to
have
several
hoses
is
based
on
a
virtualization
project.
Well,
an
upstream
project
that
is
running
on
top
of
little
bit,
which
is
the
kcli.
H
This
is
the
upstream
project.
The
the
fact
is,
this
a
little
bit
heavy
weight
to
have
these
virtual
machines
and
it's
not
trivially
set
up
and
and
unless
you
can
make
it
work,
and
we
thank
juani
for
providing
appropriate
documentation
in
order
to
make
it
work.
But
I
was
wondering,
is
it
was
a
way
for
having
a
kind
of
isolated
environment
to
test
better
theft,
ibm
not
relying
on
only
be
start.
Is
there
a
way
to
do
a
bootstrap,
but
not
messy
map
not
messing
up
with
the
laptop.
H
I
mean
I've
been
trying
to
do
like
containerization
of
the
of
the
of
the
manager
of
the
bootstrap,
sorry
of
the
sephie
dm.
So
this
would
need
like
a
docker
in
docker
or
docker,
and
it's
not
quite
easy
to
make
it
work.
So
so
I
was
wondering
an
alternative,
or
how
can
we
make
it
easy
to
provide
an
environment
that
we
have
the
possibility
to
add
some
hosts
or
or
a
more
lightweight
virtual
machines?
Or
I
don't
know.
H
C
C
E
I
think
currently,
the
it's
not
the
best
solution,
but
currently
the
best
way
to
do
this
is
using
the
start.
So
that
means
on
your
on
your
machine.
You
have
to
have
all
the
dependencies
installed
and
compile
and
build
ceph,
but
if
you
run
csv
start
with
fadm,
it
will
start
up
stuff,
adm,
yeah,
yeah,
yeah.
E
A
F
A
H
Is
what
we
are
using
in
kcli
yeah
juani
made
some
steps,
so
we
are
following
this,
but
the
fact
is
we
are
using
right
now
we
are
using
virtual
machines
yeah.
I
think
it's
more
likely
to
just
in
your
physical
laptop
to
to
to
run
the
boost
the
ffdm
bootstrap
right
yeah,
what
that
is
what
you're,
pointing
at
with
the
minimal
dependencies
installed
in
your
in
your
system.
C
C
H
H
I
E
E
H
H
F
G
F
It's
important
to
know
what
are
the
options?
Okay,
because,
probably
when
there
is
a
option
that
is
very
good.
Okay
and
we
are
most
of
us.
We
are
working
with
our
environments
and
probably
spending
some
time
in
order
to
have
the
environment
and
needed
to
to
develop.
So
maybe
we
can
create
or
schedule
a
new
meeting
in
order
just
to
discuss
this
part,
okay,
and
in
order
to
to
decide
to
to
use
the
same
environment
everybody,
because
this
is
going
to
be
a
good
thing
for
for
everybody.
A
A
B
A
C
A
Load
up
for
fossil
fuel,
so
my
my
thought
was
that
we
we
should
think
about
different
themes
that
we
that
we
should
should
think
of
like,
for
example,
looking
at
reducing
the
back
count,
then
increasing
making
the
developer
experience
better
by
looking
into
a
better
development,
environment
or
refactoring.
The
second
m
binary,
which
is
really
necessary
at
that
point,
skill
scalability,
is
a
thing.
A
A
A
C
E
G
E
I
think
there
are
still
a
few
decisions
to
make
there
as
far
as
exactly
how
to
how
to
structure
that
and
how
it
should
work,
whether
the
product
of
the
stuff,
idiom
scheduling,
would
be
config
options
and
the
resource
demons
or
whatever.
Exactly
how
I
don't
know.
I
don't.
I
didn't,
prepare
anything
beyond
what
we
were
talking
about
several
weeks
ago.
So
maybe
there's
not
a
lot
to
say
there,
but
accept
that
that
should
be
one
of
our
sort
of
deliverables
or
whatever.
