►
Description
Teodor Heggelund interviews Chris Nuernberger about his open source work in building the Clojure data science ecosystem. Tomasz Sulej and Daniel Slutsky joined the conversation too.
The interview took place on April 26th 2020, and the ecosystem has taken some huge steps of progress since then. Still, everything in this conversation is very much relevant to understanding Chris' view and the core principles and goals of his work.
You are encouraged to explore more:
* https://github.com/techascent/tech.datatype by Chris is a Clojure library for efficient N-dimensional numerics across a range of primitive datatypes and a range of container types.
* https://github.com/techascent/tech.ml.dataset by Chris is a Clojure library for data processing and machine learning.
* https://github.com/scicloj/tablecloth by Tomasz is an additional API wrapping tech.ml.dataset.
* https://clojurians.zulipchat.com/#narrow/stream/236259-tech.2Eml.2Edataset.2Edev is the relevant stream to discuss these in the Clojurians Zulip chat.
* https://clojurians.zulipchat.com/#narrow/stream/151924-data-science is the relevant stream for broader discussions of data science in Clojure.
A
B
C
A
So
our
goal
for
this
interview
is
to
explore
those
changes
and
see
if
we
can
build
a
broader
understanding
of
them,
we're
not
aiming
for
a
general
programming
audience.
We
assume
some
film
familiarity
with
with
closure
and
a
motivation
for
data
science
yeah.
So
hello,
chris.
Thank
you
so
much
for
for
joining
us.
A
Okay,
great
yeah,
so
chris
we
were
talking
briefly
offline
about
what
work
you
had
done
previously
and
I
was
listening
to
the
podcast
where
daniel
compton
was
interviewing
you
that
gave
me
some
time
perspective
on
what's
happening
the
last
the
last
time,
but
can
you
take
us?
Take
us
back
give
it!
Why
are
you
interested
in
high
performance
computing
and
what
has
brought
you
to
to
where
you
are
now.
D
You
usually
end
up
writing
a
lot
of
code
and
a
lot
of
that
just
little
bits
of
incidental
complexity
that
you're
trying
to
avoid
taking
a
lot
of
time
in
a
bad
situation.
I
I
really
like
lisp
a
lot.
I
like
common
lisp
a
lot
and
I
feel,
like
clojure,
solves
actual
structural
problems
and
common
lisp
that
can't
ever
go
away
regardless
of
what
libraries
you
use
in
common
lisp
and
the
biggest
structural
problem
is
the
platform
it
sits
on
the
jvm.
D
D
I
would
say
yes,
but
because
I
feel
that
the
performance
in
this
environment
is
very
difficult.
I
aim
for
at
least
making
sure
the
baseline
performance
is
physically
possible
and
then
backing
off
from
baseline
performance.
That's
what
I
want.
I
try
to
make
it
more
use
usable
over
time,
and
that
way
I
at
least
know
that
I'm
checking
it's.
I'm
checking
my
architecture
against
a
real
thing,
every
a
lot
of
the
steps
of
the
way,
not
every
step,
but
a
lot
of
the
steps
of
the
way.
A
D
D
D
A
Yeah,
so
I
really
want
to
go
into
the
apl
and
lisp
distinction
later.
That
was
something
that
really
interested
me
based
on
documentation
that
you've
written
for
techymal,
but
first,
if
we
can
detour
a
bit,
can
you
give
us
some
impression
of
what
you
were
doing
at
nvidia?
What
was
your
yeah?
What
was
the
goal
of
of
your
work?
There.
D
D
So
physx
is
a
physics
library
that
nvidia
produces.
That
has,
of
course,
excellent
gpu
bindings
and
some
of
the
most
advanced,
real-time
physics
you
can
ever
hope
to
do,
and
they
for
an
advanced
physics
engine
you
need
to
know
like.
Why
did
this
guy
walk
through
that
wall
like
why?
What
happened
here?
Why
did
the
physics?
Why
did?
Why
is
everything
bouncing
around
the
whole
level?
Really
fast?
You
have
to
be
able
to
debug.
D
D
So
we
would.
I
would
you'd
open
up
a
network
socket
to
the
debugger
from
the
physics
engine
and
the
physics
engine
and
your
game
engine
would
send
little
updates
of
every
single
object
in
the
engine,
and
this
adds
up
to
quite
a
lot
of
information
over
a
decent
sized
game
like
batman's
arkham
asylum,
where
it's
a
it's
a
aaa
title
and
there's
literally
thousands
of
things
all
the
time
bouncing
into
each
other
and.
