►
Description
Matt Micene (@cleverbeard), of Red Hat Tech PM, will join us for a topic that is close to his heart. There are two aspects really: a) how to make data science easier; b) how to improve the Reproducibility of data science experiments. We both think containers are an answer.
How Red Hat Enterprise Linux (RHEL) users and admins can benefit their organizations and improve their careers by learning how to use containers, Kubernetes, and Red Hat OpenShift.
Learn more at https://red.ht/leveluphour
A
Good
morning,
good
afternoon,
good
evening,
wherever
you're
handling
from
welcome
to
another
episode
of
the
level
up
hour
here
on
openshift
tv,
I
am
chris
short
host
of
open
shift
tv
and
I'm
joined
by
two
other
people,
which
you
can't
see
right
now.
Sadly
they're
in
a
little
box
off
the
side.
So
I'm
gonna
turn
around
and
fix
that.
But
I
will
let
the
illustrious
langdon
white
introduce
himself
while
I'm
fixing
that
real,
quick.
B
So
hopefully
y'all
can
hear
me,
even
if
you
can't
see
me-
and
you
know
in
the
grand
scheme
of
things-
isn't
that
better,
so
I'm
langdon
white
we're
doing
the
level
up
hour
and
we'll
talk
a
little
bit
about
the
show
in
general
and
but
we
have
a
special
guest
today,
matt
maisini
matt.
Do
you
want
to
introduce
yourself
real,
quick.
C
C
It's
the
joy
of
not
having
an
official
job
title
since
I
started
doing
research
for
our
market
intelligence
team,
focusing
on
like
technology
things.
So
I
think
that
makes
me
a
technologist.
I
I
don't
actually
know
what
other
people
think
I
do
so
yeah.
I
look
at.
I
look
at
technologies
and
emerging
technologies
and
you
know
how
they
might
come
to
market
and
how
they
might
impact
linux
things
and
other.
C
B
B
B
A
B
A
B
C
B
C
Notes,
that's
the
easiest
way
to
to
get
that
second
set
of
slides,
because
that's
usually
what
you
want
people
to
they
want
to
be
able
to
refer
back.
To
is
what
you
said,
and
most
of
us
are
usually
pretty
good
about
that
data
being
there.
Even
if
it's
not
actually
things,
we
say
out
loud
right,
yeah.
B
B
B
B
You
know
or
like
deploying
like
your
own
next
cloud
is
what
we've
done:
the
past
few
episodes,
and
today
what
we're
going
to
talk
about
is
doing
data
science
with
containers
or
arguably
science
in
general,
because
one
of
the
things
that
you
want
in
science
is
capital
r
reproducibility,
because
you
want
your
peers
to
be
able
to
review
your
stuff.
So
much
like
you
know,
quality
assurance
in
the
software
world.
B
You
need
to
be
able
to
replicate
something
somewhere
else,
so
that
you
can
prove
that
it
wasn't
some
anomaly
of
your
environment
that
that
made
your
science
turn
out
the
way
you
wanted
it
to
or
the
way
it
did
even
so
follow
us
on
twitter,
I'm
langdon
with
a
one
and
chris
short,
is
chris
short
and
matt.
Mycenae
is
cleverbeard,
even
though
I
didn't
put
him
on
the
slide,
I
probably
should
have.
B
B
Is
it
isn't
it?
Yes,
awesome
all
right,
so
I
you
know
screwed
up
my
intro
and
did
the
intro
of
today
versus
the
intro
of
the
show
first,
but
you
know
whatever
so
today
is
containers
data
science
and
replication.
Last
week
I
hadn't
finished
the
show
notes
in
time.
So
here
are
the
show
notes
for
both
episodes.
Last
week's
episode
right
now
and
the
episode
before
that.
So
obviously,
if
you
just
go
to
episodes,
you
can
find
them
all.
I
even
updated.
That's.
B
C
C
So,
let's,
let's
start
there
right
if,
if
we've
got
an
app
we've
built
and
we're
looking
to
replicate
it
like
either
we're
typically
looking
at
like
kubernetes
rebel
cassettes
like
I
need
47
copies
of
this
particular
binary
and
it's
you
know,
configs
to
spin
up
so
that
I
can
handle
load
or
something
along
those
lines.
Or
you
know
at
the
other
end
you're
talking
about
minor
variations
that
don't
really
matter
like
if
we're
talking
about
compiling
code,
minor
differences
in
the
environment.
C
C
C
Easier
to
talk
about
state
today
than
it
was
three
years
ago
at
devconf,
because
this
wasn't
on
everyone's
talk
about
was
this
study?
Reproducible
right
right
so
typically
would
sound
like
you
go
back.
