►
Description
The BlockScience team takes us through an overview of their history with blockchain research and most recently, "Content Addressable Transformers".
https://github.com/BlockScience/cats
https://medium.com/block-science/the-cats-out-of-the-bag-introducing-content-addressable-transformers-7483e61e3844
A
Test
run
all
right,
hello,
everyone,
it's
joining
remotely.
This
is
our
sixth
session
in
the
afternoon
session,
the
apac
friendly
session
and
we're
very
fortunate
to
have
the
block
science
team
joining
us.
I'm
very
excited
because
I've
been
reading
a
lot
about
the
posts
that
they
had
recently
titled
the
cats
out
of
the
bag.
B
B
So
what
is
block
science?
I
guess
one
thing
that's
unique
about
us-
is
that
we're
a
multi-disciplinary
research
and
development
firm
and
we're
specifically
interested
in
complex
systems,
and
that's
almost
at
the
stage
of
you,
know,
mapping
the
system
and
asking
the
hard
questions
and
addressing
some
of
the
hard
problems
before
the
actual.
Just
you
know
day-to-day
or
straightforward
kind
of
software
engineering
front.
B
So
a
lot
of
the
research
we
do
is
primary
research
as
a
team,
and
then
this
has
led
to
some
open
source
software
development
as
well,
so
people
might
be
familiar
with
cad
cad.
The
computer,
aided
design
software
that
block
science
has
built,
and
this
was
really
around
you
know
solving
some
of
our
own
problems
as
we
embark
on
working
on
complex
systems,
specifically
those
in
the
web3
space
next
slide
and
I'll
hand
to
david
to
speak
about
some
of
the
things
that
we've
worked
on
and
why
compute
over
data.
C
And
then
sort
of
give
it
given
who
we
are
and
and
sort
of
our
relationships?
Why?
Why
do
we
care
about
compute
over
data?
I
think
kelsey
is
by
training
an
ethnographer,
and
I
think
it's
important
to
call
out
that
these
aren't
just
technical
problems
that
we're
trying
to
solve
they're
social
problems
as
well.
They're,
the
the
the
systems
that
we're
developing
are
effectively
algorithmic
policy.
C
So
here
we're
showing
a
transforming
element.
We
call
it
a
node,
it's
divided
up
into
into
pieces,
there's
a
there's,
a
structural
and
a
functional
component.
These
can
be
further
subdivided
into
plant
and
infrastructure
on
the
structure,
side
process
and
then
for
function
on
the
on
the
function
side.
C
C
There's
a
little
definition
at
the
bottom
of
this
slide
that
sort
of
gives
a
good
description
of
infra
function,
which
I
found
useful.
Basically,
it
pretend
provides
functions
that
are
things
like
protecting
providing
supporting
connecting
containing,
including
that
sort
of
thing
it
comes
from
a
book
on
on,
in
a
more
tradition,.
C
So,
given
that's
a
transformer,
what
what
is
then
a
cat?
So
if
you
take
transformer
and
add
content
addressability
to
it,
you
get
some
interesting
properties.
One
of
them
is
that
in
a
decentralized
system,
location,
addressability
doesn't
really
work
works.
Fine
in
centralized
systems,
but
in
decentralized
systems
location
is
frequently
not
known
or
if,
even
if
it
is
known,
it
may
not
be
accessible
or
it
may
not
be
convenient.
There
may
be
better
places
to
get
the
the
stuff
that
you're
looking
for
from.
C
C
C
The
function
in
the
structure
are
there
and
what's
added
in
here
is
is
what
we
call
a
component.
We
call
a
factory
that
basically
takes
an
order,
and
in
that
order
there
are
our
links
to
through
content
ids
the
structure
as
code
to
build.
You
know
the
structure
this
thing
is
going
to
run
on
the
function
is
code
to
build
the
the
the
functional
elements
of
the
of
the
cat
and
also.
C
So
you
know
I
I
love
quotes,
so
I'm
always
going
to
have
a
slide
with
quotes
on
it.
Why?