D
Yes,
kind
of,
I
actually
got
a
really
bad
case
of
rsi
before
ever
on
the
the
project
that
got
me
hired
at
nvidia.
I
got
a
really
bad
case
of
rsi
on
working
with
a
lot
of
other
people
and
at
the
very
end
this
was
all
in.
This
was
in
c
sharp
back
in
the
day
when
you
would,
you
had
a
bunch
of
c
sharp
had
come
out
and
the
new
ui
platform
had
come
out.
B
D
Very
precisely-
and
I
remember
you
know
they
said-
I
mean
one-
the
general
response
to
the
team
was
thanks
kid.
We
got
it
which
is
fine
and
another
response
was
we
don't
think,
there's
there's
advanced
enough
compiler
technology
to
implement
the
things
that
you're
saying
and
then
one
guy
who
is
actually
pretty
high
up
said
yeah.
D
The
future
is
immutable
and
functional
for
sure,
but
we
don't
know
how
to
get
there
from
here,
and
so
that
also
had
a
pretty
profound
effect
on
how
I
thought,
and
so
I've
kind
of
been
interested
in
this.
This
overlap
of
of
functional
programming,
but
also
the
highest
performance.
You
can
possibly
get
on
a
machine.
So
given
a
problem
where
you've
you've
spent,
you
know
a
lot
of
time,
making
it
really
fast.
D
D
If,
when
you
describe
your
program
in
a
purely
functional
way,
then
the
dependency
graph
of
your
computations
is
part
of
your
program,
because
you
can't
change
something
previous,
so
you
you,
you
know
by
just
analyzing
the
tight
flow
of
the
program,
the
dependency
graph
of
what
depends
on
what
eventually,
when
you
were
talking
about
and
we'll
come
back
to
this.
This
actually
applies
to
this
conversation
very
directly.
D
Eventually,
knowing
your
dependency
graph
allows
the
system
and
knowing
what,
for
instance,
is
a
commutative
operation,
allows
your
system
to
auto
parallelize
vast
portions
of
it
or
allows
you
to
parallelize
portions
of
it
without
having
to
think
about
various
things.
So
that's
a
theoretical
argument.
D
The
concrete
argument
is,
if
I'm
a
c
plus
compiler
and
you're
passing
me
a
bunch
of
things
that
aren't
const
in
a
pointer
every
time.
I
access
that
pointer.
I
have
to
load
it
from
memory
because
I
just
don't
know
what
it's
changed.
It
outside
of
my
purview-
I
don't
even
if
you
pass
me,
two
pointers
in
a
function.
I
don't
even
know
that
one
pointer
doesn't
reference,
isn't
aliasing,
another
pointer,
so
I
might.
D
Even
if
you
have
like
two
pointers
to
a
function,
you
load
them
both
into
local
variables
to
you
that
might
be
a
local
variable,
but
every
time
you
write
to
one
I
might
have
to
reload
the
other
under
the
covers,
and
so
you
can
imagine
how,
once
I
can
make
things
past
structs
into
the
function
instead
of
passing
variables
by
pointer
where
the
compiler
is
just
copying
the
whole
struct
verbatim
into
the
function.
Now
the
compiler
can
make
the
same
reasoning
that
we're
trying
to
make
at
a
higher
level.
A
C
D
And
then
you
found
out
that
you
were
causing
all
these
horrible
assembly
level
errors
that
were
causing
the
processors
to
have
to
reload
data.
All
the
time,
because
you
were
passing
pointers
around
the
whole
program
and
you
never
knew
which
section
of
the
program
would
affect.
Another
section
of
the
program.
D
A
A
D
Yeah
I
mean
it
can
be
needed
when
you
have
rapid
changes
to
a
small
percentage
of
your
data
set,
and
you
do
need
actual
to
broadcast
those
changes
between
a
lot
of
different
entities
at
once.
So
you're
dealing
with
shared
memory
on
disk
and
you're
actually
trying
to
organize
a
couple
programs
around
it
or
you're.
You
just
have
one
program
where
you
do
have
some
reason
that
you
need
some
core
of
the
program
to
be
visible
to
the
whole
program,
all
the
time,
so
a
shared
state
of
a
board,
or
something
like
that.
D
Although
I
mean
I
know
rich's
like
original
example,
was
the
shared
ants
kind
of
wandering
around,
and
that
was
a
beautiful
example.
