►
From YouTube: CNCF Serverless 9/28/17
Description
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
A
B
A
D
B
A
E
F
F
A
A
C
C
Okay,
do
we
want
to
start
or
wait
for
someone,
no
go
ahead?
Okay,
so
what
we
wanted
to
do
is
explain
the
architecture
and
especially
sound,
more
unique
tributes
in
nuclear.
Our
intention
is
not
necessarily
to
look
at
it
as
a
holistic
solution,
but
some
of
the
components
we
will
discuss
could
actually
be
reused
by
others,
or
we
can
try
and
absorb
other
components
from
a
other
project.
So
it's
important
to
understand
sound,
a
modeling
that
we've
done.
C
G
C
We
started
doing
that.
We
had
a
previous
version
which
was
proprietary
and
then
about
a
year
ago
we
decided
to
open
source
it
and
sort
of
rewrote
the
entire
code
and
in
order
to
make
it
more
open
to
other
data
sources,
it
and
sources
and
be
more
modular
for
future
use
cases.
What's
some
essential
things
in
the
architecture?
It's
it's
a
real-time
system.
It
has
notions
of
zero
copy,
zero
context,
which
is
no
garbage
cleaning.
All
those
aspects
that
come
from
a
real-time.
This
is
why
it's
performance
is
so
high.
C
We
we
did
want
to
address
all
the
aspects
of
life
cycle
so
versioning.
How
do
we?
You
know
test
in
one
place
and
deploy
in
another
place,
because
we
have
a
lot
of
IOT
use
cases
and
we
wanted
the
data
and
events
versus
to
be
pluggable
without
impacting
the
I
had
the
function,
because
we
use
functions,
for
example,
to
do
backup
jobs
in
our
systems.
I
want
to
just
like
the
the
right
data
set
that
I
want
to
backup
as
a
parameter.
I
don't
want
to
rewrite
my
in
my
application
and
say
for
events.
C
We
want
to
trigger
the
same
function,
sometimes
for
testing
from
an
HTTP
event
and
sometimes
from
a
stream
or
a
message
queue
and
the
general
idea
is
that
needs
to
be
portable.
Pretty
much
like
with
the
Amazon
Lander.
You
can
do
a
green
grass.
You
could
do
snowball
ads.
You
could
do
in
the
cloud,
so
here
it's
even
a
bit
more
modular
because
you
can
run
it
on
a
laptop
on
a
Raspberry
Pi
on
many
other
platforms,
so
that
was
our
requirements
for
the
system
it's
divided
to
serve.
C
The
main
component
is
the
processor,
which
is
the
thing
that
wraps
the
function.
I'll
dive
a
bit
more
into
it
this
and
then
this
processor
is
wrapped
by
on
one
side,
something
called
event
listener
which
intercepts
the
events
from
the
different
types
of
data
sources.
And
then
you
have
a
notion
of
data
binding,
which
is
the
thing
that
binds
with
the
different
data
types
like
databases,
stream,
etcetera
and
on
the
controller
we
have
several
components:
it's
run
as
a
bound
to
kubernetes,
but
when
it's,
the
current
implementation
of
controller
is
over
kubernetes,
it
could
be.
C
This
is
I
think
the
only
component
that
is
tied
to
the
underlying
a
platform
there's
another
component,
which
is
the
Builder
function,
Oscar,
runs
in
a
container
and
builds
the
artifacts
there's
something
called
dealer
which
I'll
explain
it's
a
thing
that
allows
the
dynamic
association
between
various
things
like
jobs
and
scars
and
tasks
and
functions
and
there's
also
aspects
of
user
interface.
One
is
the
CLI.
The
one
is
a
pretty
much
like
I
go
playground
and
we'll
see
how
it
works.
C
So
that's
the
general
architecture
and
if
we
dive
further
into
the
processing
engine
so
the
event
listener,
you
essentially
implement
a
server
class
that
implements
the
event
listener.
There
are
four
types
of
event
listeners
because
they
define
the
behavior
of
the
of
the
in-car
workflow
and
then
under
a
class.
You
have
different
implementations.
So
if
you
go
for
the
the
code,
you'll
see
that
there
is
like
synchronous,
asynchronous
stream
and
palling
implementations,
or
so
we're
probably
going
to
general.
It
is
best
jobs.
C
The
polling,
work
and
under
those
are
different
transport
implementations,
because
synchronous
is
like
a
request
and
reply
has
one
behavior,
asynchronous
or
independent,
set
tool?
Asynchronous
messages
and
things
that
we
discussed
you
know
streaming
is
sort
of
continuous
stream
of
messages
with
ordering
and
potentially
sharded.
So
those
are
the
the
plugins
they're
implemented
in
a
way
that
you
can
actually
have
zero
copy,
all
the
way
at
zero
context,
switches
all
the
way
from
the
event
source
to
the
function
and
the
other
side.
You
have
the
data
bindings.
C
We
still
need
to
do
some
more
work
on
that.
So
currently,
today,
Amazon
API
is
in
our
release.
Rio
is
supported,
but
in
few
weeks
we'll
add
all
the
rest
of
the
things
in
the
list
here.
This
processing
runtime
have
abstraction
of
the
interface
to
the
platform,
so
this
is
why
it
can
run
over
different
platforms.
