►
From YouTube: Tech Talk | Predictive Maintenance (PdM) on IoT Data for Early Fault Detection w/ Delta Lake
Description
Join us for an online tech talk on Delta Lake. Tech talks include a technical presentation with slides and a demo, with time for Q&A at the end.
Predictive Maintenance (PdM) is different from other routine or time-based maintenance approaches as it combines various sensor readings and sophisticated analytics on thousands of logged events in near real time and promises several fold improvements in cost savings because tasks are performed only when warranted. The top industries leading the IoT revolution include manufacturing, transportation, utilities, healthcare, consumer electronics & cars. The global market size for this is expected to grow at a CAGR of 28%. PdM plays a key role in Industry 4.0 to help corporations not only reduce unplanned downtimes, but also improve productivity and safety. The collaborative Data and Analytics platform from Databricks is a great technology fit to facilitate these use cases by providing a single unified platform to ingest the sensor data, perform the necessary transformations and exploration, run ML and generate valuable insights.
Agenda: 10AM PDT - 11AM PDT (GMT-8)
10:00AM - 10:50AM - Tech Talk
10:50AM - 11:00AM - Q&A
To join the live chat, check out the meetup page: https://www.meetup.com/data-ai-online/events/270166033/ Databricks is proud to announce that Gartner has named us a Leader in both the 2021 Magic Quadrant for Cloud Database Management Systems and the 2021 Magic Quadrant for Data Science and Machine Learning Platforms. Download the reports here. https://databricks.com/databricks-named-leader-by-gartner
A
Thanks
everyone
for
joining
us:
this
is
our
online
Tech
Talk
predictive
maintenance
on
IOT
data
for
early
fault
detection,
with
Delta
Lake.
Just
a
few
housekeeping
notes,
chat.
We're
kind
of
using
that.
If
there's
any
you
know,
audio
issue
is
anything
I
can
help
with
I'll
drop
some
links
in
there
and
then
the
Q&A
is
gonna,
be
where
you're
gonna
want
to
submit
your
questions.
So
hopefully
we'll
have
some
time
at
the
end,
and
our
speakers
will
will
answer
some
questions
and
then
subscribe
today.
A
So
quick
call-out
for
our
youtube,
so
it's
linked
there
and
then
I'll
drop
the
link
in
the
chat
in
just
a
moment,
but
we
we
record
all
of
our
Tech
Talks
and
we
post
them
to
YouTube.
So
that's
the
link
to
access
our
playlists
of
one
of
our
dev
rel
managed
online
on
my
meetups
Tech
Talks
and
workshops
that
we
run
so
you're
you're
free
to
revisit
the
content
at
a
later
date.
So
we'd
love
for
you
to
subscribe
and
check
it
out
there
and
then
so.
A
B
C
Hi
there
thanks
very
much.
My
name
is
Danny
Lee
I'm,
a
developer
advocate
here
at
data
bricks,
I'm
working
with
Apache
sparks
since
0.6,
so
I
guess,
I
told
you
a
little
context,
but
how
long
I've
been
doing
this
for
Malaya
was
a
senior
director
of
data
science
engineering,
it
cooker
and
I'm.
Also
a
former
Microsoft
involved
in
projects
like
cosmos,
dB,
the
earlier
hdinsight
of
the
precursor
to
HT
insight
called
project
isotope
and
yeah.
There
we
go
so
that's
a
little
bit
about
ourselves.
C
So
hey
before
we
dive
into
him,
we're
actually
gonna
be
shifting.
If
you've
been
watching
some
of
our
sessions
from
the
past
you'll
notice
that
typically,
we
always
go
with
slides
and
so
don't
worry,
we
still
have
slides.
So
if
any
of
you
are
folks
our
slide
junk
junkies,
we
still
have
that,
but
indeed
it
has
really
great
mind
master
that
actually
allows
us
to
talk
about
some
of
the
concepts
that
we're
going
to
go
through
today.
B
So
today
we
are
going
to
talk
about
how
Delta
enables
predictive
maintenance
use
cases
using
IOT
data,
and
so
why
is
this
important?
So
we
have.
We
want
to
increase
the
reliability
and
extend
the
useful
life
of
all
these
expensive
equipment
and,
at
the
end
of
the
day,
it
translates
directly
into
cost
savings.
And
how
are
we
going
to
do
it?
Just
like
a
human
has
got
some
vital
as
parameters
you
know
said.
A
machine
also
has
got
some
parameters
as
well.
So
what
are
examples
of
such
sensors?
You
have
temperature
variations.
B
You've
got
a
vibrational
analysis.
You
have
got
oil
viscosity.
All
of
these.
If
we
were
to
analyze,
would
give
us
indications
of
failure.
You
can
diagnosis,
failure
from
there.
You
can
classify
failures
into
different
categories
and,
at
the
end
of
the
day,
you
want
to
be
able
to
recommend
relevant
maintenance
actions
at
the
time
in
a
timely
manner.
So
a
key
enabler
of
PDM,
that's
short
for
predictive
maintenance,
is
continuous
monitoring
and
that
why
do
we
need
that?
B
Because
it
establishes
baselines
for
what
is
a
normal
running
condition,
and
there
have
been
a
lot
of
idea
in
the
IOT
section.
There
are
lots
of
industries
that
have
contributed
to
it
and
so
now
that
there
is
easier
integration
with
the
sensor
data
allowing
the
health
of
the
equipment,
along
with
this
lifecycle,
to
be
transmitted
to
systems
where
we
can
start
to
baseline
and
do
this
analysis
again,
there
are
parallel
advances
in
many
fields,
including
advances
in,
say,
sensing
technologies,
better
connectivity
options
and
breakthroughs,
and
ml
and
AI.
B
One
reason
why
this
is
different
from
traditional
or
routine
time-based
maintenance
is
because
the
repairs
are
done
only
when
it
is
truly
warranted,
and
you
would
ask
what
are
the
markets
where
this
is
particularly
useful?
These
are
the
top
verticals
leading
the
IOT
and
predictive
maintenance
error.
So
there
is
a
manufacturer
right
on
top,
and
then
we
have
transportation.
We
have
energy
and
logistics
wick,
we
have
aerospace,
but
there's
no
reason
why
it
should
be
restricted
just
to
these
industries.
B
These
are
the
verticals
leading
the
sector,
but
all
other
industries
can
benefit
from
this
as
well,
and
why
should
we
care
about
it?
Let's
look
at
what
is
the
growth
projected
for
these
use
cases?