I
don't
know
2010,
probably
if
you
people
who
publish
scientific
studies,
they're,
usually
papers
right
literally
pdfs
with
you
know,
you
did
math,
you
have
tables,
you
probably
write
out.
Your
generic
formula
in
you
know
actual
mathematical
symbols
and
there's
descriptions,
maybe
of
how
they
were
used,
and
then
your
final,
like
here's,
the
fancy
table.
C
That
shows
me
my
log
graph
of
whatever
was
being
studied
and
that
sort
of
was
it
right.
If
somebody
wanted
to
reproduce
that
they
had
to
hopefully
read
all
of
your
paper
understand,
it
figure
out
your
algorithm
actually
from
your
formula
and
then
hopefully
you
may
be
lying,
didn't
put
the
data
somewhere.
That
might
be
shareable,
but
hopefully
it
wasn't
proprietary
under
some
sort
of
nda
from
a
big
pharma
company.
So,
like
there's
all
sorts
of
issues
with
like
once,
you
publish
the
paper,
how
do
you
actually
go
about
fact?
Checking
it
like?
C
A
C
At
the
time
this
executable
paper,
executable
research
paper,
starts
popping
up
in
in
various
different
places,
and
the
idea
is
we
have
a
lot
of
science,
that's
backed
by
a
lot
of
math
that
carries
a
lot
of
data,
the
best
way
to
get
to
a
state
where
it's
easy
to
peer
review.
These
things
is
publish
it
all
together
in
one
format,
so
that
somebody
can
walk
along
and
look
at
your
code
and
go.
C
You
know
what
you
that
algorithm
looks
like
it
replicates
this
formula,
but
it
doesn't
because
you
did
this
wrong
or
because
we
know
that
there's
a
flaw
here
and
also
your
data
is
there.
So
you
can
just
like,
oh
that
table
click
yep
that
table
actually
that
code.
With
that
data
reproduces
that
table.
It
wasn't.
You
know
a
publishing
error.
B
C
B
Let's
say
the
data
is,
you
know
available
like
it's
not
proprietary
or
whatever?
Is
there
any
sort
of
best
practices
emerging
about
sharing
that
data
without
you
know,
basically
mailing
me
a
hard
drive
because
it's
you
know,
terabytes
of
content
and
because
that's
that's
often
a
problem
both
in
this
kind
of
science,
replication
area,
but
also
in
kind
of
just
general
software
development,
especially
data
science,.
B
C
The
other
side
of
the
issue
right
is:
if
we
start
talking
about
containers
and
go
back
to
that
example,
most
folks
are
only
they
only
care
about.
You
know
the
binary
and
the
config
files
right.
It's
that
push
to.
I
want
my
six
meg
go
container.
That
does
one
thing
and
that's
all
it
does
the
ones
built
around
data
science
and
other
science
support
without
the
data
right,
we're
talking
about
fairly
complex
python
environments,
our
environments,
just
from
getting
the
libraries
in
place
right.
C
So
there's
a
lot
of
different
ways
that
people
have
gone
about
like
okay,
we'll
just
jam
it
in
the
we'll
jam
it
into
the
container.
That
way,
we
know
it's
versioned
and
replicated
right
because
that's
that's
an
important
part.
Other
folks
have
looked
at
that
large
file
support
in
like
online
version
control
systems
like
using
lfa
and
github,
or
something
along
those
lines
right
right,
there's
even
specialty
sorts
of
hubs.
C
If
you
will
that,
like
jupiter
hub
and
there's
shoe
others
that
I
cannot
think
of
the
name
of
right
now
that
are
designed
to
sort
of
allow
you
to
publish
your
things
there
and
then
let
other
people
pull
it
down,
binder
sort
of
does
that
with
with
jupyter
notebooks
and
data
and
then
allows
people
to
to
get
access
to
them.
So
I
don't
think
that
one's
solved
yet
because
it's
the
typical
data
issue
right
the
price
problem
with
any
data
like
no,
it
doesn't
matter.
A
B
A
B
And
it
does
work
a
lot
better
if
you
can
just
set
that
petabyte
of
data
somewhere
right
and
then
move
the
compute
to
it
and
that's,
I
think,
a
really
kind
of
much
more
interesting
model
and
I'd
really
like
to
see
you
know
kind
of
software
development
kind
of
adopt
that
in
general,
but
so
kind
of
going
back
to
you,
know
kind
of
this
replication
in
the
science
space.
You
know.
Typically,
scientists
are
not
programmers
or
not.
B
You
know
not
trained
as
programmers,
so
is
there
infrastructure
in
place
that
helps
them
in
a
sense
or
provides
frameworks,
or
you
know
whatever
I
mean
r
basically
exists
in
some
senses
because
of
science,
but
that's
that's
pretty
that's
pretty
low
level
for
for
the
kind
of
frameworks
or
support.
