►
From YouTube: Mesos Developer Community Meeting (Jan 12, 2017)
Description
Tech Talk: Debugging Support in Mesos by Kevin Klues
Tech Talk: Partition-aware frameworks by Neil Conway
A
B
D
H
Let's
just
like
it,
oh
yeah,
what
are
people
paying
and
see
going
to
go
around
and
say
your
name
briefly,
and
this,
where
you
work
and
whoa
some
of
the
demos
things
coming
together
start
first
and
they
will
not
yield,
goes
afterwards,
we'll
go
through
the
roadmap
briefly,
and
how
do
I
know
what
we
could
release?
Not
your
guy
out
right
away.
D
D
H
G
F
B
H
C
Okay
thanks.
Yes,
as
Michael
said,
I'm
going
to
present
some
of
the
new
attachment
exec
support
that
we
added
into
masers
over
the
last
few
months
at
a
very
high
level.
What
we
basically
did
is
we've
tried
to
add
functionality
in
a
basis,
it
enables
building
tools
that
mimic
functionality
similar
to
dr.
attached
and
dr.
exact.
C
So,
if
you're
familiar
with
those
tools,
you
can
now
do
very
similar
things
with
mesas
1
sort
of
big
difference,
though,
is
that,
instead
of
needing
to
have
everything
running
locally
on
your
client
you're
able
to
actually
do
this
remotely
and
attached
to
a
remote
agents
and
stream,
all
of
your
output
back
to
your
local
clients
using
streaming,
HTTP
requests
and
responses
and
there's
no
need
for
ssh
access
to
the
node,
where
your
task
is
actually
running.
Everything
is
authenticated
through
the
mesas
authorization
authentication
mechanism
and
it's
streamed
over
HTTP.
C
The
signatures
for
the
different
commands
once
so.
We
have
some
API
is
that
are
in
place
to
enable
this
and
the
ideas
that
we
built
some
command
line
tools
around
it.
They
look
similar
to
the
to
the
you
see
down
here
so
familiar
again
with
dr.
exec
and
dr.
attach
these
should
look
very
familiar
because
we
basically
mimicked
exactly
this
at
least
the
signature
of
the
commands.
For
that.
So
you
can
say
you
know
instead
of
dr.
C
Specifically
some
of
the
details,
I,
don't
throw
I,
won't
go
into
all
the
details
on
how
we
built
this,
but
at
a
very
high
level.
The
way
we
enable
the
exact
like
commands
is
what
we
basically
do.
Is
we
leverage
the
nesting
container
stuff
that
we
added
to
support
pods
back
back
in
September,
to
allow
us
to
now
launched
a
new
nesting
container
whose
life
cycle
is
tied
to
the
lifetime
of
the
connection?
C
So
when
you
do
a
masive
task,
exec
command-
and
you
access
this
to
the
API
you're,
going
to
launch
a
nested
container
on
the
agent
that
you're
connected
to
and
whenever
that
connection
dies
your
nest,
the
container
will
disappear.
So
it's
a
little
bit
different
than
the
way
you
normally
launch
contain
it,
or
these
long
running
things
running
in
the
background.
C
It's
tied
to
the
life
cycle,
of
whatever
connection
you
happen
to
make
because
of
their
nesting
containers,
they're
isolated
in
the
same
set
of
C
groups
and
the
same
name
spaces
as
your
parent
container
they're
slightly
different
than
normal
default
style.
Mystic
containers
in
that
everything
is
actually
shared.
Now,
instead
of
having
some
some
of
things
that
aren't
shared
so
the
normal
mystic
containers,
for
example,
you
might
be
able
to
provision
a
new
file
system
or
say
that
you
want
a
different
network,
namespace
and
so
on
with
these
ones.
C
You
actually
always
get
the
exact
same,
see,
groups
and
same
name
spaces
and
everything
is
shared,
and
then
you
know,
as
you
would
expect,
were
then
able
to
stream
the
input
and
output
of
the
command
back
to
your
local
terminal.
So
it's
running
on
the
remote
machine
as
an
SD
container
inside
your
parent
container
and
all
of
the
standard
m
and
standard
out
is
redirected.
As
expected.
Does
it
require.
C
What
it
doesn't
require
the
default
executor,
it
doesn't
require
task
groups,
it's
its
own,
separate
API,
call
that
you
make
against
the
agent,
and
so
instead
of
going
to
the
master,
you
have
the
launch,
a
task
group
and
get
these
things
up
and
running,
and
then
that's
the
container
sort
of
fallout
from
that.
This
is
actually
a
I
know.
The
parent
container
idea
I
wanna
be
bug
or
that
I
want
to
basically
exec
into
I
launch
I,
directly
launched
a
new
Nesta
container
inside
of
that,
so
a
special
API
call,
which
is
called
launch
less.
E
C
A
long
time
you're
not
going
to
be
able
to
executive
those
are
attached
to
those
okay
right
now,
we've
made
the
limitation.