Why
are
we
doing
this?
There's
a
guy
sort
of
one
of
the
one
of
the
patrons
of
data
driven
systems,
development,
edward
deming,
as
a
quote,
data
are
not
taken
from
museum
pieces
they're
taken
as
a
basis
for
doing
something.
C
C
C
I
think
it's
it's
it's.
I
find
it
compatible
both
with
what
I've
read
in
the
compute
over
data
literature,
I've
gone
through
and
and
it's
certainly
central
to
cats.
So
if
people
aren't
familiar
with
it,
I
would
I
would
advise
that
I
would.
You
know,
certainly
recommend
you
take
a
look
at
some
of
her
ideas.
We've
got
a
reference
here,
so,
given
all
that
sort
of
how
do
cats
and
cods,
you
know,
how
do
we
put
all
this
together?
C
C
There
are
technologies
around
that
that
provide
that
things
like
file,
storage,
web3
storage,
provides
that
streaming
ceramic
network
provides
that
so
we've
got
our
data
layer,
an
action
layer.
You
need
to
be
able
to
process
the
data
or
it's
just
sitting
there
for
museum
purposes,
and
we
don't
want
that.
C
So
that's
I
think
you
know
compute
over
data
is
producing
tools
for
for
for
that
processing
being
able
to
process
data
in
a
decentralized
system,
and
then
we
see
cats
as
sort
of
working
in
this
domain
layer
as
providing
the
business
logic
that
define
what
what
the
products
need
to
come
out
of
this.
These
transformations
need
to
be
so
that
that
that's
our
sort
of
breakdown
and
a
lot
of
what
we
want
to
get
out
of
the
interaction
with
the
compute
over
data
work
group
is
sort
of
validation.
C
C
D
Okay,
so
what
is
a
cat?
And
why
are
they
useful?
Okay
cats
is
a
unified
data
processing
framework
for
decentralized
mesh
network
for
a
decentralized
mesh
network
to
empower
collaboration
on
products
implemented
with
multi-cloud
services
that
enable
data
provenance
by
content
addressing
the
means
of
data
transport
between
services.
That's
a
wordy
explanation
of
the
sentence
below
the
title.
D
Right
all
within
this
unified
processing
framework
right,
the
platform
will
empower
the
collaboration
on
products
across
domains
between
cross-functional
and
multi-disciplinary
teams
and
organizations
right
by
reducing
the
operational
overhead
of
adding
new
data
sources
via
decentralized
and
distributing
responsibility
to
those
collaborating
within
bounded
domains
to
support
continuous
change
in
scalability.
In
other
words,
if
I
use
bugs
words
to
describe
this
kind
of
value,
it's
a
multi-disciplinary
team
working
on
a
aspect
of
a
product
as
a
micro
product
deployed
as
a
micro
software
as
a
service.
D
D
D
Okay,
so
how
are
they
useful,
specifically
the
input
data
transformation?
I'm
sorry
cats!
Well,
no
I'll
wait
for
that.
One!
Okay!
They
enable
data
process,
verification
and
transport
between
services
with
ipfs
cids,
which
are
the
content,
addresses
all
right,
so
those
enable
maintenance
and
reporting
of
data
process,
provenance
and
data
or
process
lineage
as
a
result,
because
if
you
have
data
provenance,
you
get
data
lineage
as
a
byproduct,
which
is
a
business
intelligence
view
of
data
processes
on
the
mesh
right,
so
they'll
be
used.
D
These
record
entries
are
input
and
output
necessary
to
re-execute
a
deterministic
process
in
a
pipeline
of
post-processed
in
a
pipeline
that
has
been
post-processed
a
pipeline
of
cats.
It's
wordy,
but
I
had
to
prove
provenance
this
way
with
the
I
o
itself
right,
because
the
input
and
output
is
used
to
re-execute
a
cap.
So
that's
a
very
important
concept
for
the
next
slide.