There
is
a
compounded
annual
growth
rate
of
more
than
39
percent
for
the
US
market
itself
and
a
little
over
28
percent
for
global
market.
If
you
were
to
take
it
and
that's
huge,
we
definitely
want
to
understand
where
this
industry
is
going
and
where
we
can
make
better
utilization
of
resources
and
technology.
B
As
Dan
mentioned
today,
we
are
going
to
talk
about
just
the
healthcare
sector,
but
everything
that
we
are
talking
about
can
apply
to
all
these
different
verticals.
So
the
scenario
that
we
are
going
to
consider
is:
how
can
we
predict
the
next
repair
to
an
EKG
machine?
So
what
are
some
of
the
challenges
that
we
would
face?
One
is
around
scalability,
so
we
are
talking
about
big
data
here.
B
There
is
a
variety
of
sensor
information
it
is
being
ingested
at
high
velocity
and
it
needs
to
be
stored
over
a
prolonged
period
of
time
so
that
you
can
understand
what
is
normal
working
conditions
and
all
of
this
data
sometimes
can
be
extremely
sensitive
like
to
fluctuations.
These
machines
are
highly
calibrated,
so
the
accuracy
of
the
data
is
important.
You
know,
sometimes
you
might
have
late-arriving
data.
Sometimes
you
may
have
data,
which
is
wrong,
which
needs
to
be
corrected
at
scale,
and
these
these
work
loads
do
not
operate
in
isolation.
B
You
would
have
to
combine
it
with
other
data
sources,
and
you
know
if
all
of
this
was
possible,
then
you
need
to
be
able
to
do
it
in
a
cost-effective
manner
to
handle
all
these
little
files
coming
in
in
a
continuous
stream
managing
the
end-to-end
latency
dealing
with
streaming
challenges,
and
you
would
have
to
work
with
folks
who
work
with
a
different
variety
of
tools,
so
some
of
them
could
be
just
bi
reporting.
Some
of
them
could
be.
B
You
know
machine
learning
at
different
levels
of
sophistication
and,
as
you
start
to
build,
you
not
only
have
to
manage
the
data,
but
you
need
to
manage
the
ML
artifacts
that
go
along
with
it
and
have
their
have
a
lifecycle
of
its
own.
So
who
are
the
folks
who
typically
deal
with
it?
There
are
domain
level
SMEs
for
sure,
but
there
are
data
engineers.
There
are
ml
practitioners.
B
B
As
you
are
going
to
build
a
modern
data
platform,
there
are
certain
characteristics
that
folks
look
for.
First,
is
you
want
the
speed
and
low
latency
of
a
streaming
platform?
You
need
the
scale
and
affordability
of
a
data
Lake,
and
you
need
the
reliability
and
performance
of
a
data
warehouse.
Of
course
there
are
trade-offs
like
so,
for
instance,
latency
and
cost.
B
There
would
always
be
a
trade-off,
but
at
the
back
of
our
mind,
these
are
the
three
primary
characteristics
of
a
modern
data
platform
to
be
able
to
handle
into
n
scenarios,
and
in
that
context
we
will
see
how
Delta
aids
the
scenario
how
it
addresses
most
of
these
challenges
that
we
have
talked
about.
Today's
talk
is
around
Delta
lake.
So
what
is
the
Delta
lake?
It
is
nothing
but
a
combination
of
Delta
tables.
Again
what
you
know.
What
is
the
Delta
table?
B
It's
a
set
of
data
files
if
you've
been
working
with
spark
bar
K
is
a
format
that
you
would
be
exceedingly
familiar
with.
Park'
is
highly
columnar,
it's
optimized
and
it
has
open
source
and
open
format.
So
there
is
no
locking
that
comes
along
with
it.
Delta
is
nothing
but
parking
along
with
a
transactional
file.
What
do
we
mean
by
that?
So
if
you
look
at
this
little
picture
here,
this
is
what
you
might
be
already
familiar
with.
You
have
partitions.
B
Let's
say
we
are
dealing
with
time
series
data
here,
so
you
have
a
date
partition
and
you
have
these
party
files.
This
is
where
your
data
resides.
Alongside
with
it
in
the
same
directory
as
your
table
say
my
table,
is
your
reference
table
Delta
table
that
we
are
talking
about?
You
have
something
called
a
underscore
Delta
log,
where
you
can
see
a
series
of
JSON
files.
This
is
where
each
of
your
transactions
on
the
data
are
logged,
and
why
is
this
important?
B
It's
important,
because
this
is
what
brings
in
your
asset
transaction
capability,
so
when
you
want
to
do,
updates
and
deletes
and
merges
at
scale.
This
is
what
helps
you.
It
helps
you
with
time
travel
and
all
of
us
are
guilty
of
occasionally
messing
up
a
pipeline,
making
a
human
error
or
who
knows
there
are
sometimes
upstream
processes
which
are
beyond
our
control,
which
may
have
come
down
we've
already
ingested.
It
and
you
know,
introduced
some
deficit
C
with
the
data.
B
B
It's
like
variations
of
their
data
sources,
but
at
the
same
time,
if
you
give
them
too
much
control,
then
instead
of
pristine
data
Lake
it
can
turn
into
a
swamp.
So
with
schema
enforcement,
you
can
control
and
discipline
in
a
discipline
way
an
evolution
of
the
schema,
as
opposed
to
just
a
dumpy
out
of
everybody
putting
in
whatever
they
feel
like.
These
tables
are,
of
course,
registered
with
the
meta
stir.
So
people
can
look
at
the
schema
and
understand
you
know
what
it
is
that
it
contains.
B
We
are
talking
about
cloud
here,
so
the
storage
is
important
and
we
have
storage,
so
is
suppose
you're
in
the
AWS
ecosystem.
Then
it
is
s3.
If
you're
in
the
OSHA
or
eco
system,
it's
going
to
be
blob,
storage
or
adls
or
whatever
it
is.
These
have
become
exceedingly
popular
because
of
some
advantages.
They
offer
primarily
around
affordability,
scalability
and
high
availability.
People
will
be
able
to
dump
files
in
there
and
have
direct
access
to
it,
but
there
are
a
few
shortcomings
of
it,
also,
as
the
number
of
files
starts
to
grow
listing
this
directory.
B
This
you
know
folder
and
bucket
starts
to
get
exceedingly
long,
and
so
there
are
some
alternatives
around
it,
which
is
what
we
are
going
to
see
in
the
demo
today,
primarily
around
notification.