Well,
we
might
revisit
this
later,
but
currently
it's
only
possible
to
attach
two
tasks
that
are
launched
with
the
TTYL
exec
for
the
exec
commands.
You
can
exact
things
without
it
TTYL
they
don't
have
to
be
interactive.
There
have
to
be
anything
if
you're
actually
going
to
attach
to
something
some
long-running
container
that
you've
already
you
know,
started
launched
in
the
cluster
in
order
to
attach
to
its
input
and
output.
C
C
So
how
do
we
go
about
doing
this?
This
is
just
a
real,
quick,
high
level
architecture
diagram,
showing
some
existing
components
that
we
already
have
in
the
system
and
some
new
components
that
we
had
to
add
to
enable
this.
So
the
first
thing
to
notice
is
all
the
way
on
the
Left.
We
have
a
new
set
of
agent
api's
that
we
added
I'll
show
these
on
the
following
slide:
there's
basically
three
new
api's
to
an
agent's
that
you
can
interact
with
in
order
to
run.
You
know
to
build
these
executive
type
operations.
C
I
just
had
a
handlers
that
are
able
to
you
know
interpret
these
api
calls
as
they
come
in
past
us
off
to
this
new
component,
that
we
call
the
IO
container
switchboard
and
then
that
will
in
turn
launch
a
separate
process
per
container.
That's
in
charge
of
sort
of
redirecting
all
of
the
input
and
output,
that's
being
generated
by
your
container
and
passing
that
back
out
through
any
API
calls
that
have
been
attached
to
it.
C
So
the
three
API
calls
themselves
they're
listed
here,
the
first
one
is
launched
and
it
has
empathy,
some
combination
of
extreme
requests
and
streaming
responses.
The
launch
that's
a
container
session
itself.
What
it
basically
does
is
once
you
initiate
that
you
initiate
it
with
a
single.
You
know
fixed
size,
request
that
hits
the
agent
the
agent
says:
okay,
I'm,
going
to
launch
a
pneumatic
container
I'm
going
to
run
whatever
command.
C
It's
already
lunch
with
it
and
any
output
that
that
generates
it's
now
going
to
stream
back
over
an
HTTP
streaming
response
back
to
the
client
you
optionally
want
to
attach
standard
into
this.
You
can
then
follow
that
tall
up
with
a
second
attached
container
input
call
which
actually
uses
a
new
HTTP
streaming
request
mechanism
that
we
added
so
that
you
can,
you
know,
send
chunk
HTTP
requests
over
to
the
agent
with
one
persistent
connection,
feeding
the
standard
in
into
whatever
that's
the
container
session
that
you
previously
launched.
C
C
If
you
want
to
do
an
attached,
you
don't
need
a
Longinus,
the
container
session,
because
you're
now
attaching
to
something
that's
already
running
in
the
cluster
and
instead
you
use
a
combination
of
a
tab
container
output
with
the
taps
container
input
where
the
tap
container
output
is
just
a
simple
call
into
the
agent
and
then
you
stream,
the
output
back
without
launching
a
new
container
that
makes
sense,
as
I
meant
to
before.
This
is
fully
integrated
with
basis
is
built
in
authorization
and
authentication
mechanism
going
it's
to
further.
C
It's
also
fully
integrated
with
PCOS
authentication
and
authorization
mechanism
and
they'll
go
through
a
demo
that
I
have
in
a
second
we're.
Actually,
we've
implemented
its
into
the
beach
UFC
lie
and
I'm
going
to
demo
this
with
the
running
instance
of
PCOS
and
bringing
up
an
engine
X
container
and
trying
to
debug
it
and
do
some
things
around
that,
and
because
of
this
you
know,
you're.
C
Chair
authorized
access
so
instead
of
having
the
excitation
in
the
box
and
give
SSH
keys
to
all
the
people
to
allow
them
to
log
into
the
box
and
then
try
and
debug,
however,
they
normally
would
do
it
that
way.
Now
you
can
debug
in
as
a
container
using
attachment
exact
with
my
sis's
authentication
authorization
status
of
this.
The
api's
themselves
are
fully
implemented
in
the
mesas
agent.
We
have
a
reference
implementation
of
consuming.
Those
AP
is
in
the
D
cos
CLI,
and
we
have
a
native
masive
CLI
component
to
this
coming
soon.
C
So
it's
not
built
into
the
Mesa
CLI.
Yet,
but
I
wrote
the
DC
OSC
like
portion
of
this,
so
it
should
be
a
pretty
easy
back
port
to
the
Mesa
CLI.
Once
I
find
some
time
to
do
that
in
the
next
couple
weeks,
and
so
as
I
said
in
the
demo,
I'm
going
to
use
the
D
cos,
t
lie
and
interact
with
D
cos.
He
just
so.
You
can
see
how
this
all
works.
The
demos
themselves,
I
basically
got
three
demos.
C
image.
Cuz
I
have
a
task,
and
all
that
task
going
to
do
for
is
sleep,
for
you
know,
infinite
amount
of
time,
or
you
know
999999
seconds
I'm
going
to
allocate
it
a
GPU.