C
So
all
the
things
around
logging
and
and
runtime
have
absolute
interfaces
which,
if
the
group
wants,
we
can
start
and
standardize
them
and
I'll
touch
on
some
of
them,
especially
like
logging
and
others,
and
the
function
itself
runs
in
the
runtime
engine.
If
it's
because
the
old
infrastructure
is
designed
around
my
golang.
So
if
it's
golang,
it's
sort
of
inline
go
routines.
If
it's
other
languages,
then
it's
there.
Few
invocation
mechanism,
one
is
like
a
bash,
the
other
one
is
like
unix
sockets,
and
this
is
going
to
be
a
shared
memory
implementation.
C
So
it's
not
like
some
of
the
platforms
where
it's
like
CGI,
that
you
invoke
the
process
process,
the
process
or
the
workers
are
kept
live,
and
then
you
sort
of
just
inject
the
message
and
get
events
and
other
responses
from
the
function
and
also
another
unique
thing
is
that
function?
We
we
have
multiple
workers
on
a
single
process
and
this
allows
us
to
get
very
high
concurrency
and
also
enable
this
very
high
performance.
C
H
C
This,
what
you're
saying
here
the
function
processor,
is
a
process.
Okay,
it
is
live.
The
orchestration
layer
needs
to
decide
if
it
needs
to
be
live
any,
for
example,
this
is
in
the
dock.
We
refer
to
something
called
starter,
so
if
the
function
is
down,
someone
else
intercepts
that
she
requests
loads
the
process
and
then
ships
the
message
into
the
process
in
case
of
message
queues,
it's
slightly
different,
but
the
idea
that
the
process
already
has
the
runtime
worm.
You
don't
necessarily
have
to
have
all
the
workers
up.
C
I
C
H
H
H
C
Because
what
we,
what
we
see
is
sometimes
we
deal
with
low
latency
use
cases
where
people
would
prefer
to
pay
some
more
for
the
resources
and
launch
and
have
server
zero
latency.
And
you
know
we've
seen
latencies
of
sort
of
0.1
millisecond
and
on
the
system
and
implication.
And
so
you
know
we
don't
want
to
add
another
100
or
few
hundreds
of
milliseconds
in
some
cases
to
the
invocation.
C
C
Getting
because
we
don't
have
a
lot
of
time,
let's
table
sound
the
discussion
if
some
of
the
questions
to
the
end,
okay,
so
one
unique
element
that
we've
initially
we
started
working
with
like
stateful
set
in
order
to
deal
with
the
sharding
challenge
of
how
do
you
distribute
multiple
charts
to
multiple
functions
and
eventually
we
decide
to
have
something
which
is
more
dynamic.
It's
called
the
dealer,
essentially
what
you
do
you
submit
the
job
and
you
tell
the
job.
How
many
tasks
and
task
could
be
very
different.
C
Things
could
be
a
number
of
actual
acting
charts.
It
could
be
if
you're,
sometimes
service
doesn't
feed
bad
jobs.
Let's
assume
you
want
to
do.
Etl
ETL
is
too
long
for
a
service
function,
so
you
may
want
to
break
the
table
to
a
thousand
smaller
slices
each
one
of
those
will
run
independently
in
checkpoint.
So
if
you
fail,
you
serve
start
from
the
same
point.
C
So
general
idea
is
that
you
break
a
job
into
a
bunch
of
tasks
as
many
as
you
want
there
all
sorts
of
metadata
things
like
you
know
how
many
tasks
in
a
process
do
concurrently,
etcetera
and
the
dealer
deals
with
sort
of
dynamically
distributing
the
shards
or
the
job
tasks
among
the
actual
functions.
Every
time
you
know
there
is
something
is
restarting,
it
will
reallocate
it's
actually
not
really
working
from
the
definition
of
what's
up
and
down
it's
it's
looking
at
the
deployment
expected
number
because
what
happens?
If
you
have
a
pod,
they
just
restarted.
C
You
don't
want
to
redistribute
all
the
shards
and
then
I
mean
a
second.
After
that
the
pod
went
up
and
you
redistribute
again
so
we
we
watch
the
both
events
of
up/down
events
as
well
as
deployment
scale.
They
change
events,
okay,
so
that's
a
very
generic
mechanism.
It's
actually
not
tied
to
nucleus,
so
we
can
use
it
even
and
other
micro
services.
C
They
could
run
in
parallel,
depending
on
the
metadata
you
see,
maximum
tasks
per
process
or
in
minimax
processes
and
all
that
those
parameters
or
they
could
run
in
a
cascading
like
I,
have
a
thousand
tasks
and
I'm
limited
to
only
four
processes.
So
essentially,
I
can
give
every
process
like
five
tasks.
They
finished
two
they
get
another
two
etc,
okay.
C
C
C
Also
we
demonstrate
this
week
at
strata-
is
that
by
having
you
know,
service
functions
that
first
are
so
fast
and
real-time
and
have
those
native
data
bindings
to
data
and
and
have
this
variability
to
because
go
beyond
just
simple
event-driven
to
also
sharding
and
in
stream,
processing
and
batch
jobs.
Then
essentially
it
can
replace
almost
anything
in
the
setup.
So
this
is
a
traditional
pipeline
that
we
work
with
customers.
You
know
you,
you
get
fees
from
whether
it's
operational
data
cases
like
ETL
or
CDC
or
stream.
C
You
do
all
the
pre-processing
on
those
strings
like
really
normalization
aggregation
prediction
from
chi
learning
against
a
data
source.