So
there
is
an
s3
sqs
connector,
for
instance,
that
you
can
listen
into
only
when
a
new
file
comes
then
you're
going
to
be
notified
about
it,
and
that's
when
you
pick
it
up,
so
you
do
not
have
to
list
over
and
over
again
and
there's
a
new
feature
called
Auto
loader
that
was
introduced
in
very
a
you
know.
B
B
Why
are
we
doing
all
this
because
we
want
a
single
source
of
truth
for
all
our
downstream
apps
and
all
the
inside
generation
that
folks
are
doing
can
be
done
off
a
single
source
of
truth,
and
there
are
advantages
to
it
primarily
around
the
fact
that
there
is
compute
and
storage
separation
spark
inherently
supports
it,
and
now
we
are
just
enforcing
that.
It
is
important.
B
Storage
is
far
more
cheaper
than
compute
and
if
you
can
put
it
in
inexpensive
storage,
your
storage
can
grow
independently
and
you
would
use
your
computer
only
when
you
truly
need
it.
There's
increased
reliability
and
performance.
So
we'll
talk
about
the
architecture
shortly,
but
Delta
by
itself
will
help
provide
higher
reliability
to
your
data
through
those
as
the
transactions
with
your
ability
to
do
reader/writer
isolation,
so
that
at
no
point
in
time
you
see
bad
data
and
with
multi-hop
we'll
see
how
you
can
avoid.
B
You
know
propagating
your
fault
throughout
your
pipeline,
so
there's
better
fault,
isolation,
easier,
debugging
and
there
are
different
sls
on
your
consumption
pattern,
depending
on
your
appetite
to
consume
the
data
and
we'll
see
what
those
quality
guidelines
are
and
then
structured
streaming
is
a
very,
very
important
part
of
Delta
as
well.
Those
of
us
from
their
big
data
architecture
and
Hadoop
systems
and
all
must
be
familiar
with
lambda,
where
you
need
two
ways
of
handling
data,
one
through
a
batch
method
and
one
through
a
streaming
method
and
over
time
it's
difficult
to
maintain.
B
B
So
this
is
your
platform.
This
is
your
s3
bucket.
We
are
talking
in
the
context
of
sensor
data
coming
from
your
EKG
machines,
so
as
they
land
on
the
s3
bucket,
you
can
ingest
it
directly
or
you
can
have
this
sqs
connector
only
you
will
be
notified
only
when
a
new
file
comes
in
as
the
file
gets
dropped,
it
goes
into
a
brown
zone
and
then
there
might
be
further
enrichments
and
refinement
to
the
data.
B
There
might
be
a
need
to
join
against
a
reference
table,
so
that
is
when
this
is
going
to
become
a
silver
table.
That's
enriched.
There
are
further
aggregations
and
roll-ups,
and
that's
when
this
is
going
to
be
considered
as
a
gold
table
now
remember
these
concepts
of
bronze,
silver
and
gold
are
quality
guidelines
only
so
if
your
use
case
demands
two
sets
of
silver
tables-
or
maybe
you
want,
you
do
not
need
a
gold
table.
That
is
perfectly
fine.
It's
just
a
blueprint
and
you
should
not
consider
that
to
be
like
set
in
stone.
B
B
His
use
case
should
be
allowed
to
progress
in
parallel
to
an
m/l
practitioner,
who
might
be
more
interested
in
tapping
into
the
silvers
own,
to
repeat
an
experiment
from
data
produced
a
week
back
and
at
the
same
time
a
data
engineer
may
want
to
look
into
the
brown
zone
to
reapply
a
different
set
of
business
transformations
on
that
data.
That
may
have
arrived,
maybe
two
weeks
back.
So
these
are
different
timelines.
We
are
talking
about
we're
talking
about
different
SLS.
B
Here
we
have
a
the
notebook
in
two
different
sections
to
correspond
to
these
activities
of
ingesting
and
then
read
from
bronze,
refinement
and
then
read
from
silver
and
then
finally
doing
the
roll-ups
to
have
their
gold
tables
and
tapping
into
it.
So
you
can
think
of
like
a
bi
analyst,
using
an
external
tool
like
something
like
a
tableau
or
a
power
bi
and
sitting
on
their
desktop
and
being
able
to
tap
into
this
ecosystem
as
well.
C
So
I
want
to
call
out.
Andes
has
create
this
awesome
demo
for
all
of
you,
so
these
notebooks
will
just
like
the
recording
for
this
session.
These
notebooks
will
be
available
for
you
to
download
and
test
out
yourself.
Actually,
ok,
so
that's
an
important
call.
So
we're
not
just
showcasing
like
you
know
how
to
do
it.
Good
luck,
she's
actually
created
a
really
cool
set
of
notebooks
that
you
can
go
ahead
and
you
run
your
own
multi-hop
pipeline
here,
ok
and
so
each
one
of
these
call-outs.
C
You
know
the
one
data
ingest,
one,
Avery,
bronze
and
so
forth
and
so
forth.
These
are
associate
to
a
different
notebook
that
she's
about
to
show,
and
they
are
all
like
I
said
no
before
they're
all
going
to
be
available
in
our
data,
Brooks
tech
talks
session
and
and
github
repository
there
we
go
and
that
you
can
then
go
ahead
and
download
and
run
yourself
and
just
as
Karen
noted
in
the
chat
you're.
C
B
So
this
is
the
s3
bucket
and
we're
going
to
show
how
you
would
have
to
do
it
directly
consuming
from
the
bucket
or
let's
say
you
have
an
sqs
stream,
which
you
have
given
privileges
to
receive
events
from
this
particular
folder
of
the
bucket.
Let's
say
we
have
this
folder
sensor
and
then
this
is
where
the
magic
happens
of
joining
the
two
you're
saying
that
every
time
a
new
event
or
a
new
message
comes
into
this
bucket
into
this
folder
sensor,
then
start
an
sqs
notification.
B
We'll
come
back
to
this
picture
a
little
later,
but
let's
take
a
quick
look
at
the
data
set
that
we
are
going
to
use
there.
Is
this
public
data
set
from
kaggle?
It's
a
multiple
multivariate
time
series
and
you
can
you
can
see
what
it
looks
like
there
is
a
machine
number,
it's
a
time
series.
So,
instead
of
having
an
actual
time,
one
two
three
four
says
that
okay,
one
is
coming
before
the
other.