The
main
reason
I
want
to
do
this
is
I
want
to
show
that,
oh,
you
know,
I
can
launch
some
task.
I
can
then
launch
an
interactive
bash
session
inside
of
it
and
then
sort
of
see.
Okay.
C
What's
going
on
in
here
with
I,
really
allocated
only
one
GPU
from
the
total
amount
that
I
should
have
seen
the
resource
allocations
actually
set
up
properly
and
so
on,
and
then
in
the
third
demo,
I'm
going
to
bring
up
a
full-blown
bcos
cluster
running
marathon
and
metronome
and
I'm
going
to
show
how
you
can
debug
a
running
engine.
X
instance
inside
of
inside
of
this
cluster
and,
like
I,
said
I
used
the
D
cos.
You
live
for
all
these
demos.
C
So
some
of
the
details,
real
quick
about
the
demos
I'm
going
to
do
the
hell.
The
first
one
is
going
to
be
the
simple
hello
world
demo,
where
I'm
going
to
take
this
hello
world
SH
script,
which
all
it
does
is
echo
hello
world
back
to
the
terminal,
I'm
going
to
first
upload
that
script
to
the
tasks
container
using
sort
of
a
poor
man's
SCP,
where
I'm
just
catting
the
file
and
piping
that
into
you,
know
this
to
the
standard.
C
In
of
my
dcos
task,
exec
command,
which
forwards
that
into
the
container,
because
I
then
run
the
command
bash.
So
just
copying
the
file
over
and
giving
it
the
same
name
and
then
we're
going
to
use
dcs
task
exec
to
add
proper
permissions
to
the
script
so
that
I
can
execute
it
and
then
I'm
going
to
execute
that
script
inside
the
container
by
a
dcos
task,
exec
and
then
run
the
hello
world.
C
This
age
group,
the
second
one
I'm
going
to
do
like
I
mentioned
before
I'm
going
to
quickly
run
invidi
SMI
on
the
host
to
show
you
hey
on
my
host
machine
I've
got
four
GPUs
that
I'm
capable
of
handling
out
to
some
task
and
I'm,
going
to
start
an
interactive
bass
session
inside
that
container
or
inside
that
task
and
then
going
to
set
up
some
environment
variables.
So
they
can
actually
interact
with
the
nvidia
libraries
that
are
installed
inside
that
container
and
I'm
going
to
run
invidi
SMI
and
show
you
yes
and
be.
C
I
only
am
able
to
see
and
interact
with
one
GPU.
That's
been
allocated
to
this
task,
as
I
mentioned
before
the
third
demo
I'm
going
to
bring
up
vanilla
engine
X
container
inside
running
inside
of
PCOS,
I'm
going
to
verify
it
first
that
nothing's
serving
on
whatever
port
I've
allocated
engine
X,
because
by
default
I
allocate
some
port
2
engine
X,
which
isn't
port
80.
But
if
you
just
install
the
vanilla
in
genetics,
docker
container,
it's
going
to
think
it's
running
on
port
80.
C
So
you
have
to
change
that
up
a
little
bit,
I'm
going
to
start
an
interactive
session
with
that
engine
axis
container,
I'm
going
to
update
it
to
the
port
that
it's
actually
that
we've
actually
allocated
it
too,
so
that
it
can
be
listening
and
serving
something
on
that
I'll
restart
engine
X
and
then
I'll.
Show
you
that
yes,
engine
X
is
up
is
actually
up
and
running,
mounts
to
sort
of
now
walk
through
this
process
of
something
wrong
with
my
container
in
the
cluster.
C
How
do
I
debug
that
and
then
you
know,
fix
it
and
get
it
up
and
running
again.
So
with
that
I'll
jump
to
the
demo.
So
I've
sort
of
I've
already
said
some
of
this
up,
as
I
mentioned
so
I've
got
a
Mesa
master
running
locally
I've
got
to
make
those
agent
sitting
here
running
locally
and
the
first
thing
I'm
going
to
do
is
I'm
going
to
use
the
mesas
execute
script
to
launch
this
long-running
container.
You
know
for
now
it's
just
sleep
forever
kind
of
thing.
C
C
If
I
then
pop
back
over
to
so
now.
What
I
basically
have
listed
here
is
all
this
all
of
the
commands
that
I'm
going
to
run
for
you
to
these
demos,
but
so
the
first
demo.
The
various
first
thing
I
want
to
do
is
I
want
to
configure
DCOs
to
be
able
to
run
against
a
mace
O's
only
cluster
that
doesn't
have
DC
OS
installs
I'm,
going
to
run
first
command,
which
points
the
DCOs
record.
C
Bezos
master
URL
variable
at
my
local
host,
so
I
go
ahead
and
do
that
it's
not
set
up
for
that
and
I'm
going
to
you
know,
create
this
hello
world
SH
file
locally
and
then
going
to
as
I
mentioned
before,
I'm
going
to
copy
this
up.