A
Yeah,
the
the
current
way
to
remove
host
is
to
remove
all
daemons
that
you
can
remove
easily,
except
for
maybe
the
node
exporter
and
the
crest
human
right,
because
it's
super
awkward
to
remove
it
and
then
remove
that
hose
and
and
then
cleaning
up
the
different
things
later
on,
like
manually,
removing
the
the
node
exporter
and
then
manually,
remove
it
that's
a
bad
workaround
right.
That's.
E
F
I
have
played
a
little
bit
with
that:
okay
and
we
have
the
the
host
maintenance
mode
that
I
think
that
we
can
use
the
flag
of
hosting
maintenance
in
order
to
avoid
this
problem.
Okay,
even
in
this
moment,
if
you
try
to
do
the
the
removal
of
the
host
using
the
the
host
maintenance
mode
is
more
easy:
okay,
because,
for
example,
for
the
theft,
dm
exporter
or
the
diamonds
that
are
problematic.
F
F
F
A
Yeah,
it's
not
quite
the
same.
I
mean
if
we're
setting
a
host
into
maintenance
mode
and
we
are
expecting
it
to
be
restarted
at
a
later
point,
which
means
that
we
do
not
want
to
deploy
the
demons
on
a
different
host
that
are
living
on
that
demon.
But
if
you're
removing
a
hose-
and
you
say
I
want
to
have
three
ndss
and
and
by
chance
one
of
these
nbs
is
on
that
host.
A
C
E
And
if
you
think
about
osds
too,
like
it
should
it
should
trigger
the
oc
removal
sequence
also
for
those
where
the
osds
are
drained
and
then
removed,
and
then
all
the
demons
are
removed,
one
by
one
shouldn't
have
to
do
any
of
that
manually.
It
should
be
a
single,
basically
single
command
and
then,
like
an
hour
later,
it's
like
it's
ready
and
maybe
the
host
is
even
immediately
removed,
or
maybe
it's
just
like
marked
as
it's
removable.
F
C
E
G
C
A
C
F
A
A
C
A
E
E
A
C
A
F
I
think
that,
probably
for
me,
is
the
the
most
the
most
important
so,
for
example,
to
have
a
wizard
in
order
to
start
the
cluster
okay
in
the
dashboard
and
the
support
in
the
in
the
state
of
that
we
are
almost
there
okay.
In
order
to
to
make
easy
to
to
install
the
the
cluster.
I
think
that
is
absolutely
there
to
improve,
for
example,
the
sd
management
in
in
the
dashboard
and
to
provide
the
the
tools
in
order
to
do
this
very
easy.
F
I
think
that
is
another
point
that
that
is
important:
okay
to
improve
the
creation
and
the
the
editor,
the
maintenance
of
the
services
in
the
in
the
dashboard.
So
I
think
that
probably
what
my
my
view
is
that
probably
we
must
be
focused
in
the
high
layer,
the
dashboard,
okay
and
start
to
work
the
words
down
in
order
to
provide
all
the
functionality
that
we
need
to
make
very
easy
to
install
and
to
work
with
that
f
cluster.
Using
the
theft
dashboard.
A
A
E
I
guess
on
on
my
list:
there's
two
things
I
want
to
cover.
One
of
them
is
just
really
quick.
I
think
this
cefadm
exporter
is
going
to
be
like
a
one
of
the
themes
whatever
over
this
next
cycle,
I'm
making
it
making
it
work
reliably
and
probably
the
default
model
for
house
f80,
fading
of
scraping
information.
E
A
C
A
I
mean
why
not
right,
why
not
use
this
f
adm
daemon
to
to
to
listen
to
the
one
database
and
let
the
cell
phone
exporter
detect
that
yeah.
I
need
to
deploy
a
the
domain
on
my
host
that
would
remove
the
ssh
completely
remove
the
need
to
run
ssh
connections
to
the
remote
host
and
let
the
self-admx
portal
do
it
by
itself.
C
A
E
E
Oh,
I
hesitated
to
say
future
complete,
but
at
some
level
sort
of
future
completeness
before
we
tackle
that
that
big
scalability
jump
yep,
okay.