There
are
three
operational
settings
of
the
machine
and
there
are
21
different
sensor
readings
associated
with
it.
B
We
have
a
training
and
we
have
a
test
and,
of
course,
there's
a
ground
truth
which
corresponds
to
how
much
is
the
remaining
useful
life
of
that
particular
machine.
So
when
we
are
going
to
do
some
ml
on
it,
that
will
become
handy
there's
a
little
setup
that
is
needed,
so
you're
going
to
do
a
streaming
use
case
and
streaming.
B
Not
only
do
you
account
for
data
path
where
your
data
is
going
to
reside
in
those
three
zones
that
we
talked
about,
but
you
need
a
checkpoint
location
also
and
wise
checkpoint
directory,
particularly
important
in
the
context
of
streaming,
is
because,
if,
for
whatever
reason
you
have
a
hiccup
in
the
set
up,
then
you
do
not
want
to
start
to
reinvest
from
the
beginning
or
you
do
not
want
to
lose
data.
That
checkpoint
will
help
you
rewind
the
system
to
that
period
of
time,
so
that
you
neither
have
drop
data.
B
Not
you
have
to
placate
data
and
you're
going
to
have
a
checkpoint
for
each
of
those
zones
that
we
are
talking
about
you're
doing
a
quick
check.
We
ensure
that
those
directories
have
been
created,
and
you
know,
as
we
noted
earlier,
no
no
data
works
in
isolation.
At
some
point,
you
would
have
to
combine
it
with
some
other
useful
data.
C
So,
for
example,
when
we're
talking
about
the
stream
of
EKG
data
that
Andy
has
been
talking
about
right,
so
we
have
all
the
different
operational
metrics
and
the
sensors
that
are
coming
in
from
the
KPG
machine.
That
tells
us
it's
time
to
live
or
time
to
failure
right,
because
we're
trying
to
basically
use
the
sensor
information
to
be
able
to
predict
okay,
this
this
component
of
the
EKGs,
the
sensor
rod,
is
gonna
fail
at
this
particular
point
in
time.
C
Basically,
how
old
machine
is
what
components
does
it
have
things
of
that
nature?
So
the
this
reference
table
that
he's
referring
to
is
basically
an
example
of
a
very
slowly
changing
type,
two
dimension
or
the
main
table
that
contains
the
EKG
model
information.
So
that
way,
you
still
have
that
info
and
combine
that
with
the
stream
of
sensor
data
as
it's
coming
in
and
together,
you're
actually
gonna.
You
your
need
to
join
those
two
sets
of
data
together
to
be
able
to
figure
out
okay.
C
What
ultimately
will
be
the
time
to
failure,
and
just
as
we've
noted
right
from
the
beginning,
we're
trying
to
go
ahead
and
understand
what
is
the
time
to
failure
and
trying
to
get
ahead
of
the
curve?
So
we're
not
waiting
for
the
machine
to
fail.
We're
gonna
predict
its
failure
rate.
It
will
predict
when
it's
gonna
fail,
so
that
way
we
can
repair
it
before.
We
actually
have
to
go
to
something
a
lot
worse.
C
After
all,
when
it
comes
to
EKGs
and
other
medical
equipment
takes
a
long
time
for
supplies
to
come
in,
to
fix
it
and
also
it
takes
a
long
time
to
fix
it
as
well.
So
these
are
things
that
we're.
If
we
can
get
ahead
of
the
curve,
it
actually
will
allow
smoother
operation
of
the
data
of
the
excuse
of
the
machines
that
we
have.
B
As
we
noted,
there
is
an
include
directory,
which
holds
mostly
constants
that
can
be
included
in
all
the
other
notebooks.
So
these
are
the
examples
of
like,
where
are
the
pots
and
the
schema
of
course,
so
we
talked
about
the
each
of
those
machines
having
a
time
series
associated
with
it,
three
operational
settings
and
21
different
readings
with
God.
Let
me
go
and
start
up
the
streams
and
load
some
data.
While
we
talk
so.
C
C
That's
constantly
running
and
processing
this
data,
so
as
the
data
is
coming
in
and
being
written
into
the
file
system,
the
data
books
notebook
in
this
case,
or,
basically
your
structure
stream,
if
you're
not
using
data
Bex
but
you're,
just
using
a
spark
structure
streaming,
it's
going
in
and
actually
processing
this.
So
right
now,
it's
just
a
simple
example:
the
files
are
being
dropped
directly
into
the
s3
folder
and
then
for
sensor
data
and
then,
when
we
flip
back
to
the
notebook
you'll
be
able
to
see
the
the
data
being
processed
as
you
live.
Basically.
B
C
Right,
we
might
need
to
drop
it
again
this
because
we
forgot
to
load
it
first.
So,
let's
see,
if
it
let's
see,
we
might
need
a
drop,
did
the
files
again
just
in
case
yeah
or
a
different
set
of
files,
but
nevertheless
the
context
is
that
this
is
what's
cool
about
this
auto
loader
mechanism,
as
the
data
is
being
dropped
into
the
file
system.
Just
like
that,
basically
you'll
see
the
data
being
processed
automatically
being
picked
up
by
source
for
streaming
right
from
the
get-go
and
by
the
way
there
are
multiple
mechanisms
to
that.
C
B
You
can
flip
between
the
dashboard
view
of
the
live
graph,
as
well
as
the
raw
grass.
So
you
begin
to
see
the
spike,
and
now
you
have
the
sliver
of
information
that
is
coming
in
and
you'll
continue
to
see
it
because
you
can
either
trigger
it
on
a
certain
micro
batch
interval
like
a
three
second
or
five
second
or
one
hour.
Whatever
you
consider
necessary
to
control
the
ingestion
of
inflow
of
the
data
and
as
you're
as
we
speak,
you
can
start
to
see
the
data
go
in.
B
You
can
flip
into
this
view
and
say:
okay.
This
is
the
second
batch
since
so
there
are
two
batches
the
first
batch
did
come
through,
and
this
is
the
second
batch
in
how
many
rows
it
has
and
what
what
is
going
on.
At
the
same
time,
we
have
a
right
stream,
operation
and
notice
that
the
format
here
is
Delta,
we're
using
a
checkpoint,
location
and
you're
beginning
to
see
this
data
files
streaming
into
the
into
the
right
part
of
the
string.
So
we
broken
it
into
a
read
stream
and
a
write
stream.