You
know
into
my
container
running
inside
the
task
inside
the
tasks
and
dinner,
I'm
going
to
give
that
file
some
executable
missions,
okay
and
then
I'm,
going
to
run
DCOs
task
exec
and
execute
that
hello
world
SH
script
from
inside
the
container.
C
C
C
So
what
I'm
going
to
do
is
I'm
now
going
to
open
an
interactive
bash
session
with
that
running
task
shall
I
do
a
DCOs
task,
exec
I'm
now
inside
that
container,
as
the
root
user
I
can
navigate
around
do
whatever
I
wanted,
but
the
main
thing
I
want
to
do
is
I
want
to
set
up
some
environment
variables
to
allow
me
to
interact
properly
with
the
nvidia
libraries
installed
on
this
inside
this
container.
So
I
go
ahead
and
do
that
and
now
I'm
going
to
run
natively
or
sorry.
C
Now,
instead
of
running
on
the
host,
I'm
going
to
run
this
nvidia
some
ice
command
inside
of
my
container,
and
we
see
okay
I've
only
got
what
we
to
use
been
out.
Is
it
too?
So
you
know
I
came
in
if,
for
some
reason,
something
gone
wrong,
I
would
have
been
able
to
come
in
and
debug.
Okay.
Why?
Why
do
I?
Only
why
do
I
actually
have
two
GPUs
when
I'm
supposed
to
only
one
or
did
I,
that
allocation
actually
happen
properly
and
so
on.
C
So
you
know
Interactive
a
session
go
ahead
and
exit
out
of
that.
Ok,
so
the
third
demo
I
wanted
to
do
is
now
the
full-blown
G
cos
instance.
First
thing
I
need
to
do
to
get
this
going
is
I'm
going
to
reconfigure.
My
D
cos
CLI
to
not
point
to
the
native
maces
master
anymore.
I'm
going
to
point
it
at
a
DC
OS
cluster
that
I
previously
got
spun
up
and
running
and
then
going
to
log
into
that
cluster
with
the
by
default.
C
So
that
should
log
me
in
great,
I'm
logged
in
so
I'll
go
ahead
and
go
real
quick
to
the
UI
here
to
show
you.
This
is
the
cluster
that
I
have
up
and
running.
Currently
there
are
no
services
running
at
all,
so
the
first
thing
I'm
going
to
do
is
I'm
going
to
install
some
marathon
app
on
some
long
running
marathon,
app
that
has
engine
X
as
based
on
something
on
an
engine
X
container.
C
So
if
I
go
ahead
and
open
up
this
engine,
X
that
JSON
file
real
quick,
you
can
see
what
I'm
actually
going
to
install
on
this.
So
I'm
going
to
have
a
nap
lakay
shun
called
engine
ex
demo,
I'm
running
it
using
the
unified
message
container
Iser
the
doctor
image
that
this
is
going
to
be
based
off
of
is
the
standard
library
engine
X
instance
our
docker
image
that
I
want
to
base
the
stuff
of
I'm
going
to
run
again
this
long-running
command,
so
that
I
can
come
in
and
interact
with
it.
C
Okay.
So
with
that
I'm
going
to
go
ahead
and
deploy
that
application
inside
you
see
us
so
I,
say
deseos
marathon
app.
Add
that-
and
it
says
great
I've
gone
ahead
and
created
that
deployment.
So
now
that
that's
there,
I
can
come
back
to
the
UI
and
I
see
okay,
here's
this
engine,
ex-demo,
that's
running!
You
can
see
that
it
thinks
it's
running,
that's
great,
so
it
actually
spun
up
and
is
trying
to
serve
something
somewhere
now.
C
Well,
what
is
my
public
IP?
So
what
I
would
want
to
do
is
I'd
want
to
say:
okay
well,
I'll
run
DCOs
task
exec
on
this
container
and
try
and
curl
this
special
URL,
but
amazon
gives
you
lets
you
say
hey.
What
is
my
public
IP
address
if
you
curl
this
magic
URL
amazon
gives
you
that
back
so
what
I
want
to
do
is
first
run
that,
and
it
says,
oh
no,
such
file
or
drug
that
happen.
C
It
actually
happened
because
the
default
engine
X
container
doesn't
have
curl
installed
on
it,
for
whatever
reason
so
I
say:
okay!
Well,
that's
unfortunate!
How
can
I
fix
that
I'll
come
in
run
DCOs
task
exec
to
give
them
giving
me
an
interactive
a
session
in
there
I'm
going
to
update
the
container
I
running
apt-get
to
you
know,
install
some
new.
C
Binaries
inside
here
I'm
going
to
now
install
my
curl
command
so
that
I
can
go
ahead
and
use
it
and
then
going
to
rerun
the
command.
I
tried
to
run
directly
up
here,
but
I'll
run
it
from
within
the
interactive
bash
session.
Now,
so
I
go
ahead
and
do
that
oops
and
when
I
do
I
see
some
public
IP
pop
out
of
this.