We
normally
look
for
to
help
non-programmers
program
in
a
sense
right.
B
C
I
went
to
pick
one
up
and
move
it
and
I'm
using
pip
end,
but
I've
got
a
piplock
file
and
I
dump
it
to
a
requirements.txt
and
I
pull
it
over
and
like
pip
install
like
that,
should
work
right.
It's
python
doesn't
work
somehow
something
got
version
locked
to
a
version
that
doesn't
work
with
it.
I'm
using
python36.
C
C
C
There's
so
many
different
options
and
then
you
layer
on
well,
do
I
use
a
notebook
and
then
which
notebook
and
then
do
I
use
r
and
which,
like
the
distribution,
issues
of
the
languages
alone
yeah,
so
there
are
folks
out
there
that
are
looking
at
like
okay,
like
jupiter,
does
put
out.
Containers
that
have
this
is
the
kitchen
sink
for
data
science
yeah
and
it's
good.
It
works
yeah.
B
I've,
actually,
I
think
I've
actually
used
that
before
yeah
the
one
of
the
things
actually
that's
super
useful
about
the
python
community
in
particular,
is
that
they're
friendly
in
a
sense
with
their
downstream
distributions
as
much
as
possible.
So,
unlike
a
lot
of
the
languages,
you
know
when
they
are
kind
of
designing
the
packaging
format,
they're,
actually
intentionally
choosing
making
choices
that
make
it
easier
to
distribute
in
a
linux
distribution.
B
So
you
know,
unlike
a
lot
of
languages
like
this,
is
one
of
the
things
that
people
don't
really
understand
is
like.
Why
can't
we
get
an
rpm
say
from
you
know
a
python
module
or
a
ruby
module
or
whatever,
and
a
lot
of
it
has
to
do
with,
like
you
know,
kind
of
just
automatically,
I
mean,
but
a
lot
of
it
actually
has
to
do
with
the
metadata
being
missing
and
stupid
stuff,
like
the
license
on
that
particular
module.
B
B
Well,
that's
the
whole
problem
is
that
they're,
not
a
kin
they're,
two
different
solutions
to
a
similar
or
the
same
problem
that
do
not
interact
with
each
other,
and
this
is
actually
something
that's
been
looked
at
for
a
long
time
in
dnf,
for
example,
can
we
make
dnf
essentially
a
wrapper
around
all
the
other
packaging
tools
right?
So
apparently
the
only
way
that
you
can
get
your
like.
True,
like
my
new
programming,
language,
creds
or
whatever,
is
write
a
package
manager.
So
you
know
so
every
single
language
has
a
different
one.
B
Lots
of
you
know
now
we
have
different
ones
on
different
platforms
as
well.
Like
I
mean,
there's
a
there's
at
least
two
or
three
now
for
windows,
there's
at
least
two
for
mac
os
as
well.
You
know,
then
we
have
all
the
bsd
ones
and
we
have
our
bsd
has
ones,
and
then
we
have
you
know
debian,
and
we
have
you
know,
kind
of
the
rel
and
it's
it's
friends
with
rpm.
B
None
of
these
work
together.
None
of
these
model
work
in
the
same
way,
they're
dependency
management,
they're
dependency
solving,
are
all
different,
even
though
this
is
actually
going
back
to
science.
This
is
actually
a
well-understood
problem
that
has
been
solved
with
math
and
proved,
but
instead
of
just
saying,
hey.
A
C
Well,
yeah,
and
that
was
actually
the
solution
to
my
problem,
was
I
I
was
actually
switching
between
two
different
python
package
managers
really
because
I
was
using
pip
ends
in
one
environment
which
not
only
deals
with
pip,
but
also
sets
up
the
virtual
environment,
yep
and
raw
pip
in
the
container,
because
I
was
like
well,
I
don't
need
pip-band
for
all
this.
It's
just
a
container
right
and
somehow
in
the
pip
file
lock
for
pip
end.
C
That's
where
it
inserted
this
weird
version
string,
which
I
don't
even
I
haven't
gone
back
to
look
at
the
the
that
actual
virtual
environment.
But
the
solution
was,
let
pip
be
smarter
than
me
right
and
do
the
right
thing
and
now
it
works.
I
was
like
okay,
we'll
just
we'll
just
remove
the
version
requirement
that
locked
on
that
file
in
the
requirements
text
and
lo
and
behold.
Pip
is
designed
to
solve
for
this
and
some
sort
correctly,
and
I
have
a
working
environment
in
the
container.
So
yeah.
