►
Description
Speaker: Michael Laing, Systems Architect
The Internet of Things uses Topics to tag information.
Topics are segmented named channels that are attached when information is sent or stored
IoT Brokers use Retained Storage to persistently store information by Topic.
Retained Storage is a data store that is searchable using Wildcards in Topics.
Wildcards are reserved characters that match a single level or multiple levels in a Topic
By externalizing Retained Storage to Cassandra, IoT broker instances can autoscale, potentially handling 10's of millions of clients.
This requires efficient queries using Wildcards in Cassandra to access Retained Storage.
I will present strategies for implementing fast Wildcard queries composed of sub-strategies such as:
- Cluster key shuffling and auto-inversion to help determine partition key and narrow the row slice
- Sparse secondary indexes to minimize filtering
- Stratio's Cassandra Lucene Index to augment or replace other sub-strategies
illustrated by comparative benchmarks.
I will further discuss integration with IoT brokers at scale.
A
B
B
Lately
I've
been
looking
back
mostly
I,
generally
I
look
forward,
but
that's
when
I
got
going
a
Fortran
and
you'll
see
this
show
up
every
now
and
then
because
I'm
going
to
do
some
flashbacks.
That
I
think
are
relevant
to
my
talk.
So
here's
the
first
one,
that's
the
first
computer
I
ever
worked
on
death.
Not
me
in
the
picture.
I
choose
them.
That
was
a
g2
25
8
kb
of
memory.
You
could
look
in
and
see
the
little
donuts
inside
I.
B
At
the
new
york
times
and
in
previous
jobs
that
I've
had
technology
evaluation
at
Harvard
and
others,
and
it's
kind
of
what's
what's
Alice.
Looking
for
and
I
think
she's
going
where
the
excitement
is
I,
question
everything
I
get
told.
I
was
a
Marine
Corps
brat
then
went
to
West,
Point
Ranger
airborne
was
in
Vietnam,
then
went
to
Harvard
really
got
a
question:
everything
including
authority,
and
but
go
where
the
excitement
is
kind
of
tying
in
a
little
bit
with
datastax,
but
also
we're
still
back
in
time
here.
B
Let's
see
IBM
360
at
its
introduction
and
what
was
interesting
about
that
because
they
made
and
continue
to
make
an
enormous
amount
of
this
is
that
it
was
the
first
scalable
architecture
in
hardware
and
enjoyed
enormous
success.
It
had
50
times
scalability
and
the
models
that
were
introduced
over
the
first
three
years
and
the
key
was
at
the
separated
architecture
from
implementation
and,
as
you
think,
about
Cassandra,
where
you
got
the
cap
theorem,
if
you
do
a
good
job,
they
picked
availability
and
partition
tolerance
and
gain
the
scalability
that
we're
using
today.
B
This
is
kind
of
the
summary
of
my
talk.
I'm
going
to
hit
on
these
topics
IOT
or
fabric
MQTT,
wildcard,
queries,
query
strategies
and
so
forth.
So
we'll
come
back
to
that
outline
as
I
go
through
the
talk,
so
I
work
at
the
New
York
Times.
This
is
kind
of
how
I
think
about
it.
Was
it's
an
old
company?
It's
not
like
working
for
somebody.
B
New
we'd
still
do
print
and
it's
profitable,
but
we're
successfully
transitioning
to
digital
I
got
there
after
retiring
from
IT
management
at
the
United
Nations
and
I
really
wanted
to
go
back
to
tech,
so
they
gave
me
a
corner.
They
paid
me
every
month.
I've
worked
a
lot
of
startups,
that's
pretty
nice
and
we
get
reasonable
hours.
B
That's
not
a
lot,
and
so
of
course,
we
have
to
figure
out
ways
to
get
that
out
oops.
Here
we
go
back
in
time
before
the
internet.
I
worked
at
Harvard
and
we
had
30,000
computers
that
we
could
reach
across
our
internet
now
internet
this
internet
with
a
small
I.
There
was
no
commercial
internet
at
that
time
and
those
30,000
computers
were
all
on
disparate
networks.
