►
Description
Don’t miss out! Join us at our upcoming events: EnvoyCon Virtual on October 15 and KubeCon + CloudNativeCon North America 2020 Virtual from November 17-20. Learn more at https://kubecon.io. The conferences feature presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.
An Open Platform for Trading Interconnected Equities and Assets - Walid Ali, Google
Asset Trading (Stocks, Commodities...) fuses various data sources, stat & ML models & visualization tools. K8s and Kubeflow massively scale to thousands of models, brokerage organizations & analysts We illustrate how K8s handles Stock Trading by creating two clusters (Dev & Prod) to handle: -Thousands of stocks where ticker info are ingested, processed and stored in Stateful Cassandra. -Legacy MS SQL Server accessed through containerized .NET apps. -Live Tweets are analyzed using NLP tools -News Feeds are analyzed through Thomson Reuters Eikon -Industrial sectors statistical analysis -ML Models predicting asset pricing, trained through Kubeflow &TFJobs -Inference through R-based pods, and ML models deployed through Seldon A Jenkins server handles CI/CD for automating model training and deployment. Scaling is carried out via Namespaces, RBAC, Net Policies & Resource Constraints.
https://sched.co/Zek0
A
Hi
everyone
in
today's
talk
we'll
be
addressing
creating
an
open
platform
for
equity
trading
at
scale
using
analytics
and
deep
learning
as
the
tools
for
the
underlying
platform.
The
goal
is
to
implement
both
a
development
and
a
production
component
of
the
system
in
a
production
level
trading.
We
care
about
too
many
aspects,
not
just
the
stock
price.
We
care
about
high
and
low
in
the
day
volume
which
many
primitive
trading
examples
rely
on.
A
This
is
the
workflow
of
a
single
asset,
whether
being
a
stock
or
a
metal
or
commodity,
and
so
on.
Sources
and
inputs
to
the
system
would
include
pricing
volume,
media
like
videos
and
sound,
and
news
and
text
tweets
facebook,
statuses,
company
financial
data,
in
addition
to
historical
data,
once
sources
are
available,
they
are
processed
through
stream,
processing,
polling
and
batch
processing
data
will
be
explored,
visualized,
cleansed
and
augmented,
and
in
reached
once
data
is
ready.
We
start
building
the
model
and
we
start
validating
this
model.
A
A
In
this
slide,
we're
showing
an
example
of
developing
production
clusters,
for
example,
in
the
development
side,
the
tensorflow
model
will
be
driving
the
training
service
based
on
long
short-term
memory
to
predict
the
price
and
volume
of
an
asset,
while
in
production
we'll
be
deploying
this.
This
model
at
scale.
A
To
give
a
sense
of
the
complexity
of
the
trading
floor,
there
are
multiple
types
of
trading
desks.
For
example,
equity
trading
desks
would
handle
everything
from
equity
trading
to
exotic
options.
Trading,
fixed
income
trading
desks
would
handle
government
bonds,
corporate
bonds
and
other
bonds
and
bond
like
instruments
that
pay
a
yield.
Foreign
exchange
trading
desks,
facilitate
trading
and
currency
payers.
By
acting
as
market
makers,
they
can
also
engage
in
proprietary
trading
activities.
Commodity
trading
desks
are
focused
on
agriculture,
products,
metals
and
other
commodities
such
as
crude
oil,
gold
and
coffee.
A
In
this
we
show
the
need
for
kubernetes
as
our
platform
to
handle
all
the
complications
surrounding
the
execution
of
the
prediction
service
on
a
single
asset.
Really,
the
size
of
the
box
here
is
entailing
the
complexity
as
we
captured
it
from
scully
and
others
in
their
published
reference
at
the
bottom.
But
the
idea
simply
here
is
there
are
a
lot
of
other
services
and
tasks
and
processes
that
go
before
and
around
and
after
the
machine
learning
code
is
done
and
being
able
to
comprehend.