B
B
Let's
go
to
the
second
notebook,
which
is
one
a
three
brands
here
we
are
going
to
read
from
the
branch
table
and
you
will
notice
that,
as
we
are
pumping
data
into
our
stream,
we
are
still
able
to
read
data
which
has
already
landed,
so
you
don't
have
to
wait.
You
do
not
have
to
pause
the
system
while
the
ingestion
is
going
on
in
parallel.
You
can
continue
to
work
off
it.
B
All
of
that
is
not
necessary.
There
is
no,
never
any
partial
data
nor
any
bad
data
that
is
presented
to
the
consumption
layer.
We
can
do
other
things
like
counting
from
from
the
data
that
has
just
landed
and
some
other
things
are
like.
You
can
do
a
describe
history
on
the
table.
So
this
is
where
you
begin
to
see
something
new.
There
are
these
versions
so
which
was
not
not
available
in
a
party
file
or
any
other
data
files
that
you
were
working
off.
B
B
That's
the
name,
that's
the
data
part
who
created
and
all
the
other
details,
but
it
is
important
to
know
that
these
are
very
small
details
around
the
data
that
is
stored
in
the
metadata
with
streaming
and
then
the
volume
of
you
know
changes
that
that
is
possible
on
every
ingest
of
the
data.
We
store
the
actual
schema
in
the
underscore
delta
underscore
log
directory,
and
let's
take
a
look
at
that
directory
and
see
what
it
looks
like.
B
As
you
can
see,
we
saw
the
JSON
files
and
we
will
see
them
the
party
files
in
a
in
a
parallel
directory.
Let's
take
a
peek
at
one
of
these
JSON
files
to
see
what
it
looks
like
and
sure
enough.
You
see
the
schema.
You
see
the
CMAs
string
of
having
a
name
of
a
unit
number
of
integer,
the
metadata,
the
cycle
time
and
each
of
those
sensors
an
operational
settings,
so
this
was
just
to
you
know
emphasize
that
it
is
not
stored
in
the
hide
meta
store.
B
So
the
next
table
here
is,
which
is
data
refinement
and
as
before,
we
are
going
to
have
a
different
mindset,
a
different
hat,
a
different
set
of
rules
playing
in
here.
This
is
where
we
want
to
join
against
a
a
new
reference
table.
We
want
to
apply
some
business
transformation,
we
want
to
cleanse
the
data
and
then
persist
it
back
into
the
silver
zone
so
like
before
we
read
from
the
bronze.
This
looks
very
similar
to
the
data
that
was
coming
in
P
because
we
did
not
make
any
transformations
major
transformations
on
the
bronze
table.
B
This
is
the
reference
table
that
Danny
was
talking
about
earlier,
which
has
got
information
about
serial
numbers
specific
to
a
machine
so
that
when
you're
joining,
you
know
in
what
context
you
are
looking
at
the
data
and
that's
the
joint
right
here.
You've
done
some
basic
filtering
and
joining
against
it.
It's
the
ancient
number
or
the
machine
ID
against
which
you
have
joined.
B
So
you
come
up
with
a
data
frame
which
is
yet
another
data
frame,
but
it
has
got
these
additional
information
associated
with
it,
and
this
is
what
we
persist
in
to
the
silver
zone
like
thank
the
goals
on
this
is
what
is
gets
persist
into
the
silver
zone.
Let's
move
on
to
the
next
one-
and
this
is
an
ml
practitioner,
for
instance,
we
would
want
to
come
and
do
some
analysis
on
the
data,
so
this
is
exploratory
data
analysis
you
might
want.
B
You
I
see
what
are
the
correlation
between
the
different
sensor
readings
you
might
want
to
use
a
sequel
to
do
this,
I
mean
you
might
be
a
Python
programmer.
Who
wants
to
do
this
or
it
might
be
an
hour
library
that
you
want
to
use.
So
all
of
these
different
languages
is
easily
supported
and
the
results
are
just
the
same.
So
if
we
were
doing
something
very
simple,
like
a
count
in
sequel
versus
in
our
versus
in
Python
they're,
all
going
to
be
the
same.
So
what
it's
trying
to
say
is
at
this
point.
B
You
have
a
lot
of
tools
at
your
disposal.
You
have
certain
frameworks
and
languages
that
you're
comfortable
with,
and
you
can
use
any
one
of
them
to
work
off
the
same
data.
You
do
not
have
to
copy
it
anywhere
else,
and
now
that
we
have
our
data
well
taken
care
of
it's
reliable,
it
is
has
landed.
Now
we
can
do
some
modeling
activities
on
on
the
data
we
had
earlier
referred
to.
B
We
have
our.
We
have
our
fields
already
defined.
What
we
are
trying
to
do
here
is
have
an
additional
column
called
the
time
to
failure
which
we
determine
by
finding
out
what
is
the
maximum
cycle
time
that
a
machine
like
that
usually
encounters
and
figuring
out
in
what
transition
period
of
that
machine
cycle,
we
are
in
and
doing
a
very
simple
calculation
to
find
out
what
is
the
remaining
cycles
left?
B
Let's
say
we
are
considering
this
analysis
for
the
next
30
days,
so
our
period
here
is
30
days
and
we
want
to
see
whether,
in
this
30
day
period,
we
are
going
to
have
a
TT
F
of
0
and
1,
and
that
is
what
is
going
to
determine
our
boolean
value.
That
is
the
label
that
they
are
going
to
attach
to
the
data.
So,
at
the
end
of
this
period
we
have
got
two
additional
columns.
This
is
the
T
TF
and
the
label.
B
We
do
the
same
thing
with
the
trink
data
set
and
right
now
we
have
our
data
ready
for
analysis.
The
usual
thing
would
be:
let's
establish
a
baseline
model,
let's
do
a
very,
very
simple
regression
model
to
determine
1
or
0
of
failure,
so
that
then
we
can
add
on
to
it
other
models
and
compare
and
contrast,
as
are
we
getting
better
with
our
prediction-
is
one
algorithm
more
suited
than
the
other
for
this
particular
scenario?
B
Again,
if
you
have
done
these
type
of
scenarios,
there
are
fairly
boilerplate
code.
You
have
a
set
of
features.
You
have
an
assembler
where
you
are
saying
okay,
these
are
my
input,
features
that
I'm
going
to
consider
as
part
of
my
experiment.
You
have
a
scalar
so
that
all
the
features
are
within
a
certain
range.
You
apply
the
logistic
regression
and
there
are
certain
how-to
parameters
to
it.
Let's
say
we
have
established
this
to
be
ten
and
point
five
and
you
scale
down
some
of
these
features.