B
Well,
I
mean-
and
that's
that's
often
the
big
problem
right
as
long
as
like,
if
you're
not
mixing
and
matching
it
often
works
so
you'll
have
python
developers,
for
example,
who
will
not
install
anything
coming
from
anywhere
besides
upstream
pip.
The
downside
to
that
is
that,
if
you
is
that
you
know
kind
of
going
back
to
that
version,
locking
problem
you
know
there's.
This
is
one
of
those
really
arguable
positions.
It's
like
and
one
of
the
things
actually,
we
were
trying
to
solve.
B
Actually
with
modularity
and
app
streams
is
I
want
guarantees
that
the
version
I
worked
with
will
still
work.
So
in
other
words,
so
I
go
inversion
lock
everything
you
can
do
this
in
basically
every
package
manager,
somehow
you
can
do
it
with
you
know
dnf.
You
can
do
with
apt-get,
you
know,
but
the
problem
with
version
locking
is
you
get
exactly
what
you
version
lock
to
well,
that
means
no
security
updates.
B
B
I
think
it
was
yeah
and
if
you
did
that,
wordpress
and
drupal
wouldn't
run
because
they
were,
they
only
worked
with
5.2
with
them,
and
so
you
also
have
different
like
kind
of
rules
depending
on
the
language,
depending
on
the
environment,
about
what
things
will
be
backwards,
compatible
et
cetera,
et
cetera.
So
that
was
a
bit
of
a
digression
into
one
of
my
pet
peeves
of
package
managers,
but.
C
It's
exactly
that
level
of
version,
locking
that
we
need
to
make
sure
we
have
available
so
that,
oh,
I
don't
know
if
I'm
trying
to
reproduce
an
experiment
that
chris
built
and
he
likes
the
open,
blas
linear,
linear,
algebra,
set
up
for
for
scipy
and
numpy.
But
I
prefer
atlas
that
when
I
take
his
code
and
his
data
right,
we
can't
just
stop
at
code
and
data
because
we
don't
want
version,
mismatches
or
developer
choice
in
and
what's
my
implementation
to
actually
impact
that
outcome
right.
C
So
these
things
really
do
need
to
be
like
super
specifically
version
locked
and
then
to
complicate
it
further.
We
go
back
to
like
our
title
of
data
science,
like
let's
look
at
machine
learning
and
artificial
intelligence
packages
right
that
the
data
science,
one
that
we
just
talked
about
from
jupiter
hub.
That's
I'm
going
to
call
that,
like
traditional
data
science,
that's
it's
mostly
math
libraries
and
scikit-learn
and
those
sorts
of
things.
It's
got
some
machine
learning,
but
it's
not
pi
flow
or
tensorflow,
or
some
of
these
larger
abstractions
around
deep
learning
and
the
rest.
C
So
once
you
start
trying
to
coordinate
15
different
sets
of
requirements
and
libraries
and
modules
and
they're
accompanying
data
sets
and
corpuses
and
lexicons,
and
like
not
the
experiment,
data
just
the
things,
it
needs
to
run.
To
know
that
I
shouldn't
count
the
word
the
every
time
it
shows
up
as
significant.
C
C
They
just
need
to
push
the
button,
get
their
environment
and
then
attach
it
to
some
data
and
start
going
right.
So
there's
like
the
one
end
is
there's
one
off
onesie
twosie
people
building,
here's
a
catalog
of
containers,
of
six
different
options
and
then
there's
the
here's.
This
all-singing,
all-dancing,
orchestrated
massive
beast
of
an
operator
that
can
do
magic
for
you.
B
Right
right,
yeah
interesting,
I
actually
didn't
realize
that
was
one
of
the
kind
of
goals
of
open
data
hub,
but
that
is
that
is
definitely
interesting.
One
of
the
things
I
think
you
know
kind
of
my
own
experience
that
we're
we're
also
seeing
is
universities
kind
of
starting
to
recognize
this
problem
and
collaborating
more
in
order
to
with
other
universities
in
order
to
try
to
resolve
some
of
these
challenges.
B
B
Oh,
it's
actually
yeah,
so
collaboration
of,
like
I
think
it's
like
boston
university
mit,
I
think
harvard's
involved.
You
know
like
a
whole
bunch
of
universities
and
they
basically
put
a
whole
bunch
of
hardware
in
a
data
center
and
they're
starting
there
and
what's
interesting
is
the
project
itself
is
its
own
science,
as
well
as
supporting
other
people's
science.
B
So
what
they're
trying
to
do
is
make
an
open
stack
or
some
sort
of
basically
public
or
cloud
that
you
can
share
hardware
into
and
get
hardware
out
of
when
you
need
it,
and
so
you
know,
the
problem
with
your
typical
cloud
right
is
that
you
have
to
go,
buy
it
kind
of
on
demand
and
if
you
have
existing
assets,
you
know
especially,
you
know.