You
can
do
all
the
complex
event
processing
with
those
you
could
do
all
the
data
movement.
You
know
from
one
repository
to
the
cloud
and
back
you
could
do
even
the
API
is
that
serve
the
dashboards,
and
essentially
you
also
have
other
services
in
our
platform
that
deal
with
like
machine
learning
or
spark
or
other
things
which
are
you
know,
those
will
be
continuous
services,
they
have
their
own.
You
know
api's
and
tcp
ports,
etc.
C
Okay-
and
this
is
one
example
that
I'll
show
in
a
minute
that
how
we
essentially
take
an
event,
do
parallel
enrichment
across
multiple
tables.
Do
a
complex
event,
processing
work
against
like
four
different
streams
and
tables
and
geo
maps
60
lines
of
code.
All
of
that
example,
I
want
dive
into
the
next
few
slides.
It
will
will
be
something
that
you
could
look
at
after.
This
is
the
definition
of
the
like
job
spec
and
because
there
are
different
points
into
the
unique
aspects
of
you
know
things
like
data
binding.
C
You
could
notice
it's
a
kubernetes
CRD,
so
essentially
you
can
do
Kubik
at
all,
create
of
a
bunch
of
things
and
functions
in
the
same
and
command.
There
is
this
definition
of
events
that
are
shipped
into
the
function.
This
is
also
something
worth
discussing,
because
this
is
part
of
how
we
see
generalizing
the
event
structure
across
multiple
forms
of
events.
C
This
is
the
logging
interface.
We
also
suggest
to
sort
of
standardize
the
logging
interface
today
we
have
different
implementations
for
the
logging
interface
like
bring
to
the
screen,
pass
it
for
HTTP,
write
to
a
file,
push
it
to
event,
event
stream.
This
is
another
thing
that
I
wouldn't
like
to
have
us
maintaining
everything,
and
if
we
can
form
a
generalized
model
here
and
everyone
agrees
to
it,
then
we
can
have
people
very,
creating
plugins
for
logging
for
stream
logging
and
one
uniform
interface.
C
C
Okay,
so
good!
So
where
Iran
will
show
us
model,
it
will
show
us
the
demo
and
one
of
the
things
that
I
mentioned
it.
Nuclear
comes
with
a
built
in
sort
of
a
playground,
pretty
powerful.
So
you
could
you
don't
it's
essentially
a
kubernetes
deployment,
you
just
launch
it
and
it
works
standalone.
You
could
launch
as
many
as
you
want
in
parallel
and
and
what,
let's
so
Iran,
let's
start
with
no
show
a
simple
fight
on
you
could
see
right.
You
write
the
function,
a
simple
fight
in
a
getting
booked
by
the
event.
C
In
the
context,
the
context
also
gets
you
the
logging.
The
logging
is
booked,
structured
and
and
has
different
levels
while
and
then
you
just
compile
the
function,
run
it
from
the
same
playground.
You
can
actually
also
invoke
it
and
entity
that
the
return
response
could
be
either
structured
or
unstructured.
If,
instead
of
putting
a
structure
the
response,
I'll
just
put
a
string,
it
will
also
respond.
C
You
just
won't
have
the
flexibility
is
to
control
all
those
things
like
return,
status,
etc,
and
with
the
logging
level,
you
can
actually
control
in
real
time
what
kind
of
log
messages
are
going
to
be
shown
to
the
to
the
user
or
be
written
to
the
log
stream.
So
if
I
have
a
same
function
with
different
levels
of
logging
and
I
happen
to
run
into
a
challenge,
I'll
just
turn
on
debugging
and
run
okay,
and
so
that's
around
that.
Maybe,
let's
show
a
different
function
now
and
go
okay,
and
this
you
could
see
again.
C
Also
all
the
notion
of
the
logging
structures
start
to
set
Sarah,
and
maybe
you
can
show
us
the
performance
in
Roger.
So
what
you
see
in
nuclear
is
also
integrated
with
prometheus
and
we
have
our
own
tool
for
HTTP
benchmarking.
It's
pretty
fast
tool
generates
about
800,000,
HTTP
requests
per
second,
it's
open
source
and
what
you
see
well,
while
we're
running
the
function,
we
can
run
parameters,
also
look
at
the
HTTP
blaster
event
and,
and
you
see
the
the
results,
hopefully
yeah-
we're
running
an
empty
function
that
just
returns:
nothing:
okay,
yeah.
C
C
Okay,
so
that's
around
that
and
maybe
around
yeah,
can
you
show
us
also
how
you
invoke
some
function
from
different
sources
yeah.
So,
as
we
mentioned
the
same,
the
things
same
function,
signature
is
the
exactly
the
same
across
multiple
types
of
sources.
You
can
subclass
it
and
extend
stuff,
but
it's
pretty
comprehensive.
So
the
nice
thing
is
I
can
take
the
same
function
that
I
just
invoked
with
HTTP
and
now
invoke
it
through
rabbitmq.
Okay,
because
I
asked
the
function
to
listen
on
the
fae,
synchronous
and
synchronous
data.
C
C
So
until
now
we
sort
just
play
it
with
sort
of
stuff,
but
you
can
see
that
you
I
mentioned
to
you
this
real-time
IOT
demo,
that
you
know
intercepts
core
messages
and
and
does
a
real-time
enrichment
with
other
datasets.