You
build
a
pipeline,
but
then
you
decide.
B
Maybe
that's
not
going
to
be
right.
Maybe
I
want
to
experiment
with
some
other
set
of
hyper
parameters
and
that's
when
you
introduce
a
cross
grid
across
validator-
and
you
say:
okay
over
five
feelings:
I
want
to
experiment
using
a
pattern
grid
of
a
few
parameters.
This
is
just
a
demo,
so
these
are
very
small
examples.
Small
ranges,
but
you
get
the
drift.
You
would
be
experimenting
with
hundreds
of
these
and
not
just
across
a
single
model,
but
across
several
models,
but
it's
important
to
establish
a
baseline
and
so
yeah.
B
B
So,
LST
M
stands
for
a
long,
short-term
memory.
It
is
in
the
family
of
noodle
nets,
and
sometimes
it
works
very
well
for
this
type
of
use
cases.
Why?
Because
it
is
time
series
and
if
you
take
consecutive
time
series-
and
you
study
it
to
say
that
when
you
have
previous
values
like
this
it,
there
is
a
chance
that
a
machine
fails.
But
when
you
have
previous
values
like
this,
the
Machine
does
not
fail.
B
So
it's
not
a
single
value,
a
single
row
by
itself
that
you
are
considering,
but
you
are
considering
a
group
of
data
points
all
consecutive
in
time
to
help
make
that
decision
while
be
doing
this,
because
it
has
shown
that
these
sometimes
work
better
and
so
as
an
ml
practitioner.
You'll
have
to
do
your
due
diligence
to
see.
Does
it
truly
work
or
not
work?
And
that's
where
you
have
set
it
up
like
before?
You
have
the
TTFN
and
the
label
column.
B
You
have
your
validate
data
set,
which
you
use
to
see
how
well
your
training.
Was
it
an
overfit
or
an
under
fit?
And
then
you
do.
Similarly,
you
do
your
feature
scaling
you
reshape
it
because
then
lsdm
has
to
match.
You
come
up
with
your
network
architecture.
So
in
this
case
we
have
so
many
units.
So
many
epochs
your
drop
out
whatever
layers,
and
then
you
start
with
the
Train.
B
So
these
are
Findlay
boiler
plate
code,
but
you
have
to
do
it
over
and
over
and
over
again,
and
that
is
where
the
rigor
that
you
put
in
the
number
of
experiments
that
you
do
you
to
figure
out,
which
is
your
best
model
matters.
So
we
come
up
with
an
with
a
set
of
parameters.
Now
we
have,
we
have
given
it
its
training
data
set
and
it
has
gone
through
the
epochs.
It
has
come
up
with
an
accuracy
and
a
confusion
matrix.
So
what
do
we
want
to
do?
B
You'll
determine,
should
I
schedule
a
maintenance
for
that
machine
right
now,
or
should
it
be
next
week,
or
should
it
be
in
the
next
month,
or
maybe
it's
perfectly
healthy
and
I
do
not
even
have
to
worry
about
this
while
you're
doing
all
of
this,
let's
show
you
what
is
happening
with
ml
flow.
You
have
all
these
runs
recorded
here
and
I'm,
going
to
switch
back
to
my
mind,
map
here,
to
show
you
the
companion
to
Delta
Lake,
which
is
the
ml
flow
and
the
analytics
that
we
have
done
so
far.
B
B
So
what
is
ml
flow
ml
flow
is
needed
for
the
end-to-end
lifecycle
management
of
the
artifact
itself,
and
if
we
look
at
this
little
chart
here,
there
are
four
components
to
it:
tracking
projects,
model
and
registry,
so
tracking
is
what
people
often
go
to
because
in
the
past,
the
first
step,
because
this
is
what
helps
them
record
their
experiments,
so
we
just
ran
through
a
few
of
them.
Now
can
we
remember
a
week
from
now
as
to
what
we
learn
and
how
we
land
it
and
what
was
the
code
behind
it?
What
was
the
configuration?
B
B
So
tracking
server
is
constantly
recording
all
of
this
for
you
so
that
you
can
go
back
in
time
and
search
and
say:
okay,
I
want
to
find
what
was
the
best
model
that
had
produced
a
week
back
because
now
I
am
actually
regressing
and
I
want
to
understand
what
changed
did
I
do
cause
that
regression
to
happen
or
I'm
done
and
I
can
justify
it?
I
can
show
it
to
somebody
else
and
say
why
I
have
chosen
this
to
be
my
best
model
under
the
condition
under
present
circumstances.
B
So
you
choose
your
model,
you
know
the
best
one,
and
now
you
have
the
task
of
taking
this
to
another
environment
because
you
may
be
built
it
in
a
dev
setup
on
your
laptop.
Now
you
have
to
take
it
into
a
staging
environment
so
being
able
to
package
the
model
for
it
to
be.
Reproducible
in
many
platform
becomes
your
next
challenge
say
you
already
did
it
now.
The
next
step
would
be.
There
are
different
consumption
patterns,
like
maybe
somebody
expects
this
to
be
a
simple
Python
function
that
should
be
exposed.
B
Somebody
wants
it
to
be
a
batch
job.
Somebody
wants
it
to
be
a
streaming
one.
Somebody
wants
it
as
a
tensor
flow
model,
so
you
have
a
lot
of
deployment
options
which
is
completely
orthogonal
to
how
the
model
itself
was
built,
and
that
is
why
it
is
necessary
to
make
a
distinction
between
the
two.
The
last
one
is
around
registry.
So
in
the
coding
world
we
are
all
familiar
with
Gibbs
as
like
a
source
of
truth,
we
go
there
to
see
who
did
what?
B
What
is
the
latest
version
and
if
you
make
it
make
a
bad
judgment
call
you
can
always
go
back
and
to
a
different
version.
You
know
who
did
what?
If
somebody
breaks
the
nightly
regression
series,
then
you
can
go
back
and
pinpoint
exactly
what
happened
and
you
can
recover
from
it.
The
same
thing
is
needed
for
the
ml
world
as
well.
B
There's
a
need
for
a
centralized
location
for
this
artifact
that
is
precious
to
you,
because
these
artifacts
are
living
breathing
things
they
you
know
once
it
is
set
in
motion
and
it
is
taken
into
production.
Somebody
needs
to
be
accountable
for
it.
There
is
auditing
around
it,
there's
governess
oscars
around
it
and
then
every
morning
will
become
stale
at
some
point.