Big
iron
assets
there's
no
way
to
kind
of
have
them
participate
or
when
they're
slow
share
them
with
others.
A
C
Yeah,
it's
not
so
much
that
the
math
varies.
It's
that
the
implementation
of
the
algorithm.
So
there's
we
get
into
the
into
the
nitpicking
details
between
the
formula
and
the
algorithm
right.
The
formula
is
the
math
written
on
the
whiteboard,
and
that
is
what
it
is,
and
the
algorithm
is
how
you
implement
it
and
so
implementation
details,
depending
on
what
you're
doing
and
for
data
science,
it's
a
little
bit
different
because
there
we're
looking
for
things
like
when
we
talk
about
some
of
these
differences.
C
C
So
for
data
science,
it's
it's
a
little
more,
not
a!
I
read
your
paper,
I
recoded
your
stuff
and
I
got
a
different
answer.
So
much
as
it
is,
I
took
your
model.
I
took
your
training
data
and
it
wasn't
as
accurate
as
you
claimed,
or
it
took
47
times
as
long.
So
therefore,
that's
that's
not
good,
because
that's
not
what
you
said.
You
did
in
your
in
your
paper.
B
Well-
and
one
of
the
points
I
would
make
too
is
like
michelle
orlondi:
maybe
it
brings
up
in
the
chat
when
we
do
math
on
computers.
We
make
trade-offs,
especially
when
those
numbers
are
really
really
big
and
rounding
errors,
and
so
their
example
right
is
four
followed
by
a
whole
bunch
of
zeros
and
then
one
you
know,
rounding
errors
are
a
big
deal
where
you
take.
The
rounding
error.
Pane
also
has
an
impact.
The
other
another
one
that
comes
up,
which
is
kind
of
weirdly
interesting,
is
most
random.
B
Number
generators
on
a
computer
are
not
actually
random,
and
so
when,
if
you
use
a
you
know,
what
do
they
call
pseudorandom
totally
blank
if
you
steal
random
inputs
that
can
actually
have
an
effect
on
your
outcome
because
you
actually
are
because
of
the
lack
of
randomness.
You
actually
get
some
variation
in
the
result,
so
there's
so
like.
If
we
were
doing
all
this
work
by
hand,
it
would
be
very
replicable.
C
It
was
at
the
rounding
point
where
in
python
two
it
would
round
down
and
then
python
3
it
would
round
up
so
yeah
and
again,
I'm
blaming
python.
But
I
don't
actually
remember
that
it
was
python,
but
it
was.
It
was
a
pretty
common
language.
It
was
on
a
major
version
boundary
like
that
yeah
I
mean,
even
even
on
the
short
end,
like
simply
a
programmer's
like
yeah.
We
now
have
floats
so
I
can
round
longer,
like
I've
got
better
access
to
64-bit
hardware.
C
B
Yeah
yeah
yeah,
I
I
don't
like
math
math
on
computers
is,
is
really
really
hard
if
you've
never
done
kind
of
real
math.
When
you're
doing
two
plus
two
equals
four,
you
know
it's.
It's
not
too
big
a
deal.
It's
funny
right
because
you
know
when
I
was
in
college
and
then
shortly
thereafter.
I
worked
in
r
d.
You
know
I
did
you
know
the
math
in
college
and
I
did
a
lot
of
that
stuff
in
computer
science,
land
right
and
then
I
went
and
worked
in
an
r
d.
B
B
You
know
you
do
all
this
kind
of
really
high
level
math
and
how
to
implement
that.
Even
in
software
in
computer
science,
you
know
classes
or
whatever.
But
if
you
go
on
to
be
a
software
developer,
you
actually
don't,
generally
speaking,
have
to
do
very
much
of
this.
So
the
it's
kind
of
roaring
back
with
the
push.
A
I
mean
the
we've
been
having
a
general
conversation
about.
You
know,
bioinformatics
genomics
data
science,
that
kind
of
thing
we
talked
about
protein
folding
a
little
bit
and
the
recent
advancement
there
kind
of
thing,
which
is
awesome
for
folks
that
haven't
heard
was
deepmind,
did,
has
made
some
quote
unprecedented
progress
on
protein
folding,
and
it's
like
when
you
think
about
protein
folding
and
how
that
can
cure
cancer
or
genetic
diseases.
It's
you
know
having
any
kind
of
advancement.
C
B
A
B
When
I
was
doing
math,
I
did
a
lot
of
like
math.
That
was
not
straight
up
computer
science
math,
because
I
thought
it
was
fun
and
one
of
them
one
of
the
spaces.