This
is
actual
the
code
you
can
see.
The
first
line
is
the
data
binding
we're
actually
having
in
the
definition
language,
we're
saying:
DB
0
equals
to
certain
data
type,
certain
security
credentials
Sarah,
and
then
that
already
is
for
bound
to
my
context
and
I.
C
What
I
do
here
is
we
also
have
serve
like
pandas
or
data
frame,
as
Park
data
frame
implementation,
similar
to
that
in
Python.
In
go
sorry
for
real-time,
it's
also
going
to
be
open
source,
and
you
could
say
that
what
I'm
here
doing
here,
I'm
doing
far
a
little
read
from
four
different
tables.
You
know
cars
driver
is
wrote,
information,
doing
all
sorts
of
decision-making
underneath
you
know,
and
cop
the
geo
max
drop
date
card
driver
violation,
streams
all
that
stuff.
The
last
line.
C
By
the
way,
you
could
see
that
it's
or
waits
for
completion,
because
everything
here
happens,
asynchronously
all
those
five
different
updates
to
five
different
streams
and
tables
happens
concurrently.
So
that's
why
the
last
row
is
essentially
like
waiting
for
all
those
things
to
complete.
You
could
see
that
very
few
lines
of
code
I
managed
to
do
something
that
in
traditional
stream
processing
will
take
multiple
components
and
you
know
one
be
as
elastic,
and
this
is
an
elastic
function.
C
C
Hope
we
have
two
more
minutes
for
that.
So
just
show
us,
you
know.
So
there
is
a
full-blown
CLI.
You
can
see.
There
are
so
many
options
when
you
want
to
invoke
a
function
and
you
could
decide
where
do
you
want
to
have
the
repository?
So
you
know
things
like
IOT
use
cases
you
deploy
and
run
a
function,
create
an
artifact,
and
then
you
want
to
spread
the
same
artifact
to
multiple
edge
devices.
You
could
do
that.
C
C
You
know
maybe
show
us
the
publishing
feature
for
the
versioning,
just
it
okay.
So
the
nice
idea
is
that,
let's
assume
I
tested
the
function,
everything
is
cool.
I
want
to
now
label
it
and
tag
it.
So
that's
the
publishing
feature
and
what
I'll
do
with
the
event
source
I
will
define,
which
sort
of
label
selector
or
which
version
or
alius
I,
want
the
event
to
drive
so
now
I
could
have
event
source.
C
You
know
extreme
that
the
production
per
stream
and
matches
with
a
production
function
and
in
parallel
I'll
have
a
different
function
just
for
testing
and
I'll,
invoke
it
with
HTTP
or
another
stream.
Okay.
So
that's
the
that's
the
general
right
here.
You
can
update
a
bunch
of
things
to
each
one
of
the
versions
like
environment
variables.
You
could
label
functions
and
all
those
things
it's
very
comprehensive,
we'll
probably
won't
have
time
to
get
into
all
those
things
good
any.
So
thanks
Ron
for
okay.
C
A
C
A
All
right
cool,
thank
you
all
right.
So,
first
on
the
agenda
is
a
gentle
reminder:
people
who
have
been
trying
their
best
to
address
comments
in
the
doc
itself,
ignoring
this
section
of
what
is
service,
the
rest
of
the
doc
I
think
there's
less
than
about
a
handful
of
comments
that
are
outstanding.
So
if
there
are
any
comments
suggested
that
is
or
something
in
your
section,
please
try
to
address
them.
Don't
just
leave
the
comments
there.
A
If
you
think
you
could
rest
it
whether
that
means
actually
do
what
they
asked
or
ignore
it
or
something
or
reject
it,
go
ahead
and
resolve
the
comment,
so
we
get
disappear,
some
doc
itself.
My
goal
is
to
try
to
get
zero
comments
in
the
doc
really
really
soon.
So,
just
a
reminder
of
that.
So
let's
now
jump
to
the
conclusion
section,
the
doc
and
actually,
let
me
go
ahead
and
try
to
share
my
screen
here-
hold
on
a
sec.
A
Cool
all
right,
so
the
conclusion
section
first
I
wrote
this
I
guess
maybe
two
weeks
ago
and
I
find
it
very
hard
to
believe.
There
are
zero
comments
that
it
was
actually
perfect,
so
I'm
gonna,
just
haven't
had
a
chance
to
review
it
yet
so,
rather
than
taking
the
time
to
review
it
right
here
right
now,
what
I
would
like
to
do
again
is
just
encourage
people
to
please
read
it
as
soon
as
possible,
because
this
is
probably
the
most
critical
part
of
the
document
itself.
A
We
need
to
catch
this
as
soon
as
possible
because,
aside
from
the
what
is
service
section,
this
section
should
be
the
one
we
spend
most
of
our
time
on,
and
it's
worrying
me
that
no
one's
commented
on
there
because
I
know
I
didn't
do
that
good,
a
job
so
and
are
there
any
high
level
things
about
this
people
like
to
discuss
now
or
if
not
how
late
it
and
we
can
just
go
back
and
discuss
it
next
week.
But
please
take
a
look
at
some
point,
so
they
pause
here
any
high-level
comments
or
questions.
G
A
So
the
basic
goal
of
the
recommendations
section
is
to
provide
us
a
set
of
proposals
for
the
CN
CF
COC
to
consider
for
next
steps,
whether
it's
do
nothing
at
all
or
its
look
for
some
sort
of
standardization
effort
or
something
in
between
that's.