D
The
cids
are
generated
by
the
cat
right
by
clients,
ipfs
clients
on
the
worker
nodes
of
a
software
as
a
service,
in
this
case
it's
spark
which
will
also
transform
the
input
data
cid,
that
provenance
record
entry
right
to
produce
an
output
which
is
a
combina
combination
of
another
record
entry
and
the
result.
Data
and,
as
you
can
see,
the
cloud
service
model
is
encapsulated
in
a
cat
for
which
a
multi-disciplinary
team
collaborates
on
a
single
attack
on
a
single
cat
queue.
D
This
is
an
example
of
david's
slide.
Above
that's
collapsed
between
the
cap,
processor
and
the
node,
so
the
processor
is
software
to
be
deployed
on
a
single
node.
I
thought
that
was
important
to
mention,
so
you
have
structure
which
is
a
user-specified
computation
or
data
transformation
framework
such
as
spark
desk,
etc.
D
This
is
an
example
of
this
is
an
example
of
that
isolation
in
the
user
experience
where
one
is
using
spark
sql
to
transform
a
data
set
abstracted
by
a
spark
data
frame
using
spark
sql.
So
I
decided
to
have
a
combination
of
like
you
know,
ansi
sql
and
just
like
spark
api,
just
to
show
you
that
you
know
it's
both
okay
cube
next
slide.
A
D
D
Okay,
so
I
just
want
to
explain
the
demo
before
I
demo:
okay,
just
to
tie
it
all
together.
So
this
okay,
I'm
using
cube
control,
okay
in
order
to
monitor,
monitor
execution
right.
This
control
plane
the
service,
the
secret
identity
and
access
management
is
in
terraform.
For
example,
cluster
roll
binding
is
handled
by
a
pi
spark
api
behind
the
hood
that
I'm
using
to
build
a
con,
a
docker
image
that
I've
modified
to
include
an
ipfs
client
on
that
will
be
built
from
docker
to
be
worker
nodes
of
a
spark
cluster.
D
Okay.
Here's
an
example
in
this
example:
there
are
two.
So
each
of
these
spark
workers
have
an
ipfs
client,
that's
being
used
to
as
a
storage
hack,
because
we
don't
have
web3
storage
right
yet,
but
there's
a
cache.
So
once
it's
ip
once
it's
cid'd
by
the
first
cat
that
cluster
is
non-ephemeral
for
the
purposes
of
you
know,
because
we
don't
have
storage
yet
right
and
the
second
cat
gets
those
cids
of
the
data
set
and
other
uris
necessary
to
process.
D
Itself,
I
guess
so.
I
just
needed
to
note
that
right-
and
this
is
an
example
of
I
would
say-
that
record
entry
I
referred
to
earlier-
we're
calling
it
like
a
bill
of
materials.
It's
a
supply
chain
term
that
we
still
so
you
have
input,
transfer,
transformation,
invoice,
data
output,
all
cid.
The
data
output
is
partitioned
in
the
case
for
distributed
processing,
everything
that
all
the
partitions
rcid.
So
this
is
kind
of
like
an
example
bomb
or
that
provenance
record
entry
that
is
retrievable
and
re-executable
from
the
ipfs
network.
D
So
that's
an
input
right,
so
the
first
cat's
purpose
is
to
content,
identify
the
data
set
because
it's
not
on
the
network
and
we
don't
have
storage
yet
okay.
This
is
an
example
of
an
m
of
everything,
critical
in
the
provenance
record
entry
or
the
bomb
necessary
to
execute
a
cat.
In
this
case
it's
creating
input
for
the
second
one.
These
are
the
same,
so
it's
like
cat
0
produces
the
input
for
the
subsequent
cat.
So
this
is
the
input
for
the
subsequent
cat
spark
transforms
it.
D
The
ipfs
clients
cids
everything
necessary
for
re-executing,
both
the
input
and
the
output,
and
then
you
have
an
output
bomb,
so
it
produced.
So
the
the
execution
of
the
second
cat
produces
an
input
bomb
and
an
output
bomb.
Everything
re
necessary
to
execute
re-execute
the
entire
pipeline
to
the
demo.