Well,
the
other.
The
other
big
item
that
I
wanted
to
just
to
talk
about-
and
this
might
bleed
into
the
work
discussion-
is
how
we
should
be
what
the
orchestrator
interface
for
aha
should
look
like.
E
First
of
all,
and
it's
it's
sort
of
combined
with
this:
it's
a
property
of
the
rgw
service
d,
whatever
it
is,
and
so
I
wonder
if
the
orchestrator
interface
should
do
something
similar
where
for
any
service
at
least
any
service,
where
it
makes
sense,
there's
an
optional
property.
That's
like
external
ip
and
whether
or
not
I
guess
this
is
really
just
the
external
led
and
then,
if
that's
set,
then
fadm
will
deploy
it
to
proxy
and
or
keep
the
live
d
as
needed
in
order
to
expose
that
service.
C
E
A
C
A
E
C
A
A
E
G
E
So
I
guess
in
the
rgw
case
I
mean,
I
wonder
if
whether
whether
we
have
this
intermediate
deployment
layer
is
sort
of
an
implementation
detail,
I
wonder
if
it
makes
sense
first
to
just
settle
on
what
the
orchestrator
api
should
look
like
for
this,
and
then
we
can
figure
out
what
the
most
elegant
or
expedient
way
to
implement.
That
is
in
the
rgw
case.
It
seems
like
that.
The
first
thing
that
we
would
want
is
in
the
kubernetes
for
kubernetes
and
rook.
E
E
E
C
E
E
F
I
am
with
the
with
the
option
to
start
with
something
simple
like
adding
the
virtual
ip
to
the
specification
of
the
service
and
try
to
to
use
that
to
install
it's
a
proxy
plus
keep
alive
okay
and
provide
only
one
kind
of
configuration.
Okay,
because,
for
example,
in
the
case
of
rcw,
what
we
have
done
is
just
try
to
mimic
what
we
have
in
this
moment
in
defensible.
F
Okay
and
the
same
parameters
that
we
have
in
the
defensible
playbook
is
the
parameters
that
we
have
in
the
specification
for,
let's
say
for
a
hazard,
okay
and
what
it
seems
that
it
was
the
the
thing
that
was
needed.
Okay.
So
what
I
think
is
that
probably
what
we
can
improve,
that
we
can
separate
the
a
and
use
keep
a
life
and
proxy
in
other
services.
But
maybe
the
first
step
is
to
call
this
thing
to
make
easy
to
use
this
to
the
final
user.
So
using
a
virtual
id
is.
F
I
think
that
the
only
thing
that
we
need
and
and
and
user
is
going
to
be
very
happy
with
that.
Okay
and
after
that,
in
the
background
we
just
to
deploy
it's
a
proxy
and
keep
alive,
it
is
going
to
be
more
complicated
and
it
is
going
to
be
the
layer
that
you
do
not
want
states.
But
I
think
there
is
no
no
other
possibility.
C
E
Okay,
I
think
that
that's
probably
fine
and
if
there's
actually
a
need
for
somebody
to
not
have
h8
proxy
for
some
reason,
h,
just
a
quick
question:
aj
pricey
can
do
the
ssl
termination
right,
yeah,
okay,
so
that's
actually
good
anyway,
because
it
it
combines
all
that
logic.
So
then
we
know
not
to
configure
rgw
with
ssl,
whatever
it
like
ignores
all
of
those
papers
over
all
those
details,
I
think
the
other.
The
other
key
use
case,
though,
is
for
nfs
and
that's
the
other
like
key
deliverable.
E
I
think
that
we
need
to
get
into
stuff
adm
supported
as
soon
as
possible,
back
porter
specific,
and
in
that
case
I
think
they're
at
the
very
minimum.
We
need
a
virtual
ip,
a
minimum
at
the
minimum.
We
need
the
vertical
ip,
but
my
understanding
is
that
the
fancible
is
also
deploying
itchy
proxy
in
order
to
balance
nfs
specifically.