What
we're
trying
to
lay
down
here
as
a
set
of
recommendations.
A
They're,
all
any
of
the
things
you
just
mentioned,
there
are
valid
possibilities.
They
keep.
The
key
point
here
is
people
need
to
write
down,
identify
what
they'd
like
to
see
as
an
option,
and
then
we
as
a
group
can
decide
whether
that's
worthy
of
us
saying
to
TOC.
Yes,
we
like
this
idea
strongly
enough
that
we
want
you
guys
to
seriously
consider
it
great.
G
H
A
H
A
Goal
yep
definitely,
and
those
are
the
kind
of
things
that
I'd
like
to
see.
You
guys
add
to
the
dock
itself
if
I
have
already
covered
it,
wherever
you
think,
I
had
done
a
poor
job
of
covering
it.
Please,
edit
the
text
directly
there
and
like
with
all
other
suggestions.
Please
just
don't
leave
a
comment.
Saying:
hey
it'd,
be
nice.
If
you
talked
about
this
actually
add
the
actual
text
in
there,
it
makes
it
so
much
easier
to
review
and
and
to
understand
exactly
what
you're
looking
for
yeah.
C
I
think
another
point
beyond
the
user
is
also
the
echo
system
and,
yes,
we
work
with
guys
developing
edit
or
and
the
buggers
and
whatever.
If
we
maintain
some
some
ability
to
create
a
uniform,
the
ammo
files
or
unit
for
name
event,
structure
or
uniform
debugging
mechanism,
or
at
least
close
in
terms
of
the
model
that
makes
life
easier
for
the
echo
system.
And
if
the
echo
system
fosters
everyone
wins,
I
would
agree
with
that
part.
E
A
G
A
Don't
think
we
missed
them
out
on
this
call
right
now,
because
I
think
everything
you
guys
are
mentioning
your
great
ideas:
let's
get
them
in
the
doc
as
actual
textual
edits.
That
way,
we
can
look
at
them
in
context
and
and
we
look
at
them,
review
them
and
in
essence,
vote
or
approve
or
reject
them.
Yeah.
H
That's
right,
I'll
add
something
to
the
doc
of
another
suggestion
is
in
the
first
bullet.
We
say
we
encourage
more
service
technology
to
try
CSF
and
more
open
sources
party
places.
So
we
should
keep
this
window
open
right
because
there
could
be
more
providers
open,
source
solutions
coming
in
the
future.
So
how
we
should
I
mean
one?
Is
this
going
to
be
like
reported
to
the
CSF
TOC
or
what's
the
timeline?
Could
they
be
made?
Could
we
make
it
this
time
man
longer
well.
A
So
I
don't
want
to
push
out
Nestle.
Oh,
we
don't
have
a
formal
timeline
for
the
white
paper
other
than
try
to
get
it
done
as
soon
as
possible.
But
you
are
correct.
We
don't
want
to
necessarily
lock
down
everything
so
that
if
new
projects
come
along,
they
can't
be
added
to
something,
and
that's
that
quote
something
that
we
need
to
figure
out
right.
So
do
we
take
this
document?
Put
it
into
github,
so
we
can
have
people
issue
PRS
to
get
their
product
listed
in
there.
A
Or
do
we
talk
about
the
next
subject,
which
is
which
is
create
a
table
someplace
where
we
actually
list
these
things,
and
this
is
a
lead,
living
breathing
document
of
the
various
projects
and
tools
and
service
and
stuff
like
that.
That
might
be
the
place
to
list.
It
is
in
this
landscape
document
rather
than
the
white
paper
itself.
So
that's
something
else.
We
need
to
discuss.
Yeah.
K
Yeah
we
wanted
this
to
be
kind
of
a
living
document.
The
Google
Doc,
that's
kind
of
updated
and
we're
basing
this
on
what
the
service
I'm,
sorry,
the
storage
working
group
has
done
and
that
they
had
success
with
according
to
Chris,
yeah
I.
Think
the
the
key
shift
that
we
have
right
now
is
getting
the
white
paper
in
a
good
enough
state
to
get
out
the
door,
so
we
can
actually
start
working
on
some
some
plantations,
some
code
first
type
of
stuff
rather
than
too
much
in
terms
of
you,
know,
generic
wording
and
stuff
right.
A
A
So
when
we
talked
about
creating
this,
this
spreadsheet
did
we
agree
to
remove
the
appendix
for
the
document
that
listed
some
of
these
things
out
and
only
have
the
spreadsheet,
or
did
we
say
the
document
was
gonna
be
a
sort
of
like
a
snapshot
right
now,
but
the
living
document,
the
living
breathing
side
of
things,
would
be
the
spreadsheet.
My.
A
A
K
That
that's
why
I
was
agreeing
with
that.
Yeah
I
agree,
okay,
and
even
if
the
to
get
even
if
this
thing
moved
from
being
a
document
to
being
a
living
document
and
github
I
think
it
would
be
a
separate
item
anyway.
So
to
keep
it
external
from
this,
it
makes
sense
on
two
fronts:
okay,
so.
A
E
Sure,
instead
of
having
things
that
had
all
on
the
same
sheet,
it
made
made
sense
to
us
to
break
out
what
are
the
the
projects
that
are
delivering
more
serverless
types,
types
of
orchestration
platforms,
function,
schedulers,
executors,
split
out
development
tools
and
then
there's
a
services
tab
as
well
to
list
out.