B
B
B
What's
the
metrics
a
unique
ID
for
each
one
and
and
at
the
at
the
same
time,
if
you're
not
happy
with
it,
you
can
come
back
and
and
try
something
else,
but
you
have
the
option
of
tracking
it
as
well
as
going
back
to
it
at
a
future
period
of
time.
Let's
move
on,
we
have
gone
through
our
silver
model,
a
silver
zone
which
is
where
ml
practitioners
usually
come
then
comes
the
gold.
This
is
where
data
roll-up
happens
for
bi
reporting
purposes,
so
you
can
do
a
read
stream
and
a
write
string
together.
B
This
is
just
an
example
of
it
and
at
this
point,
if
the
table
is
already
created,
then
you
can
go
and
query
it.
So
this
is
your
gold
stable.
This
is
simulating
an
analyst
coming
in
and
reading
curated
data,
using
sequel,
which
is
typical
tooth
for
analyst,
and
you
can
you
can
see
the
unit
number
in
the
count
and
just
for
fun
sake.
Let
me
go
to
my
tableau
dashboard
here.
B
B
I
can
refresh
it
and
from
this
point
onwards
you
know,
if
you
are
into
reporting,
then
you
would
know
you
know
what
are
the
best
charts
that
you
want
to
draw
out
of
these
data
points,
how
you
want
to
refresh
it
and
so
on
so
forth.
But
the
thing
to
point
out
is:
you
are
actually
looking
at
all
the
data
that
you
would
want.
It's
sort
of,
like
you
know,
pulling
a
sliver
of
data
somewhere
else
pulling
it
down.
This
is
as
quickly
and
to
end
as
you
can
possibly
get
to.
B
To
summarize,
we
had
the
use
cases
we
concentrated
on
a
single
one.
We
saw
what
were
the
unique
requirements
of
these
are
the
advantages
of
Delta
and,
along
with
the
Delta,
you
have
ml
flow,
which
helps
you
manage
your
artifacts.
So
it
would
be
fair
to
say
that
there
are
several
advantages
to
materializing
data
frames
in
this
multi-hop
pipeline,
and
that
includes
fault,
recovery,
easy
troubleshooting
and
ability
to
cater
to
consumers
with
a
different
access
pattern
SLS.
So
we
Delta
and
structure
streaming.
B
It
gives
you
a
very
clean,
simple
architecture,
blueprint
and
that
can
be
applied
in
very
similar
use
cases.
We
saw
the
loading
and
how
incremental
ingestion
at
scale
with
Delta
can
be
achieved
and
Delta
can
works
well
as
a
source
as
well
as
a
sink,
and
so
we
saw
how
we
can
change
different
stages
of
the
pipeline.
The
other
thing
to
consider
is
your
jobs
sometimes
start
out
as
batch.
B
Your
ended,
but
if
you
design
it
this
way,
then
you
can
easily
transition
from
a
batch
job
to
a
streaming
job,
with
just
few
configuration
changes
as
opposed
to
real
designing
recoding.
All
of
that
there
is
always
trade-offs
between
cost
and
latency.
So
you
have
the
levers
in
place
in
the
count
in
the
in
the
injection
and
the
read
stream
constructs
to
use
something
like
trigger
once
so,
if
you
say
trigger
once
and
give
an
interval
of
12
hours
that
essentially
becomes
your
batch
job.
B
You
have
processing
interval,
which
you
can
control
your
cost
also,
so
you
can
have
a
processing
interval
of
5
seconds
versus
15
seconds
and
so
on.
So
delta
lake
is
indeed
the
next
generation
data
lake
for
building
your
pipelines
and
the
analytic
offerings,
as
it
combines
all
the
desirable
characteristics
from
all
the
three
different
platform
types,
namely
the
low
latency
of
streaming,
the
scalability
and
affordability
of
data
lakes
and
their
reliability
and
performance
of
warehouses.
B
We
also
saw
ml
flow,
how
it
offers
several
building
blocks
to
manage
the
analytic
asset,
lifecycle,
tracking
hydrations
of
model,
training
model,
selection,
packaging
and
subsequent
deployment
and
promotion
to
higher
environments.
So
with
that,
we
can
see
if
there
are
questions
that
we
can
address
actually.
C
Do
me
a
small
favor
switch
back
to
the
slide,
so
I'm
gonna
go
ahead
and
just
finish
it
up
a
little
bit
and
then
also
I
will
go
ahead
and
either
answer
the
questions
directly
myself
or
I'll.
Ask
you
to
answer
the
questions,
if
that's
okay,
sure
alright
cool
so
justly
yeah,
let's
start
with
the
dock
like
containers,
if
you
can
click
on
present,
so
we
can
show
the
full
screen.
Thank
you
very
much.
C
Thank
you
very
much
perfect.
So
don't
worry,
I'm
going
to
answer
some
questions
and
there's
a
very
good
good,
important
call-out
from
in
the
zoom
webinar
chat
that
actually
I'd
like
to
go
ahead
and
dress
to
because
it's
actually
a
very
good
point.
So,
for
starters,
just
to
give
you
some
context,
there
are
Delta
Lake
connectors
that
actually
will
talk
to
not
just
spark
as
you've
been
seeing
here,
but
will
also
talk
to
hive
presto
and
other
systems
as
well.
I
flipped,
the
next
slide.
C
Please
there
are
many
partners
and
providers
that
are
already
working
with
Delta
link.
So
this
is
the
reason
why
we
talk
about
doesn't
like
as
much
as
we
do
and
also
if
we
can
jump
to
the
next
screen.
Where
again,
please
there
are
many
users
of
Delta
like
okay
and
yeah,
just
flip
the
next
screen
and
that's
not
for
the
market,
feel
okay.
So
again,
if
you
want
to
provide
some
comments
after
the
sessions
over,
we
just
would
love
you
to
describe
to
the
data
books
YouTube
channel,
so
you
can
keep
the
conversation
going.
C
So
now
looks
let's
address,
because
we
actually
have
two
very
different
views:
both
valid
okay,
one
view
was,
for
example,
that
there's
a
wide
range
of
topics.
That's
really
interesting
lots
of
food
for
thought,
that's
pretty
cool,
yet
at
the
exact
same
time,
there's
some
other
feedback,
which
is
also
valid.
You
know
that
oh,
the
stitches
too
densely
packed
either
make
it
you
know
90
minutes
long
or
cover
fewer
topics.