But
as
you
try
to
scale
that
up
you
end
up
with
so
many
you
know
so
many
conflicts
of
people
trying
to
write
that
you
actually
can
get
a
lot
more
performance
by
just
being
more
disciplined
about
blocking
certain
things.
B
D
I
got
into
common
lisp
and
then
similar
to
rit.
I
got
into
common
lisp
and
then
closure
came
out
and
you
know
I
watched
rich's
simple
made
easy
and
that
really
meant
a
lot
to
me,
although
you
may
not
see
it
because
I
tend
to
write
big,
complex
libraries.
Well,
that
was
close,
but
it
did
actually
mean
a
whole
lot
to
me
and
I
wasn't
a
fan
of
the
jvm.
D
D
D
D
But
expert
systems
are
essentially
big,
rules-based
kind
of
clever
prology
kind
of
things
and
those
some
of
those
are
still
around,
but
a
lot
a
lot.
A
lot
of
those
companies
failed
as
various
server
architectures
that
they
had
tightly
bound
their
lisp
implementation
to
went
out
of
vogue,
and
they
actually
had
such
hardcore
implementations
of
various
different
details.
D
They
could
not
port
those
big
list
systems
to
x86,
and
so
there's
there's
a
huge
amount
of
loss
there
and
loss
of
knowledge
there
due
to
the
failure
due
to
how
tightly
bound
those
lisps
implementations
were
to
the
underlying
hardware.
I
mean
they
had
list
machines
that
they
had
bound
it
to
at
the
assembly
level
to
do
things
and
I'm
sure
that
worked
great.
But
then,
when
everybody
in
the
world
didn't
want
to
risk
list
machine,
you
lost
your
market
base,
really
quick
and
there's
all
sorts
of
mistakes.
D
I
mean
most
of
those
companies,
oh
man.
This
could
get
me
in
trouble.
Let's
because
I
have
an
engineering
run
company.
If
most
of
those
companies
are
engineering
runs,
so
they
were
more
interested
in
research
and
exploring
what's
possible
than
they
were
on
market
capitalization,
and
so
when
downturns
hit-
and
there
were
two
relatively
I
mean-
there's
one
in
like
75
and
there's
another
one
in
83-
they
weren't
in
a
position
to
survive
the
hard
periods.
D
B
D
A
D
Agree
with
that
survive
probably
do
run
on
everything,
but
the
x,
it's
just
a
rough.
It's
a
rough
system
to
use
common
lisp.
Now
a
lot
of
the
things
that
you
take
for
granted
like.
Oh
there's,
a
collec,
at
least
in
java.
You
have
a
collections
interface
at
all,
and
it's
not
just
like
the
lisp
is
completely
the
list.
Data
structure
is
completely
unrelated
from
the
vector
data
structure,
which
is
completely
unrelated
from
the
map,
data
structure
and
oh
yeah.
D
D
D
No
so
I
first
started
doing
clojure
in
2008,
and
I
have
on
my
the
oldest
repository
I
still
have
access
to
on.
My
personal
github
is
a
little
graphics
editor,
where
I
could
run
a
little
3d
scene
and
I
could
edit
the
shaders
and
it
was
written
front
to
back
enclosure
and
uses
all
sorts
of
horky
things,
because
I
was
like
still
in
the
mode
of
c
plus
plus
so
like.
It's
got
some
weird
architectural
concepts
in
there,
but
it
did
work
and
that
was
in
2008.
A
D
D
Maybe
I
can't
remember,
but
it
was
really
hard
to
use
and
it's
since
fallen
out
of
favor,
and
so
eventually
we
just
decided
to
go
further
down
the
route
into
neural
networks,
and
actually
I
wish
we'd
have
chose
to
go
just
down
the
route
of
basic
machine
learning
first
and
then
neural
network
second,
because
you
can
actually
get
really
far
with
basic
machine
learning.
Just
like
xg
boost.
D
All
of
machine
I
mean
pretty
much
all
machine
learning
is
some
some
level
of
statistical
inference.
I
mean
you
have
some
amount
of
data
that
you're
updating
the
model
with
and
then,
when
you're
done,
updating
the
model,
you
run
it
through
other
data
and
see
how
well
it
predicts
the
output
and
it's
just
a
matter
of
how
powerful
that
model
is.
D
D
Chained
together
before
you
get
to
the
final
output,
is
a
very
general
way
of
kind
of
putting
it
in
each
black
box
can
transform
that
problem
into
a
higher
and
higher
dimensional
alloy
space,
and
so,
as
I
add,
more
linear
layers
to
a
network,
I
can
start
mapping
it
to
non-linear
more
and
more
non-linear
and
more
higher
dimensional
problems.