You
know
what
are
what
are
the
recommended
or
preferred
applications
that
you
would
you
would
tie
into
for
event,
sources
that
are
known
to
be
interoperable
with
with
some
of
the
some
of
the
function
executors,
you
know
just
in
terms
of
the
venting
etc.
A
A
Actually
it's
in
the
dock
itself,
but
you're
writing
me.
Did
you
do
so?
Here's
the
link
in
slack
and
it's
in
the
meeting
minutes
as
well
right
here,
yeah
but
as
I
know,
you
guys
haven't
a
chance
to
really
look
at
this,
but
as
a
starting
point.
This
seemed
like
the
least
headed
in
the
right
general
direction
and
we
just
need
to
do
some.
Some
tweaks
I
think.
K
A
A
Okay,
I
didn't
think.
I
saw
him,
okay,
so
Sara.
Thank
you
for
your
peer
edits.
There
I
personally
like
the
direction
you
took
it.
The
question
is
I
think
it's
going
to
be
more
for
Chris
innate
than
anybody
else,
because
he
was
on
the
other
side
of
the
fence
in
terms
of
what
the
present
here
is,
whether
this
is
going
to
be
still
along
the
lines
where
he
can
live
with
in
terms
of
Amazon's
view
of
things.
A
G
There's
this
perspective
of
like
making
something
that
can
run
anywhere
or
making
something
that
can
run.
I
can
run
myself
where,
if
I'm
running
it,
myself,
I'm
kind
of
acting
as
both
the
provider
and
the
developer,
but
those
are
still
two
separate
concerns
and
and
I
was
kind
of
recalling
also
I'm
something
I
didn't
know
quite
how
to
put
in
the
dock,
because
I
couldn't
find
a
good
reference
to
was
unholstered,
where
that
seemed
very
relevant
in
some
way.
G
So
this
was
just
a
way
of
attempting
to
clarify
what
was
already
there
and
I
refrained
from
actually
I
believe
changing
anything
rather
to
separate
the
things
that
are
clearly
not
controversial,
which
is
like
zero
ops.
Everything
is
managed
for
you
as
the
developer,
deploying
the
function
right
and
then
there's
this
other
concern,
which
is
as
the
developer
using
this
platform.
G
What
are
my
costs
and
so
I'm
really
interested
in
the
working
group?
Thinking
like,
what's,
what
is
the
business
model
have
to
do
with
anything
and
and
at
first
I
wanted
to
strike
it
because
I've
been
most
of
the
dock,
I
was
considering
under
a
technical
lens
and
then
I
realized
that
anecdotally,
a
lot
of
the
developers
that
I
talked
to
are
and
seen
use
our
products
are
excited
about
the
it's,
the
business
model
of
server
list
that
drives
their
usage
of
it
as
much
as
the
like.
Its
developer
cost
and
dollar
cost
right.
G
J
K
I
think
we
had
that
originally
somewhat
in
the
attaced
reactor
right.
We
wanted
to
call
out
who
the
target
audience
was
for
this
paper
and
I
think
we
did
yeah.
We
agreed
that
it
was
Sara,
like
you
said,
there's
two
different
audiences
right.
There's
that
developer
with
the
point
of
view
of
serverless
and
there's
also,
especially
given
the
open
source
projects
right.
K
Those
people
that
are
providing
and
may
want
to
host
and
learn
about
that
and
I
think
that
yeah
that
could
be
crisper
in
the
paper,
but
I
think
we've
gone
back
and
forth
on
it
a
couple
times,
but
I
definitely
think
it
should
be
there
right.
So
clearly,
this
is
we're
talking
targeting,
and
this
is
who
we're
talking
about
when
we
mentioned
certain
things,
yeah.
G
C
I
was
a
little
buckle
on
the
slack
on
that,
but
you
know
again
whatever
we
decide
I'm
I'm
open
to
that.
But
you
know,
let's
assume
that
now
a
cloud
provider
goes
to
a
large
customer
says
you
know
what
you're
gonna
get
a
million
invocations
for
free
every
month.
Ok
or
you're
gonna
get
service
site
license
for
invocation.
If
you
put
a
million
dollar
in
front
of
me
so
isn't
now
you're
not
gonna.
Call
it
service.
G
E
A
Does
it
have
to
so
you
think
we
can
even
I
see
the
word
cost
I.
You
know
I
can
understand
why
someone
would
look
at
that
and
see
dollar
signs.
But
to
me
when
I
see
the
word
cost
I
actually
tend
to
think
more
about
resource
usage
and
I'm
wondering
whether
it'd
be
less
controversial
to
talk
about
resource
uses
instead
of
using
the
word
cost,
which
implies
monitor
things
but.
C
But
it's
very
indicated
there
is
actually
always
resources,
whether
it's
a
entropy
gateway,
listening
on
in
event,
whether
it's
some
stuff
that
is
waiting
on
cold
start
or
whatever.
There's
always
some
resources
and
I
think
that
what
here
is
more
about
the
business
model
am
I
going
to
charge
just
for
invocation
or
even
today,
that's
not
correct,
because
in
Amazon
provides
X
and
X
amount
of
free
invocations,
you
know
per
month,
okay.
So
how
does
that
fall
into
there's
no
cost
when,
when
idle
right.