Is
this
space
called
tilings,
which
is
where,
like
what
kind
of
shapes?
Can
you
put
on
a
plane
that
will
cover
the
plane
completely
yeah?
Well
right
exactly-
and
this
has
like
this-
is
a
problem
space
for
that's
really
interesting
when
you're
trying
to
like
coat
things
in
you
know
material.
B
For
example,
you
know
like
the
shuttle
with
tiles,
for
you
know,
heat
dissipation,
but
this
woman,
who
was
like
you,
know,
a
housewife.
You
know
in
her
like
30s
or
40s,
or
something
like
that
discovered
something
like
seven
of
the
12
just
doing
them
on
her
kitchen
table
when,
in
her
like
free
time,.
C
C
C
There's
a
lot
of
people
out
there
that
you
know
as
a
newbie
data
scientists,
there's
a
lot
of
other,
very
accomplished
folks
who
are
doing
these
challenges
to
to
to
put
these
to
do
this
work
and
it's
it's
a
very
much
a
it's
a
competition,
but
at
the
same
time
it's
hey.
I
put
this
out
there,
and
this
is
my
code
like
what
do
you
think
and
you'll
get
comments
from
very,
very
experienced
folks
saying?
Well,
this
is
an
interesting
approach,
but
have
you
looked
at
xyz
or
this
is
off?
C
Yeah,
I
know
red
hat
has
actually
put
out
a
couple
of
challenges
around,
I
think
was
one
of
the
ones
I
looked
at
was
like
sentiment,
analysis
and
some
you
know
in
some
tweets,
so
some
of
them
are
are
just
community
driven
challenges.
Others
are
actually
businesses
putting
out
like
bounties
on
hey.
We
need
an
approach
to
do
this
sort
of
analysis
on
this
data
and
we
don't
have
the
resources
that
sort
of
thing.
C
So
it's
a
very
interesting
place
to
go,
lurk
and
learn,
and
then
the
other
one
is
it's
sort
of
the
the
idea
that
you
find
that
you
brought
up
about
compute
locality.
I
think
that's
sort
of
an
old
school
thing
in
the
high
performance
compute
environment
that
I
used
to
deal
with
from
a
customer
standpoint
and
a
long
long
time
ago,
in
a
galaxy,
far
far
away
far
far
away,
and
that
was
one
of
the
like.
C
That
would
be
amazing
if
you
could
ever
figure
out
how
to
do
that,
because
the
traditional
model
of
install
and
application
didn't
work
correctly,
but
in
a
modern
world
you
know
fire
up
a
couple
of
containers
on
a
data
lake
like
this
is
a
much
different
sort
of
sort
of
proposition
and
to
be
able
to
actually
do
those
sorts
of
things
in
a
controlled
fashion
without
completely
destroying
data
and
nodes
and
all
that
sort
of
management
right.
So
right,
yeah.
B
A
B
A
data
mart
is
yeah
or,
or
you
know
it's
one
of
those
things
like
lots
of
things
in
software.
There's
87
different
definitions,
depending
on
who
you
ask,
but,
generally
speaking,
a
data
mark
is
considered
a
you
know,
sometimes
a
right
through
cache,
but
you
know
often
just
a
read-only
cache
of
a
lake
so
because
the
way
data
is
stored
in
a
data
lake
generally
speaking,
is
not
a
good
design
for
that
data
for
the
applications
that
would
consume
that
data.
B
B
C
C
Three
years
is
a
long
time
in
the
container
space.
The
the
idea,
though,
I
think
is,
is
along
the
lines
of
what
we've
been
talking
about.
Is
how
do
you?
How
do
you
build
for
for
the
people
who
who
are
the
the
scientists
and
doing
the
experiments?
How
do
you
build
a
set
of
tools
that
are
a
set
of
toolboxes
that
they
can
use
consistently
without
having
to
be
experts
themselves
in
packaging
and
dependency
management
right?
C
And
it's
the
same
sort
of
challenge
that
you
know
the
jupiter
hub
images
that
we
talked
about?
Are
there
I'm
currently
poking
around
with
vs
code,
because
they
recently
drop
some
support
for
podman
into
their
container
remote
execution,
plugin
module
extension,
that's
what
they
call
them,
they
call
them
extensions.
You
pick
a
word.
C
So
I
think
you
know
as
an
example
as
an
exemplar
of
like
what
the
challenge
space
is.
It's
it's
still
a
valid
repo
as
an
actual
workable
example,
not
so
much
anymore,
but
it
it
essentially
set
up
a
the
literature.
The
linear
algebra
example
that
I
brought
up
earlier
is
that
you
know
somebody
wants
to
see
how
the
different
implementations
of
that
are
available
for
python.