C
So
great
I'll
go
ahead
and
grab
that
guy
come
back
to
the
UI
and
I'll
try
and
now
access
it
over
its
public
IP,
but
again
because
engine
X
by
default
serves
on
port
80,
not
on
this
random
port.
That
I
happen
to
have
allocated
it,
I'm
going
to
need
to
go
in
and
reconfigure
engine
X
to
serve
on
this
port
instead,
in
order
to
actually
get
this
up
and
running
and
see
some
output
from
this,
so
I
come
back
to
the
set
of
commands.
I
want
to
run.
C
C
I'm
now
able
to
run
some,
you
know
more
sophisticated
command
that
requires
a
TTYL
on
the
remote
end.
I'm
going
to
modify
that
now.
Instead
of
listening
on
port
80,
I'm
going
to
listen
on
port
80,
31
16,
so
I
go
ahead
and
modify
that
close.
That
out,
once
that's
closed
out,
I'm
going
to
restart
engine
X
and
Gen
X
is
restarted
and
assuming
everything
worked
right,
I
should
be
able
to
refresh
this
and
engine.
C
X
is
now
running
on
this
port
refresh
and
there
you
are
welcome
back
welcome
to
engine
X,
so
I
was
able
to.
You
know
see
that
there
were
some
problems
in
my
cluster
launch,
interactive
bash
sessions
launched
and
random
commands
inside
my
container
fix
things
up
and
get
things
up
and
running
without
having
that
ever
ssh
onto
the
agents
where
this
container
was
running.
C
So
those
were
all
three
of
the
demos,
they'll
jump
back
to
the
slides
here
and
just
really
quickly
give
special
thanks
to
everyone
that
helped
out
with
this.
It
was
a
huge
effort
between
a
lot
of
people
to
actually
get
this
going.
There's
people
working
on
the
CLI
there's
people
working
on
streaming
for
a
request.
There
was
people
working
on
the
security
aspects
of
this
lots
of
different
pieces
had
to
come
together
to
make
this
happen
in
the
time
frame
that
it
did
so.
C
I
C
K
True
yeah
and
the
user
is
so
you
first
do
the
login.
But
of
course
that's
cruelty
see
the
slogan
if
33
with
that's
how
that
I
don't
know
how
that
works.
But
how
that
interacts
with
basis.
Is
that
message
because
you
said
it's:
the
oulton
of
n,
which
is
just
a
still
a
list
of
principles
in
cigarettes
right,
so
Terry,
free
predefined,
so
a
fixed
list
of
principles
and
secrets
right.
So
this
is
separate
from
the
the
actual
unix.
I
Is
I'm
baking
powder
and,
and
the
user
that
you
want
to
specify
is
something
that
you
protect
by
the
maesters
acl's.
We
had
an
oasis
for
that,
but
that
says
which
user
you
can
login
asked.
Maybe
web
container
essentially-
and
I
think,
is
healing
just
hard
codes
into
root
root
user
doesn't
set
it.
I
don't
think
it's
such
a
at
all
and
even
a
chabad
is
wonder
if
they
accept
ryu.
Third
part,
it's
a
prototype,
but
it's
protected
by.
So
there
are
acl's
that
we
are
related
measures
for
these
api
calls
as
well
thinking
protects.
K
C
K
I
J
E
I
J
E
C
C
H
D
A
Oh,
he
do
all
right
great
okay,
so
this
talk
is
some
slides
on
the
work
been
doing
on
partition
awareness
of
remarks,
so
what
problems
of
a
kind
of
sovereign
solve
three
kind
of
distant
problems?
So
the
first
is
that
tradition
basis
is
defined
only
one
policy
for
what
happens
when
he
should
be
construction
from
from
the
master.
So
you
know
daily,
ms,
you
know
try
to
avoid
applying
policy
for
applications.
I'm
going
to
let
frameworks
handle
people
of
what
someone
what
happens
in
past
inspiration
I.
They
respond
that.
A
Second,
we
use
this
task
that
has
tasked
lost
to
define
to
meet
a
lot
of
different
things
that
make
16
different
places
represented
and
those
places
have
fairly
different
semantics.
So
we
wanted
to.
We
want
to
break
that
down
this
in
a
little
more
granular,
so
that
it's
easier
interpret.
What
how
you
said
to
respond
to
the
status
of
it.
You
got
it
and
the
third
one
is:
there's
no
signal
right
now
we
tell
you
when
a
task
becomes
partition,
but
it
can
be
useful
to
know
some
of
the
time.
A
Lasers
can
tell
you-
or
they
already
tell
you
that
has
the
worst
part-
is
he's
never
going
to
come
back,
so
we're
looking
at
how
to
communicate
that
information
framework.
So
it
has
is
not
just
I
mutable,
but
it's
gone
forever.
So
let's
go
look
at
these
in
order.
So
how
does
person
handling
working
lazos
one
point
earlier?
So
there's
a
pretty
simple
fix
policy,
so
it
basically
moves
us
about
health,
checked
agents
when
an
asian
fails
the
health
check.