A
H
I
think
I
agree
with
that.
You
know
because
I
think
different
I'm,
just
wondering
whether
we
need
to
be
so
restrictive,
I
think
different
vendors
proper
will
have
different
the
cost
model,
but
the
resource
will
be
on
usage
will
be,
has
I
mean
be
common
right.
You
know
no
matter
what
what
kind
of
you
know
on
Cosmo.
Even
the
function
is
not
running.
H
There
will
be
some
resource
being
your
system,
especially
if
it's
a
complicated
workflow
right
there
will
be
the
platform
the
provider
will
have
resources
to
manage
that
workflow
and
I,
and
you
know
at
one
side
is
you
know
the
provider
might
not
charge
for
some
resource
usage.
Email
function
is
Ronny
and
on
the
other
side,
some
some
provider
might
charge.
You
know
for
the
platform
usage
of
managing
the
whole
workflow.
I
Not
I
think
it's
problematic.
If
you
say
it's
it's
for
developer
resources,
then
warm
start
cannot
be
counted
as
surveillance
because
he
already
has
the
RIS
was
afraid,
no
matter
what
so
yes
said
that
even
someone
does
worms
that
it's
not
service
anymore,
because
it
costs
the
rhesus
for
the
developer.
I
agree.
H
A
That's
really
an
interesting
question,
because
it
seems
to
me
that
a
lot
of
the
a
lot
of
the
back
and
forth
we've
had
recently
is
around.
Is
it
absolutely
zero
costs
last
resource
users
as
seen
by
developer
or
is
a
minimal
amount
as
seen
by
a
developer?
Just
because
the
things
warm
is
that
okay
yeah?
So
we
called
service
my
correct
and
that's
basically
the
gist
of
the
disagreement
here.
I.
H
Yeah
I
think
that's
key
I.
Think
I'll
agree.
You
know
we
do
not
say
absolutely
no
cost
when
idle,
because
there
could
be.
You
know
when
the
function
is
not
running,
there
might
be
some
cost
minor
cost.
It
depends
on
the
you
know:
the
law
even
I'm
alone.
Currently
it
charges.
You
know,
as
I
said
on
one
site
it
charges.
It
will
not
charge
you
when
you
your
function
is
really
running,
not
idle.
They
give
you
some
free
free
runs,
but
on
the
other
side
for
the
step
function,
it
charges
a
platform
cost.
H
A
So
all
right
so
I
know
that
there
definitely
some
people
in
this
working
group
who
believe
serverless
equals
absolute,
zero
right
and
I
believe
there
are
other
people
on
the
call
who
who
don't
want
to
be
quite
as
strict
and
want
to
allow
for
something
greater
than
zero,
and
what
that
number
is
it
greater
than
zero
I,
don't
know,
I
think
it
might
vary,
but
they
won't
allow
for
at
least
something
a
little
bit
greater
than
zero.
Is
that
fair?
No.
C
I
think
I
wear
again,
whatever
we'll
decide,
I'm.
Okay
with
that,
but
I
think
we're
missing
the
point.
The
experience
the
operational
experience
is
the
thing
that
important
is
not
it's
not
what
you
pay
is
as
an
operation.
Guy
I,
don't
need
to
manage
server
is
that's
the
important
point.
I
want
to
run
a
function,
I
click,
something
I,
I
deploy
a
function
without
thinking
about
the
OS
that
I
need
to
deploy
without
thinking
about
the
docker
containers.
I
think
that's
the
core
thing,
not
not
how
you're
gonna
charge
me
and
business
models
are
flexible.
C
E
So
straight,
but
also
we're
looking
at
this
purely
on
a
technical
point
of
view
in
nuancing
it,
whereas
serverless
is
also
being
used
as
a
generic
term
throughout
the
industry
to
encompass
both
and
I.
Think
that
that's
getting
that
that's
playing
into
this
as
well,
but
I
know
I've
talked
to
some
of
the
open
source
projects
and
Nathe.
They
feel
that
you
know
they
provide
server
list
from
a
developer
point
of
view,
regardless
as
to
whether
they
can
you
know
who's
operating
it.
Yeah.
A
You
at
least
speaking
from
the
IBM
side
of
things,
I
think
if
we
were
to
take
the
approach
of
dropping
the
I
guess
it's
out
the
second
bullet
there,
the
absolute
no
cost
when
idle
or
dropping
the
discussion
of
that
from
this
from
this
definition,
as
you're
I,
think
you're,
proposing
I
think
we
would
probably
have
a
problem
with
that
because
to
us
part
of
service.
Is
this
getting
down
to
zero
resource
usage,
zero
cost?
Where
you
want
to
call
it?
A
L
This
is
a
I
I
know.
Chris
is
not
on
the
call,
but
for
Amazon
perspective
you
know,
Chris
made
those
three
points
very
clear.
You
know
zero
servers
to
manage
absolutely
no
cost
when
idle,
except
state
food
storage
and
instant
scalability
I'm
trying
to
ping
Chris
is
a
customer
meeting
right
now.
So
I
think
those
are
three
tenets
that
we
would
like
to
stick
to
as
well
right.
A
So
let
me
ask
you
this,
because
I
I'm,
not
what
I'm
trying
to
figure
out
is
I
mean.
Let
me
pick
on
your
own
for
a
sec,
it's
good!
It's
fun
to
come
up
with
a
good
person.