A
All
of
this
is
extremely
important
for
the
valid
execution
of
the
system
from
a
machine
learning
perspective.
This
is
the
machine
learning
path
where
the
data
gets
ingested.
Analyze
transformed
validated,
you
split
it
in
the
model.
You
build
the
model
you
validate
it.
Then
you
augment
with
the
training
at
scale.
You
start
rolling
out
the
model
and
serving
it,
and
then
you
monitor
and
log
it
and
we'll
be
talking
about
each
stage
in
the
coming
few
slides.
A
A
A
We're
just
getting
you
back
to
the
machine
learning
path,
just
to
keep
in
mind
that
not
only
you
have
to
isolate
the
different
commodity
stocks
and
others,
but
you
will
have
to
ensure
this
process
goes
forward
with
every
cluster
you
haven't
had
the
first
path
which
the
data
is
ingested
in
and
processed.
For
example,
the
thicker
data
are
fed
into
the
ticker
processing
pods,
which
in
turn
are
ingested
into
alexandra
interface
pod.
A
Similarly,
the
tweets
are
processed
through
pods,
which
are
later
saved
into
the
cassandra
database.
We
relied
on
already
published
libraries
from
google,
for
example,
like
click
to
deploy
for
the
ease
of
managed
services.
The
scalability
is
achieved
through
vertical.scalar,
which
actually
helps
us
a
lot
that
we
don't
have
to
worry
about
it.
Since
it's
a
managed
service
from
cloud
provider.
A
Data
analysis
segment
is
extremely
important
and
it's
really
the
foundation
of
building.
What
comes
next
in
the
model
so
process
data
are
stored
inside
cassandra,
including
twitter
analytics,
which
have
their
servers,
read
the
process
data
and
provide
the
analytics
around
it.
Many
are
based
functions
based
on
chinese,
provided
pre-defined
analysis
for
development
part.
A
We
use
dataflow,
where
parallel
execution
of
algorithms
to
determine
the
best
correlation
and
protection
here.
The
multiple
here
there
are
multiple
services,
one
that
development
service,
pods
read
ingested,
tweets
analyze,
the
cinnamon
analyze
the
sentiments
through
communication
with
natural
language
processing,
parsers
and
finally,
recording
the
analyze
tweets
into
cassandra
database
data
scientists
test
their
algorithms
through
running
jupiter
hub
and
communicating
with
the
sentiment
analysis
service.
A
A
Some
of
these
filters
are
given
here
as
examples
in
this
list
of
filters
that
we'll
be
talking
about,
for
example,
an
arima
filter,
autoregressive
integrated
moving
average
on
all
stocks
and
testing
them
against
actual
return
for
every
filter.
We
run
many
parallel
algorithms
and
tested
the
best
possible
outcome
from
analysis
of
variance
anova.
A
A
A
And
now
getting
into
the
next
part
as
we're
creating
multiple
sets
of
chrome
jobs
want
to
transform
data
into
formats
used
by
tensorflow
another
to
is
to
do
the
transformation,
for
example,
pca,
whitening,
normalization,
outliers,
and
so
on,
so
being
able
to
run
the
cro.
The
cron
jobs
is
very
important
for
us
to
be
able
to
fork
off
these
different
tasks
simultaneously.
A
Examples
of
splitting
is
actual
splitting
is
more
complicated
than
just
dividing
them,
but
it's
extremely
important
to
try
different
sets
of
the
splitting.
That
would
result
in
directly
influencing
the
model
as
we
build
it,
and
this
part
is
executed
using
tensorflow
jobs,
tf
jobs
created
on
kubeflow,
where
the
longshore
terminary
lstm
for
price
prediction,
the
random
forest
models
for
asset
market
prediction
and
the
natural
language
processing
analysis
for
the
media
and
the
tweets
and
textual
information.
Multiple
simultaneous
jobs
are
created
and
validated.