Whereas
you
know
a
tree
based
method
like
xg
boost
is
essentially
just
a
big
tree
of.
D
If
statements-
and
it
has
some
interesting
capabilities,
but
it
can't
map
to
a
really
high
dimensional
space-
well,
like
facial
rack
or
the
linear
models,
all
boil
down
to
a
dot
product,
so
they
can
only
do
a
linear
mapping
at
the
end
of
the
day
and
you
have
to
have
enough
features
going
into
them
that
the
output
can
match
the
the
dot
product
of
the
your
dot
product
against
your
feature.
Set
can
actually
produce
your
mapping
that
you
need,
and
so
with
the
like
and
those
would.
D
Those
would
be
like
elastic
net
or
they'd,
be
like
basic
linear
regression,
multivar
like
not
even
multivariate,
just
basic
linear
regression,
there's
kind
of
a
whole
range
in
there,
but
the
neural
nets.
The
theory
was,
you
would
never
have
to
do
feature
engineering,
because
the
neural
net,
you
could
just
add
more
layers
and
it
would
learn
the
features
it
needed
and
my
experience
in
practice
has
been.
That
has
not
been
the
case.
All
the
time.
D
Yeah,
so
this
is
actually
getting
into
the
core
of
almost.
I
think
the
most
interesting
parts
of
machine
learning
is
you.
Can
you
can
get
a
good
answer
in
two
different
ways
of
if
you
want
to
take
a
black
box
and
produce
an
answer,
and
I
could
have
a
more
powerful
black
box
or
I
could
manipulate
my
problem
such
that
it
matches
the
black
box
closer
feature.
D
D
B
D
D
Yep
you're
trying
to
do
that
and
you're
trying
to
use.
You
are
definitely
using
your
knowledge
to
make
the
result
better,
and
so
it's
dependent
on
the
machine
learner
a
lot
as
you
get
further
and
further,
and
I
I
think
it's
also
a
matter
of
yeah.
You
find
out
little
edges
of
the
data
set
and
you
make
inferences
about
how
the
variables
must
relate,
based
off
your
mental
model
of
reality
and
kind
of
push
that
into
the
machine
learning
engine
which
can't
know
what
you
know
about
reality.
A
C
D
D
D
So
you
being
the
person
who
who
launched
the
docker
container,
it
would
work
fine,
but
when
somebody
else
tried
to
to
ssh
into
that
docker
container
docker
machine
would
just
shut
down
the
container
like
that
was
their
solution
to
something
I
don't
remember
what
it
was.
But
that's,
obviously
your
clients
are
not
really
happy
with.
B
C
D
Well,
when
think
topic
ended,
I
wanted
to
continue
my
research
into
kind
of
binding
these
numeric
worlds
to
clojure,
and
I
some
of
it
was
just
working
on
libraries
that
I
knew
were
interesting
to
think
topic,
but
are
not
meaningful
to
discussion,
they're
just
utility
stuff,
and
then
I
think
I
pulled
out
the
actual
math
portion
of
cortex,
which
was
sort
of
a
I'm
gonna.
It
was
sort
of
a
very,
I
would
say
a
poorly
done
neanderthal
as
neanderthal
is
done
right
now,
so
it
was
like
a
low-level
math
library.
D
You
could
add
two
vectors.
You
could
do
gem,
you
could
do
all
the
things
you
needed
for
cortex,
which
sounds
like
a
lot,
but
it's
really
not
in
terms
of
math,
and
it
would
work
on
cpu
and
gpu,
and
so
in
that
sense
it
was.
It
was
a
little
bit.
It
was
always
targeted
towards
really
high
performance,
which
made
it
really
difficult
to
use
in
a
lot
of
ways
and
not
useful
for
generalized
cloture
programming,
and
that
became
data
type.
D
So
the
data
type
library
is
a
graduated
progression
from
general
closure
sequence,
oriented
programming
down
the
path
towards
numeric
programming
of
the
type
represented
by
languages
like
apl
and
to
a
lesser
extent,
numpy,
and
that
form
of
programming
has
a
lot
of
advantages
and
in
terms
of
performance
and
often
in
terms
of
readability.
If
you're
doing
fundamentally
numeric
algorithms.
D
And
so
data
type
is
me
trying
to
add
this
layer.
It's
almost
like
it's
almost
more
like
it's
typed
at
the
lowest
level
and
it
it's
expecting
typed
containers.
So
I
have
a
long.