Do
you
think
it's
possible
to
find
wording
that
you
could
live
with?
That
still
keeps
a
discussion
of
that
second
bullet
point:
yeah.
C
I
wrote
you
first
time:
it's
okay,
I,
don't
get
funded,
usually
I,
think
I
think
the
compromise
could
be
to
strike
out
the
absolutely
on
one
end
and
also
at
the
usually
as
I
wrote
you
and
I
again.
It's
not
my
definition.
I
think
the
definition
should
focus
should
not
focus
on
whether
to
price
and
I
am
willing
to
bet
with
most
of
you
that
the
pricing
models
will
change
over
time,
but
I
think
an
acceptable
thing
is
to
strike
out
there
absolutely
and
right
usually
on
there
no
costs.
So
then
no
one
is
violating.
C
K
A
L
C
A
And
I,
like
everybody
else,
obviously
on
the
call,
because
obviously
there's
more
than
two
players
here
beside
your
on
and
Chris
I'd
like
everybody,
something
called
to
think
about
whether
they
cannot
say
this
is
perfect,
but
whether
it's
they
can
live
with,
because
I
think
we're
probably
reading
it
far
more
precisely
than
most
people
probably
read
it,
but.
A
L
G
This
is
Sarah
from
Google
and
I
noticed
from
that
Antonio
who's
been
more
active
in
this
working
group
was
strongly
advocating
absolutely
a
cost
for
an
idol.
So
I
I'd
like
to
check
in
with
him
and
get
his
thoughts
on
you
know,
sort
of
the
backstory
there
I
could
see
it
argue
different
ways,
but
I
think
I'm
still
I
still
don't
understand
what
we're
talking
about
like
are.
G
We,
we've
talked
about
resources
and
we've
talked
about
business
model,
and
so
it's
not
clear
to
me
whether
this
paragraph
is
actually
about
the
cost
model
for
the
developer
like
what
they
you
know,
whether
it's
in
dollars
or
in
deals
or
in
like
whatever
like.
Is
it
actually
about
the
business
model
which
can
apply
to
like
internal?
G
You
know
things
right
or
I
could
say
like
when
the
developer
is,
is
not
the
same
as
the
provider,
then
blah
blah
blah
right,
because
I'm
just
curious
about
that,
because
because
of
this
persistent
like
we're
seeing
this
huge
adoption
based
on
actual
dollar
cost
interest
right
and
having
things
scale
to
zero
is
an
important
characteristic
for
a
lot
of
firebase
developers
right
that
have
their
either.
They
have
Axton
development
or
they
have
mobile,
apps
or
spikey
traffic.
They
have
only
weekend
traffic
or
only
weekday
traffic,
and
so
the
dollar
cost.
A
Okay,
we're
almost
at
a
time
and
I
wanna
be
respectful
of
the
clock.
So
let
me
do
a
couple
of
things
here:
first,
because
Chris
spent
a
lot
of
time
and
effort
trying
to
word
this
section
very,
very
carefully.
That's
why
I'm
very
nervous
about
accepting
Sarah's
edits,
even
though
I
really
like
them
so
Aaron.
A
Can
you
check
with
Chris
to
see
if
he's
okay,
with
the
edits
that
Sarah
made
because
I
don't
think
she
actually
changed
any
of
the
controversial
bits
either
all
she
does
move
things
around,
I
think
I,
really,
okay,
so
because
we
can
at
least
get
the
the
syntactical
changes
that
she
made
in
there
agree
to
or
not
I
think
I
do
really
useful.
That
way.
People
can
can
make
their
own
minor
tweaks
if
they
see
a
Sarah
to
you
to
your
question
about
what
this
ex,
what
the
paragraph
is
actually
means.
G
A
G
Yes,
I'm
Sarah,
Allen,
I
joined
Google
about
a
year-and-a-half
going
about
nine
months
ago,
took
responsibility
for
the
parts
of
firebase
that
overlap
with
Google
cloud,
so
firebase
is
Google's
mobile
platform
and
gives
direct
from
mobile
access
to
Google
cloud
storage,
as
well
as
the
firebase
real
time
database
and
a
number
and
like
twelve
different
products
that
are
not
pop
rocks.
So
so
my
excitement
about
this
working
group
is
really
around
our
functions.
Effort
we're
originally
from
the
mobile
perspective.
We
came
to
like
well.
G
Mobile
developers
want
just
to
have
a
little
bit
of
function
in
the
cloud
and
then
what
I'm
seeing
is
that
mobile
developers
have
exact
same
needs
as
everybody
else,
because
the
backends
of
mobile
apps
are
incredibly
complex
and
almost
every
enterprise
has
mobile
app.
So
it's
just
a
really
interesting
space
to
me.
We're
very
active
in
you
know,
sort
of
we
have
a
lot
of
different
events
that
trigger
functions
and
I'm
in
a
long
time
on
the
internet,
but
new
to
this
space,
so
I'm
really
excited
to
be
involved.
All.
A
Right
cool
well,
thank
you
happy
to
have
you
here,
especially
with
all
the
great
edits
you
make.
The
doctor
really
appreciate
that
and
with
that,
I
can
have
to
call
it
time,
because
we're
definitely
out
of
time
and
we'll
talk
again
next
week,
and
please
look
at
that
conclusion
section
make
some
edits.
We
need
that
get
done
relatively
soon.
All
right.
Thank.