For
numpy
to
do
linear,
algebra.
You
can
easily
set
up
two
different.
C
A
B
So
it
really
depends
on
the
system,
as
I
kind
of
respond
to
the
chat,
it
has
a
lot
to
do
with
the
design
or
the
goal
of
the
of
the
tool
or
the
the
implementation
rather
than
the
product
set.
You
would
choose
the
product
based
on
the
goal.
So
you
know
redis,
is
you
know,
name
value
pairs,
so
very,
very
good
if
you
have
very
disparate
data,
but
it's
it's
kind
of
in
that
construct,
so
you
know
think
like
environment
variables,
but
lots
and
lots
of
them.
B
So
if
you're,
if
you
have
that
kind
of
data,
then
redis
is
a
good
choice.
If
you
know,
if
you
take
what
we
used
to
call
document
stores
that
are
often
referred
to
as
nosql
databases
now
like
mongodb,
for
example,
mongodb
is
really
really
good
if
you
need
a
big
block
of
data
based
on
a
small
piece
of
data,
so
especially
if
that
block
of
data,
the
structure
of
it
is
not
that
important.
B
So
this
a
great
example-
I
was
actually
talking
about
this
with
my
brother,
who
works
for
a
museum
museum
websites,
are
good
or
can
be
a
good
place
for
something
like
a
because
you
say
hey,
I
want
art,
exhibit
x
right
and
you
get
back
this
whole
block
of
images
of
it.
You
know
curator's
descriptions
why
it's
important.
You
know
the
history
of
it
all
the
stuff,
the
individual
pieces.
The
structure
of
that
data
is
not
that
important
you're
not
going
to
have
duplications
across
like
the
curators
comments
across
pieces
of
art
right.
B
B
When
you
don't
know
what
the
structure
is
going
to
be
or
where
the
structure
is
going
to
fall
and
then
basically
migrate
data
into
a
sql
database.
As
you
understand
the
structure
of
what
you're
actually
looking
at,
when
you
talk
about
a
data
mart,
I
would
say
in
my
experience:
they
are
traditionally
sql
databases,
because
the
the
lake
is
well
structured
generally
speaking,
but
you
want
to
do
some
manipulation.
So,
for
example,
you
know
most,
you
know
a
true.
This
goes
back
to
going
to
computer
science.
B
When
you
talk
about
sql
one
of
the
things
they
tell,
you
is
always
put
it
in
fourth,
normal
form
which
basically
reduces
the
duplication
of
data
across
your
data
set
by
as
much
as
humanly
possible.
But
that
means
you
have
a
ton
of
different
tables.
So,
for
example,
your
customer
record
might
be
made
up
of
like
57
different
tables.
B
All
with
you
know,
five
columns
each
because
that's
the
way
the
data
is
structured
and
that's
the
way
you
get
the
most
deduplication,
but
for
a
front-end
application
who
needs
to
display
the
customer
profile,
that's
very
expensive
to
go
and
query
that
entire
57
set
of
tables.
For
that
one
record
really.
So
what
you
do
in
a
data
mart
is
actually
create
like
a
view
right
in
the
sql
sense,
except
maybe
not
literally
implemented
this
way.
B
Usually
it's
it's
a
there's,
an
old
term
for
it's
like
an
implemented
view
or
something,
but
basically
where
you've
already
crashed
the
57
tables
together.
Now
you
have
a
customer
record
in
one
table,
but
it
has
lots
of
duplicated
data,
which
is
fine
because
you're
only
pulling
out
a
piece
of
your
lake,
so
you
know
you're
talking,
you
know
a
gig
right.
Instead
of
you
know
five
petabytes,
and
so
that's
that's
kind
of
what
the
idea
behind
marting
is.
B
I
would
say
you
know,
and
then
basically
it's
like
you
know
going
back
to
the
museum
example.
Your
lake
might
be
all
the
different
art
that
you
have.
You
can
actually
think
of
a
physical
museum,
usually
something
like
20
or
less
of
the
art
is
actually
on
display.
Most
of
it
is
in
their
basement
and
they
rotate
it
through
because
they
don't
have
enough
physical
space
to
show
it
all
off
at
once
same
kind
of
idea.
B
So,
like
you
know,
your
lake
might
be
have
all
of
the
things
that
you've
ever
heard
of
any
special
exhibits.
You've
ever
had
things
that
were
on
loan
from
all
these
other
places.
You
know
et
cetera,
et
cetera,
whereas
your
mar
in
front
of
that,
even
if
it's
the
mart
is
now
mongodb,
let's
say,
is
just
the
stuff
that
is
currently
things
that
you're
showing
on
the
website.
So
that's
another
example,
so
that
was,
I
don't
know.