A
So
if
it
just
fails,
respond
to
five
innings
in
a
row
within
15
seconds,
each
by
default,
Mason
declares
that
agent
removes
it
from
the
cluster
losing
the
Registrar.
Any
frameworks
that
are
running
tasks
on
that
agent
will
get
a
pass
gloss,
status,
update
and
the
slave
lost
callback
will
be
invoked.
A
Then,
when
the
agent
reconnects,
if
it
does
reconnect,
then
the
agent
will
be
shut
down
and
all
the
tasks
on
the
agent
will
be
terminated.
No
I,
notably
the
flavor,
is
not
consistent.
If
the
national
filter
or
flashes
build
over,
the
agent
will
be
allowed
to
reconnect
and
its
chances
will
go
back
one.
So
what
are
the
totals?
Are
coming
semester,
things
here?
A
Well,
the
first
one
is
that
there's
only
one
policy,
though
all
forums
have
to
use,
so
you
thought
well,
she
probably
makes
sense
if
you're
running,
you
know
it's
a
stateless
applications
where
you're
happy
to
just
spawn
I'm
against
ass
somewhere
else,
but
a
different
frameworks
might
well
want
to
handle
unreachable
casting
different
ways.
They
might
need
two
different
kind
of
before
they
launch
a
place
with
task.
They
might
not
always
want
loss,
contra,
costa
county.
A
Another
variant
of
that
is
that
a
single
framework
might
want
to
treat
different
tasks
that
it
starts
in
different
ways,
so
suppose,
yet
any
service
framework,
the
way
in
which
you
handle
the
data
node
failing
it's
probably
very
different
than
way
that
you
hear
on
the
internet
fail.
You
might
have
many
hundreds
of
data
nodes,
you
might
have
a
handful
of
magnums.
A
Often
you
know
some
frameworks
are
going
to
want
for
consistency
in
that
it
might
certain
tasks
you
might
use
a
resource
or
in
in
you
know,
research
that
need
to
be
accessed
the
mutual
screw
support.
So
they
might
all
logic
at
most
one
instance
of
a
framework.
Other
can't
other
frameworks
I
want
to
maximize
availability.
They
want
to
say
you
know:
I
need
to
have
at
least
five
copies
of
this
task.
I
have
more
that's
not
a
big
deal.
A
A
15
hours
over
the
right
now
36
learn.
Okay,
I
am
so
so
this
in
your
coaching
days
was
1.1,
but
it's
only
enabled
if
you
opt
into
your
capability.
I'll
talk
about
design,
so
the
new
behavior
is
to
say
that
mazes
will
no
longer
sharp
enough
protection.
Agent
eat
chocolate
capability.
Mrs.
will
also
no
longer
kill
tasks
on
those
person
agents
when
they
rear
again.
Instead,
you'll
get
a
status
update
that
your
task
is
unreachable,
which
is
not
a
terminal
status
with
you.
A
So
you
find
out
okay,
this
task,
our
joining
an
agent,
the
agent,
the
Masters
lost
contact
in
the
agent,
the
task,
maybe
I'm
running
again,
and
it's
up
to
you
to
decide
how
to
handle
that
transition,
either
from
running
immutable
from
a
neutral
back
right.
So
you
know
it
can
be
a
little
bit
more
work
if
you
were
just
using
the
default.
A
The
default
whole
thing,
but
I
doesn't
the
flexibility
to
choose
how
to
handle
those
events,
and
actually,
if
you
really
did
it
proper
because
we
always
allow
lost,
has
to
come
back
after
the
natural
field
over
really
to
get
this
waiver
correct,
you
would
have
need
to
handle
that
transition.
Anyone,
so
you
know
actually
correctly
written
production
framework
should
have
very
little
changes.
The
bib
anemic,
so
a
lot
of
compatibility
that
that
behavior
is
only
enabled
if
you've
stuffed
by
this
particular
capability.
A
So
if
you
don't
specify
it
when
the
end
agent
comes
back,
I'm
so
remember:
well
is
the
capability.
When
the
agent
comes
back,
when
every
registers
the
agent
will
never
be
shut
down,
so
that's
a
behavior
change
that
is
applied
regardless
of
disability
and
when
the
agent
rear
editors
forgot
that
ability
we
will
kill
non-production
where
tasks
on
the
engine
unless
the
agent
I
less
the
master
scale
over.
A
So
it
is
we're
trying
to
emulate
the
ultimate
where,
in
the
past,
your
alliance
uncertain
terms
on
your
hats,
getting
killed
so
we'll
try
to
kill
tasks
in
the
same
set
of
circumstances
going
forward.
We
might
make
this
person
workability,
equal
and
some
futuristic
missiles,
so
somebody
else
he
added
he
is
the
ability
to
find
out
when
accessing
declare
an
interval,
so
this
would
be
filled
with
us
as
message.
Any
idea
there
is
that
your
firm
is
down
when
and
ask
for
an
agency
competition.