C
In
fact,
this
is
the
exact
debate
in
DITA
and
I
were
talking
about
it
could
should
we
shrink
this
thing
down
or
make
it
a
two
part
or
a
three
parter,
the
reason
we
actually
decided
that
it
made
sense
to
still
keep
it
this
place,
because
you,
as
the
data
engineer,
data
scientist,
you're
gonna,
have
to
do
all
this
right.
This
is
what
you
just
saw
here
today
and
indeed
of
went
ahead
and
had
to
do
this
in
order
to
be
able
to
build
up
for
her
set
of
customers.
C
Okay,
I
OT
customers
and
all
of
those
components.
Are
there
now
don't
get
me
wrong?
Youyou
could
just
easily
turn
out
said:
hey,
but
dude
you
could
have
just
went
ahead
and
made
this
into
a
multi-part.
The
part
of
the
reason
why
we
did
not
do
is
multi
parts
because
we're
in
realistic
as
we
dive
into
each
one
of
these
issues,
what
we're
planning
we
had
actually
need
five
different
sessions.
We
did,
we
thought
none
of
you
would
actually
be
tolerant
of
that.
So
what
we're
doing
instead,
actually
is
we're.
C
Gonna,
be
reading
a
blog
post
to
Oh
up
this
session.
That
actually
will
contain
not
just
the
full
notebooks
in
the
full
context,
but
also
links
to
all
the
various
different
videos,
either
already
pre-existing
ones
or
some
new
ones
that
allow
you
to
actually
fall
through
and
build
this
entire
system.
C
But
what
ultimately
comes
down
to
is
that
it
is
densely
packed
because
you,
as
a
date,
an
engineer,
a
data
scientist
when
it
comes
to
building
these
multi-hop
pipelines
with
IOT
you're
gonna
actually
have
to
be
familiar
with
each
and
every
one
of
those
those
concepts
so
again
great
feedback
from
everybody.
We
really
enjoy
it.
It's.
So
it's
not
like
we're
saying.
C
What's
right
or
what's
wrong,
it's
just
that
we
were
trying
to
find
that
right
balance
of
how
to
make
sure
that
at
least
the
context
or
the
primer
is
there
and
again
that's
why
the
notebooks
are
going
to
be
there
right
away
for
you
to
download
as
we're
writing
the
blog
and
once
the
blog
goes
out,
we'll
make
sure
to
update
everybody
accordingly,
so
that
information
is
there
for
so
you
can
run
it
yourself.
Okay,
all
right,
so
there's
also
an
excellent
kind
of
EDA
that
an
all
credit
to
you
for
this
one.
C
Some
folks
really
love
your
mindmap,
yeah
I,
absolutely
agree
that
was
pretty
that
that
is
absolutely
sweet.
Okay,
all
right!
So,
let's
see
okay!
Well,
let's
we've
got
about
two
minutes
left,
okay
and
so,
like
I,
said
before
subscribe
to
the
YouTube
channel.
So
that
way
you
can
go
ahead
and
we
can
continue
the
conversation
there,
because
this
video
will
go
up
there
and
we'd
love
to
actually
continue
talking
to
folks.
C
B
Yes,
you
have
a
lot
of
luck
in
any
system.
You
have
a
lot
of
options
for
querying
faster,
like
Delta.
Has
things
like
file
level
operations,
Z
ordering
compacting
optimized,
auto
optimize
bin
packing?
All
of
these
are
geared
towards
your
use
case,
and
so
you
have
to
understand
exactly
what
your
access
and
make
sure
that
you
are
following.
B
You
know
alongside
it,
so
your
question
is
definitely
very
loaded
in
terms
of.
Is
it
geared
for
a
fast
access
it?
You
know
if
you
do
not
use
the
right
partitioning,
for
instance,
if
you
do
not
use
the
right
Z
ordering
in
your
where
clause
and
you're
doing
a
full
table
scans.
Of
course
it
can
be
extremely
expensive,
but
that
is
why
you
have
all
these
options
at
your
disposal
to
be
able
to.
You
know,
make
the
most
out
of
it.
C
Actually,
I
don't
but
I
do
want
to
answer
one
more
question
before
can
we're
gonna
switch
right
back
to
you
and
and
then
the
session,
because
I
want
to
make
sure
we're
cognizant
everybody's
time.
There's
a
quick
question
about
how
do
we
connect
a
real-time
application
that
renders
predictive
maintenance
in
the
real
time
right,
where
the
calls
can
be
a
sickness,
okay
and
so
the
the
quick
context
that
I
would
basically
go
ahead
and
call
out
and
then
Nvidia
I
would
love
you
to
add
to
it.
C
It's
basically
that
is
the
most
the
context
of
structure
streaming
right.
The
the
idea
is
that
with
Delta
Lake
and
we
actually
had
processions
in
the
data
Burke
YouTube
channel,
which
talked
about
how
you
can
have.
Instead
of
the
lambda
architecture,
we
can
have
the
Delta
architecture
which
allows
us
to
deal
with
both
streaming
data
and
also
with
batch
data
at
the
exact
same
time.
C
In
fact,
during
today's
session,
that's
exactly
what
was
shown,
which
is
that
there
were
some
queries:
that
of
the
IOT
data
that
you
were
literally
reading
the
data
as
it
was
coming
in
live
well.
In
some
cases
we
were
going
ahead
and
looking
at
historical
data
or
joining
the
historical
data
to
provide
that
context.
So
anything
else,
do
you
think
I
that
I
miss
then
Anita
that
should
be
called
out.
B
C
Cool
all
right.
Well,
then,
hey
Karen
I
think
we're
good
to
go
I
in
terms
of
the
session
and
before
we
close
up
I
just
want
to
say,
I
apologize
for
anybody
else.
We
were
not
able
to
answer
the
questions,
but
please
do
ping
us
on
the
YouTube
channel,
because
we
are
active
there
to
answer
questions.
Okay
and
back
to
you,
okay,
angst.
A
Any
thanks
in
edita
does
a
great
presentation.
Like
Danny
mentioned,
this
talk
is
recorded
and
it's
gonna
be
uploaded
to
YouTube
and
we'll
also
make
sure
you
have
the
to
the
notebooks
so
watch
out
for
your
emails,
if
you're
with
us
on
zoom'
you'll
get
any
more
than
24
hours
with
all
the
the
links
as
a
follow-up
email.
So
thanks
everyone
for
joining
us
and
have
a
great
rest
of
your
day
take
care
of
bye.