Does
that
answer
that
question
that
was
kind
of
long.
B
It's
it's
a
space
where,
if
you
play
in
it
it
comes
like.
Oh,
like
it
gets
really
clear,
very
fast.
Like
you
know,
like
the
first
time,
I
tried
to
pull
a
query
against
a
57
table
customer
record.
I
discovered
very
very
fast
that
I
wanted
my
response
times
to
be
not
minutes
right
so
yeah,
so
it
makes
it
much
more
much
easier
once
you
actually
participate
in
the
space
here,
because
it's
like
10
of
and
say.
B
B
A
B
All
right,
so
sweet,
sweet
internet
points
for
anybody
who's
new
to
the
show
we
award
points
for
coming
to
the
show
or
for
doing
a
pr
on
our
repo.
You
know-
or
you
know,
filing
issues,
there's
actually
like
there's
an
activities
page
on
the
repo
itself.
Let's
see,
can
I
copy
pieces
from
anywhere
now
that
I'm
not
allowed
to
actually
open
the
slides
independently
of
the
presentation
but
netherlands
hackum
2900
points
narendev
with
2400
points.
A
B
B
And
so
yeah,
so
you
know
collect
the
internet
points.
We
think
they're
fun
and
entertaining
way
to
kind
of
say
you
know
hey.
I
came
to
the
show
I'm
engaged
with
the
show
someday.
Maybe
we
will
have
prizes
for
the
internet
points
at
this
point.
We
just
like
to
keep
them
away
and
we
like
to
say:
hey
thanks,
netherland
hackham
thanks
narendev
thanks
noah
thanks
joe
fuzz
thanks
jp
dade
for
coming
to
the
show,
and
we
really
appreciate
your
questions.
B
A
B
B
A
B
C
I
mean
you
know:
internet
points
are
always
good
and
hoarding
is
a
is
a
very
valid
and
and
useful
thing
to
do,
especially
in
especially
in
the
dark
times
of
2020.
You
know,
I
think,
grabbing
as
many
as
you
can
and
keeping
them
for
for
future,
something
like
a
squirrel,
it's
winter
we
should.
We
should
do
that.
Yeah.
Definitely.
B
C
B
C
B
A
B
We
already,
as
we've
talked
about
on
the
show
before
we
have
a
cool,
a
bunch
of
cool
things
like
the
learn.openshift.com
or
try.openshift.com,
which
kind
of
give
you
more
of
a
guided
experience.
You
can
also
contact
a
like
whoever
your
red
hat
representative
is,
and
they
have
even
more
access
to
kind
of
guided
content
if
you
want
to
play
around
with
openshift,
but
the
we
just
launched
this.
B
This
developer
sandbox,
which
I
think
is
super
cool
and
you
should
go
check
it
out
and
because
matt's
here
and
is
part
of
rel.
We
will
also
pitch
the
developer
program,
which
gives
you
you
know,
free
or
sorry,
no
cost
access
to
rel.
For
many
many
use
cases,
one
of
the
things
that
we're
actually
working
on
is
trying
to
update
the
like
fac
kind
of
not
the
t's
and
c's.
The
t's
and
c's
are
pretty
good.
B
The
problem
is
people
read
the
terms
and
conditions
around
the
usage
of
rel
in
the
dock,
and
they
think
that
there's
lots
of
scenarios
that
they
can't
use
it,
which
are
not
actually
accurate.
They,
you
can
actually
use
it
in
a
ridiculously
large
number
of
ways
and
still
be
licensed
compliant,
so
some
free
and
like
they
said,
starter
and
free
options.
I
think-
and
so
I
just
wanted
to
kind
of
share
that,
because
I
think
the
dev
summer
box
literally
went
live
like
this
week.
C
B
C
A
B
Yeah
well,
so
I
basically
want
to
do
some
kind
of
examples
of
why
it
probably
doesn't
matter.
So
that's
so
I'm
going
to
try
to
prep
that
for
the
next
show,
if
I
don't
finish
prepping
it
for
the
next
show,
then
we
will
be
doing
something
else,
but
so
watch
for
you
know
kind
of
my
twitter
feed,
chris's,
twitter
feed,
the
openshift
twitter
feed,
the
red
hat
twitter
feed.
B
A
B
B
B
A
A
For
cost
and
performance,
optimization,
that
is
with
profit
stores,
the
name
of
that
company
that'll,
be
joining
us
for
the
openshift
commons
briefing.
Then
after
yeah
1400
eastern
1900
utc
we've
got
red
hat
enterprise.
Linux
presents
security,
so
that'll
be
a
good
show
to
kind
of
finish
the
day.
Talking
about
the
all-important
thing,
security
so.