I
have
a
typed
container
of
doubles
and
each
container
has
a
count.
So
if
you
take
clojure
like
the
core
of
clojure,
I
feel,
like
is
a
sequence.
C
D
Yep,
it's
an
abstraction
of
a
array
because
I
didn't
want
to
have
an
interface
that
could
both
read
and
write,
and
I
also
wanted
a.
I
wanted
a
transformation
step
between
a
generic
object
and
a
reader,
and
so
an
array
you
can
always
read
from,
but
in
the
data
type
world
it
before
you
read
from
something
you
often
have
to
to
ask
for
the
reader
of
that
object,
because
things
can
have
a
non-trivial
translation
to
the
reader.
That
might
actually
be
time
consuming.
C
D
C
D
D
A
reader
really
consists
of
two
functions
and
one
of
those
functions
is
in
a
base
class,
a
base
interface,
so
you
can't
see
it,
but
the
main
function
is
read.
You
basically
have
two.
You
have
like,
what's
the
size
of
this
thing
and
the
read
function,
so
if
I'm
in
code,
which
might
be
a
better
way
to
actually
do
this,.
D
Absolutely.
It
returns
longs,
because
I
want
to
use
these
things
on
things
that
are
a
lot
bigger
than
four
gigabytes,
potentially
really
big
things,
so
your
size
has
to
return
along.
But
anyway,
in
this
case
we'll
say
our
size
is
five
and
we'll
just
return
the
index
and
that's
going
to
make
it
look
like
a
persistent
vector
to
closure
and
readers
implement
all
of
the
interfaces.
D
A
D
C
B
D
D
Correct
yeah,
because
data
set
and
the
the
tensor
library
and
the
dataset
library
always
kind
of
make
things
the
tensor
extension
of
the
data
type,
which
I
don't
actually
think
is
that
useful.
Maybe
I
won't
mention
that
again
for
now,
but
the
I've
been
careful
so
that
when
you
print
things
to
the
ripple,
they
are
only
taking
fixed
size
of
things
at
most
20
or
something
like
that.
F
D
F
B
D
If
I
need
it
to
be
countable,
I
make
sure
it's
countable,
but
I
stop
after
that,
so
for
persistent
vectors,
they're
they're
countable,
so
they
have
a
translation
to
reader.
That
does
nothing.
Okay
or
it
does
very
little.
I
mean
essentially
a
persistent
vector
implements
java.util.list,
so
that
makes
it
a
readable
object
and
I
might
change
it
or
I
might
just
use
it
as
a
list.
It
just
depends,
but
for
a
sequence
that
means
I
have
to
wrap
it
in
something
so
an
object.
D
D
Readers
are
meant
to
model
random
access
memory,
so
I've
moved
from
a
definition
of
computing
that
is
based
off
pure
sort
of
s,
expression,
sort
of
type
computing
to
more,
like
a
von
neumann
type
machine.
Where
there
are,
I
mean
your
computer,
your
memory.
This
is
an
interesting
thing
to
think
about.
The
ram
on
your
computer
is
untyped
and
your
operations
are
typed,
but
ram
has
a
fixed
size.
D
So
that's
just
kind
of
an
interesting
thing.
Your
assembly
instructions
are
all
typed,
but
your
ram
is
untyped
and
that's
the
basis
of
computing
that
we
use
in
a
hardware
level.
I
don't
know.
I
just
always
found
that
kind
of
interesting
thing,
but
remember
earlier
I
was
talking
about
how
to
make
things
fast.
D
B
A
A
D
D
I
think
that
is
the
most
small
graduated
step,
but
but
the
the
thing
is
is
when
not
to
rip
on
job
when
java
people
think
of
that
they
think
of
an
arraylist.
That
is
realized
in
a
concrete
sense
and
what
I'm
trying
to
do
is
is
draw
an
analogy
between
a
sequence
which
is
of
a
virtual
realization
of
a
bunch
of
data
and
a
reader,
which
is
also
a
virtual
realization
of
a
bunch
of
data.
But
the
reader
is
random
access.
C
D
I
can
oh,
and
I
also
do
a
lot
of
a
lot
of
compile
time
optimization.
So
the
first
time
you
acquire
a
data
type,
if
you
don't
aot,
there's
a
hit
there,
and
so
that's
that's
one
thing
that
does
pop
up
now
now,
but
that's
kind
of
a
side
point.
So
if
I
does,
if
I
have
a
new
reader,
that's
some
mathematical
abstraction
on
the
first
reader.