A
It's
actually
pretty
difficult
to
tell
how
long
its
be
controversial
I've
been
on
each
before
so
you
know
the
kind
of
way
we
expect
things
to
respond
to
unusual
task.
We
have
to
start
a
timer
and,
depending
on
how
sensitive
the
task
is,
how
important
that
ability
is.
You
know
the
cost
of
you're
going
to
work
out
of
multiple
copies.
You
might
wait
some
period
of
time
and
then,
after
that,
waiting
I've
heard
of
time.
G
A
Okay
yeah,
so
we
store
this
set
of
the
set
of
agents
that
are
so
in.
You
know.
One
changes
that
we're
now
storing
the
set
of
a
neutral
agents
in
the
registry,
so
there
are
some
kind
of
performance
since
your
honor,
if
you
doubt
that
I'll
talk
about
community
so
yeah,
so
that's
the
first
major
feature:
it's
basically
changing
the
way
that
we
deal
with
Gretchen.
She
passes.
A
A
You
know
16
places
task.
Log
large
failbook
has
become
unreachable,
the
task
of
running
the
nation,
of
a
shutdown,
explicit
reconciliation
for
an
unknown
task,
and
so
on
and
more
than
just
sending
this
status
in
different
places.
The
semantics
like
how
you
would
respond
to
that.
As
of
that
event,
as
a
permanent
author
would
have
been
different,
you
know
the
way
that
you
respond
to
a
gala.
Castellan
is
probably
to
be
different.
A
The
way
the
response
you
a
racial
caste
or
it
right,
so
we
want
to
clarify
this
so
going
forward
yeah,
but
really
definitely
task
lost
yet,
but
in
the
future
we
will
be
deprecated
in
when
you
opt
into
this
part
from
our
capability.
Instead
of
getting
task,
lost,
you'll
get
a
set
of
more
granular
past
states,
so
we're
going
to
send
your
calf
dropped.
A
So
the
last
thing
that
we
wanted
to
address
part
of
this
work
is
in
frameworks
warm
for
me
on
when
tasks
are
gone
forever,
so,
as
I
said
kind
of
the
way
that
a
lot
of
inertia
can
handle
rh
bill
task
means
they're
going
to
wait
to
see
if
the
past
comes
back,
because
you
know
when
a
possible
problem
stripping
system
is
very
particle
in
between
that
pass.
That
is
just
slow
and
passed
it
to
actually
fail.
A
So
if
it's
go
to
respond,
you
know
best
responder
than
70
seconds
will
say
that
it's
unreachable,
but
it
may
or
may
not
still
be
running.
It
may
and
may
not
go
back
tomorrow.
So
the
way
you
handle
that,
as
you
use
a
timeout,
try
to
say
you
know,
maybe
I
want
immediately,
replace
it,
but
I'll
replace
it
after
waiting
a
little
bomb.
There
might
also
be
cases
where
it's
not
safe
to
replace
the
task
at
all.
You
know,
maybe
you
are,
you
know,
accessing
some
crucial
resource,
so
I'm
shared
disk,
for
example.
A
Now
you
don't
want
to
disrupt
our
turquoise
and
ask
unless
you're
very
certain
that
the
original
task
is
a
preacher,
so
in
some
circumstances
we
can
do
a
little
better
than
this,
so
in
general
we
can
distinguish
between
slow
tasks
or
counseling.
We
can't
contact
the
past
later
failed,
but
there
are
gonna
be
some
cases
where
make
those
knows
that
a
task
is
definitely
not
running
so,
for
example,
if
the
agent
will
shut
down,
we
know
the
past
and
graceful
shutdown,
tremendous
tasks
we.
D
A
That
it
has
two
definite
on
Brian,
then,
when
we
communicate
that
homeworks
they
could
respond
more
aggressively
example.
They
could
cancel
their
trend
up
and
do
something
more
properly
or
maybe
they
could
replace
a
task
that
they're
really
sensitive,
and
you
want
to
make
sure
that
the
previous
coffee
is
definitely
gone.
A
So
we're
looking
at
at
sending
this
task
on
status,
update
when
names
of
snows
the
cast
is
definitely
no
longer.
So
again,
that's
an
optimization.
It's
not
something
in
the
line,
but
as
a
framework,
you
could
actually
respond,
give
a
better
user
experience.
So
we
could
send
that
basically
agents,
gracefully
shut
down
are
there
by
a
zinc
user
one
or
when
it's
taken
down
by
the
Machine
down
network
so
from
be
involved.
I'm
modifying
the
shutdown
sequence
also
that
the
master
knows
for
sure
the
agent
shut
down
before
presenting
estas
Vonnegut
now
there's
another
case.
A
A
We
are
you
kind
of
using
casa
updates
to
describe
transition.
There
lies
the
whole
eight,
so
so
every
test
you
get
this.
That
means
every
task
that
have
been
running
on
that
agent
on
right,
and
you
know
you
could
make
it
where
you
learn
and
again
about
situations
where
you
know
just
the
cast
down.
That
would
be
more
expensive
for
a
minute.
First
oral
test
ideas
out
there
shaking
reviews.