D
A
D
D
Maybe
a
side
by
side,
okay,
there,
okay
in
this
example,
I've
already
executed
cat
0,
because
I
want
to
show
you
that
the
bomb
cid
is
the
same
okay.
I
just
want
that
to
be
seen
hold
on.
Let
me
check
the
spark
cluster
right,
so
this
is
the
cat
zero
spark
cluster.
I
made
it
non-ephemeral
because
we're
hacking
storage.
D
D
D
D
There
is
an
input
bomb
because
this
cat
is
creating
the
imp.
The
output
of
this
cat
is
the
input
for
the
subsequent
clad,
so
there's
no
input
bomb
cid.
This
log
contains
like
ipfs
ur,
like
ipfs
uris,
which
change
okay.
So
that's
why
they're
not
in
the
bomb,
otherwise
the
cid
would
change.
This
is
just
the
idea
of
the
infrastructure
is
code.
D
D
Okay,
right
now,
the
dependencies
of
the
cat,
zero
application.
Okay,
right
now,
the
cat,
xero
application-
is
being
packaged
into
a
virtual
environment.
These
dependencies
are
being
installed
into
that
virtual
environment
as
well.
Okay,
so
the
virtual
line
is
created.
It's
installed
in
the
virtual
environment,
the
virtual
alignment
is
zipped.
D
D
C
A
A
A
D
Okay,
so
cats
in
this
example,
which
is
a
proof
of
cons,
concept,
the
cat.
The
reason
why
the
cat
is
non-ephemeral
is
just
the
storage
hack
cats
hasn't
been
deployed
right,
it
hasn't
been
deployed.
The
only
reason
why
it
sticks
around
is
because
I
don't
have
storage,
I
don't
have
content
address
storage,
so
what
you're
describing
is
a
cat
on
a
mesh
network
that
has
been
deployed.
D
D
D
B
D
D
D
D
D
D
A
Well,
fair
enough!
Well
I
I
want
to
say
I
I
really
appreciate
you
guys
taking
us
through
a
demo
because
we're
all
in
the
various
stages
of
building
and
so
we're
all
you
know
we're
all
building
together
and
just
seeing
what
you
guys
are
putting
together.
Just
seeing
the
you
know,
some
of
the
the
raw
inside
pieces
of
it
is
super
helpful.
So
I
I
very
much
appreciate
that.
D
A
Just
one
last
question:
for
you
in
terms
of
workloads,
you
know
so
in
it
separate
from
the
computer
data
working
group.
One
of
the
other
projects
I
work
on
is
bacco
yao
and
we're
trying
to
serve
deterministic
and
non-deterministic
workloads,
which
is
tough.
Do
you
guys
make
any
distinction,
whether
it's
you
know
deterministic
and
it's
always
going
to
return
the
same
output?
Or
do
you
in
this
case
primarily
promote
primarily
focus
on
non-deterministic
workloads,
just
whatever
the
data
says
at
the
time
it's
going
to
be
unique
and
different
for
each
other.
D
D
D
D
C
For
the
purposes
of
being
able
to
track
provenance,
immutable
data
and
item
item
potent
process
is,
is
a
is
the
goal
that
that
you
know
we'd
like
to
achieve
to
the
extent
possible.
You
know
what
what
do
you
do
to
a
process
that
has
a
random
number
generator
in
it?
Well,
you
can
get
around
that
by
passing
a
random
number
in
as
an
input
parameter,
something
along
those
lines.
What
to
what
extent
can
you
make
a
process?
D
D
D
A
Especially
if
you
look
at
like
a
lot
of
blockchain
processing,
you
know
like
popular
ethereum
type
processing,
it's
deterministic,
it's
the
same
output,
every
time
when
you
run
the
smart
contract,
but
then
in
our
world
of
broader
use
cases
it's
entirely
open-ended,
so
we're
always
kind
of
trying
to
figure
out
use
cases,
but
we
haven't
found
a
strongly
deterministic
use
case
either.
But
it's
helpful
to
hear
your
guys
perspective.