If
I
do,
I
can
do.
D
C
D
B
C
D
C
C
D
Yeah
that'll
be
the
highest
that
would
be.
My
guess
is
what
would
perform
the
best
or
and
or
use
the
least
amount
of
memory
as
we
move
to
bigger
and
bigger
data
sets,
honestly,
I
feel
like
performance
becomes
less
of
an
issue
and
the
amount
of
ram
you
use
becomes
more
because
you
know
if
you
run
out
of
ram,
and
you
start
getting.
Oh
out
of
memory
exceptions
you're
done
like
your
whole.
D
D
D
D
D
D
A
A
D
Whole
lot
going
on
here
that
maybe
maybe
we
shouldn't
or
shouldn't
talk
about,
but,
for
instance,
the
stocks
data
set,
has
a
symbol
and
there's
not
many
different
symbols
in
the
whole.
I
think
there's
five
of
them
and
what
the
data
set
library
does
underneath
the
covers
is
it
creates
a
string
table
and
then
stores
that
column
in
a
byte
array.
D
So
what
it
means
is
that
there's
five
strings
in
the
symbol,
column
and
the
actual
column
data
is
an
index
in
a
byte
is
is,
is
is
a
set
of
numbers
in
a
byte
array,
depending
on
the
order.
The
strings
came
into
the
system,
so
there's
a
string
table
which
is
a
map
of
string
to
integer
and
back
so
two
maps,
one
to
string
to
integer
one
integer
and
back
and
there's
a
byte
array
with
indexes
which
tells
you
which
string
is
supposed
to
be
in
that
index.
D
D
D
D
Yep
exactly-
and
I
think
I've
got
a
basic
first
step
for
a
lot
of
people
and
probably
the
only
step
they
ever
really
need
honestly
is
load.
The
data
set
once
do
like
a
little
bit
of
data
set
processing
on
it
and
then
convert
it
to
a
map.
Secretor
and
finish
up,
I
think,
for
a
lot
of
people.
Your
initial
step
with
the
tekkenmail
dataset
library
might
be
map
sequences
of
maps
to
datasets
and
back.
A
D
That's
a
good
question.
I
don't
I
don't
concretely
know
I
just.
I
know
that.
I
know
that
in
the
last
two
weeks
before
I
mean
honestly
before
daytime
support,
the
dataset
library
was
interesting
but
hard
to
work
with,
because
a
lot
of
the
things
we
work
with
are
log
streams
or
things
that
are
very
date
oriented.
D
So
you
had
to
convert
all
that
information
into
like
second
since
epochs,
or
something
like
that,
but
now
that
it
has
dates
and
it's
efficient
it's
kind
of
now
we
can
use
it
every
day
for
a
lot
of
our
problems
so
and
plus,
I
feel
like
now.
There's
community
support
for
it,
whereas
before
it
had
too
many
sharp
edges
and
the
people
who,
as
thomas
said
earlier
when
he
tried
it,
just
didn't
work
so
and
that's
because
there
are
just
too
many
sharp
edges
in
it.
D
D
D
C
B
D
And
so
now,
oh
and
it
did
an
integer
space,
that's
a
bug
that
should
have
worked
in
floating
point
space,
though,
because
the
first
yeah
yeah
that's
a
bug.
So
not
all
the
rough
edges
are
out,
but
any
case
that
happened
in
constant
time.
It
didn't
actually
do
math
on
the
whole
column.
It
just
created
an
equation
and
set
the
column
to
that
equation.
A
A
Okay
in
in
that
case,
I
would
just
like
to
say
a
huge
thanks
chris
for
taking
the
time
to
do
this.
It
was
really
great
to
to
learn,
learn
the
whole
thinking
behind
behind
what
you've
been
doing,
and
that
has
been
helping
me
a
lot,
because
I
have
had
cases
where
I
wanted
to
use
the
data
frame
library,
but,
but
I
haven't
had
a
conceptual
understanding
of
where
it
fits
and
understanding
the
readers
and
your
thoughts
on
how
to
keep
this
efficient
has
really
helped
with
that.
D
I
think
what
I'd
like
to
say
is
I
I
don't
want
to
oversell
anything.
It's
a
young
library.
You
know
numpy
took
15
years
before
people
even
tried,
pandas
and
then
pandas
took
another
eight
years
on
top
of
that
before
they
supported
dates
or
something
I
mean
it's
all
very
young
stuff,
so
be
patient
with
it.
You
know
and
file
issues.
I
am
not.
I
do
not
take
issues
personally.