So
the
idea
is
mostly
the
case
for
right
away.
A
Is
also
basically
reading
it.
Never
we
mean
case
for
one
past
generation
is
reachable
the
other
ones
so
ritual,
so
we
I
mean
we've
also
talked
about
having
a
waiter
to
look
at
the
life
cycle
about
the
whole
machines
or
means,
but
right
now
we're
kind
of
all
through
this
one
is
always
thought
everything
was
lip
service,
so
get
some
task.
Lawyers
for
the
case
where
Mays
those
know
is
that
the
agent
is
gone.
They
can
prove
that,
but
the
other
cases
where
mazes
won't
know
that.
A
A
So
love
or
you
know,
some
criticisms
will
actually
build
in
mechanisms
where,
let's
say
you
nourish
your
filesystem,
you
want
to
take
over
a
shard
of
the
house,
Lewis
home
by
one
node,
to
really
make
sure
that
that
the
the
note
that,
on
the
previous
previous
task,
that
on
that
chart
is
gone,
you
might
us
another
bad
way
to
make
sure
that
had
the
previous
test
is
definitely
are
made.
So
you
know
you
can
skirt
ugly
turn
off
the
power
yoga
change.
E
A
A
90-minute
yet
we're
still
kind
of
a
finding
these
accidents,
and
that's
one
thing
we're
trying
insider
you
know
there's.
Basically
we
could
say
that
the
agent
ID
is
gone
so
that
age,
maybe
you've,
no
you're
running
all
the
castle
rock
on,
but
there's
also
also
people
would
like
to
know
when
the
coal
mines
the
resources
that
agent
had
our
are
definitely
not
just
for
me
so,
depending
on
which
way
we
go
here
like
you,
get
also
had
some
kind
of
volume,
reconciliation
or
volume
lexical.
So
that's
kind
of
sucking
air
is
exactly.
A
G
G
I
Booking
agent
on
the
past
and
ethnic
is
gone.
Ok
and
one
dollar
can
murder,
wait
guys,
let's
evolution
knowledge,
the
monkeys
days
are
doing
with
keeping
dating
a
game
because
really
good.
So
that's
a
lesson
I
use
guess
we're
useful,
but
a
vegetative
have
to
definitely
gone,
but
your
voice
and
stuff
might
still
be
previous
regulation.
I
A
So
yes,
I
was
saying
like
at
least
what
we're
talking
about
now,
wouldn't
wouldn't
give
you
wouldn't
guarantee
that
the
resources
in
particular
for
some
volumes
associated
that
agent
don't
come
back
with
the
defamation,
I
things,
but
we
actually
might
change
a
key
cuz.
That's
certainly
something
that
people
care
about.
So
so
one
last
thing,
so
we
careful
about
how
we
manage
staking
so
before
we
do
this
work
makes
us
kept
a
list
of
registered
agent
in
the
rugged.
A
Log
is
arbitrary,
so
that's
that
list
is
going
to
be
bounded
by
the
size
of
your
cluster.
So
now,
we're
also
now
will
also
be
keeping
a
minimum,
the
set
of
currently
unreachable
agents,
and
if
we
had
gone
in
my
bar
operator,
we
will
meet
the
list
of
gone
from
barbarians,
so
on
with
the
set
of
registered
agents
set
of
elite
relations
could
grow
without
bound.
A
Know
whether
might
see
unknown
when
you
reconcile
for
the
state
of
a
very
old-
and
so
you
know,
the
idea
here
is
the
frameworks
would
at
least
be
able
to
handle
unknown
somewhat
gracefully
to
the
schedule,
so
1.1
ship
with
the
change
of
the
Persian
ballsier.
One
point
you
get
most
of
the
new
past
lettuces
and
then
143
overlooking
a
cast
on
right
and
I,
usually
get
a
little.
You
I
won't
put
yogurt.
So
thanks
very
much
happy
the
big
questions.
K
A
G
So
all
the
you
so
keen
it's
an
older
game
would
you
go?
Are
you
hungry
general
schedule
a
beer
leave
anyway,
we're
working
with
that
same
is
at
one
point.
Oh
and
it's
do
something
at
one
point.
Your
uncle
even
starts
at
the
time,
so
that
favor
I
think
would
understand
only
ask
launched
yet
because
it's
an
order
paper,
but
it's
going
to
get
these.
You
know,
that's
know
how
would
the
master
take
it
out
with
it
should
be
turning
us
lost
muscle,
tees,
so
moves
will
the
framework
presents
the
additional
work
ability?
A
That
it
has
to
okay,
we
will
see
some
differences.
I
give
you
a
script
that
we're
looking
at
metrics
in
the
master
of
example,
and
looking
at
how
many
lost
tasks
they
were
sure
if
you
have
a
frame
of
the
launcher,
the
task
that
is
partition,
aware
that
has
tons
of
mutual.
He
already
has
a
neutral
metric
want
enhance.
What's
the
matter,
leave
your
kitchen
Michael
yeah,.