Token
ring
you
name
it
all
kinds
of
networks
before
tcp/ip
became
ubiquitous,
so
that
was
our
reachable
internet.
B
If
we
kind
of
jump
into
the
rabbit
hole,
a
little
bit,
I
was
lucky
I
developed
a
system
that
is
really
the
progenitor
of
many
systems
that
I've
worked
on
that
deal
with
messaging
and
databases
and
I
got
to
compelled
a
complete
system
called
the
information
utility
that
connected
all
of
those
machines
using
c
and
db2,
and
mainframe
and
I
wrote
all
the
pc
and
mac
software
too.
So
with
a
lot
of
fun
for
me
and
save
Harvey
a
lot
of
money,
but
back
for
the
New
York
Times.
B
B
When
I
came
to
the
times,
then
SAT
me
in
the
corner:
I
work
on
infrastructure
or
new,
a
good
bit
about
Amazon,
Web
Services.
It's
just
why
they
brought
me
in
and
but
they
also
gave
you
the
license
to
think
and
so
I
thought
about,
publish
and
subscribe,
and
about
that
time
websockets
started
to
come
on
the
scene
and
become
usable,
and
so
we
thought
well
what?
B
If,
what
if
we
could
take
what
we
knew
about
web
socks
is
a
vague
idea,
which
is
why
you've
got
kind
of
a
vague
slide
up
there
and
we
could
sort
of
flatten
everything
out.
We
could
come
up
with
a
way
to
auto
scale,
make
it
global,
make
it
reliable
and
manageable,
and
I
was
just
one
guy
with
an
idea,
but
they
said:
oh
okay,
keep.
You
know
working
on
that
in
your
spare
time,
so
I
thought.
Well,
we
need
to
make
a
mesh
a
communications
mass.
It's
got
to
be
resilient.
B
It's
got
an
auto
balance,
use
lease,
latency
routing
all
these
goodies.
That
I
have
learned
about
in
n
AWS
and
we
kept
pushing
on
that
and
we
said,
okay,
we
could
put
in
nodes
with
roles.
Have
it
be
self-organizing
using
things
like
today
where
we're
moving
to
console
as
an
example,
but
we
wrote
a
lot
of
code
to
to
manage
this
in
Python
and
then
what
we
can
do
is
connect
up
using
web
sockets
and
other
messaging
technologies,
amqp
primarily
and
put
messaging
everywhere.
B
Multiple
zones
and
looks
like
this,
so
this
is
the
New
York
Times
fabric,
and
it's
really,
you
know
I
laugh
to
think
about
it,
but
I
think
about
it
as
kind
of
a
just
a
greasy
layer
that
distributes
messages.
But
it's
got
a
memory,
and
so
you
can
think
of
all
those
little
green
dots
in
there
being
Cassandra
notes,
and
so
we
have
a
cluster.
B
Our
production
cluster
is
running
in
every
region,
with
data
centers
in
those
regions
spread
out
across
the
zones
and,
as
a
result,
we've
never
had
a
failure.
Everything
else
is
also
distributed
in
the
same
manner.
This
whole
thing
scales
up
and
down
on
daily
basis,
at
about
a
15
to
1,
to
10,
to
1
ratio,
and
you
can
see
the
messaging
move
around
the
world
as
as
this
happens,
so
that's
pretty
cool.
So
it's
not
just
me
running
this.
B
There's
actually
three
of
us
now
and
we
do
all
the
DevOps
and
all
the
development
honest,
but
just
jump
down
the
rabbit
hole
a
little
bit.
I
think
this
is
an
aspect
of
Cassandra
doesn't
get
quite
enough
attention,
and
that
is
that
it's
a
staged
event
driven
architecture
and
I've
been
interested
in
this,
because
our
fabric
uses
the
same
kind
of
architecture
asynchronous
and
the
key
to
it.
B
Is
it
defers
processing
if
it
doesn't
have
to
do
it,
doesn't
do
it
right
away
queues
it
up,
and
so,
in
my
experience,
people
get
into
trouble
when
all
of
a
sudden
compaction
looks
like
it's
going
wild
on
and
they
didn't
anticipate.
Ince
hints
repairs
all
of
these
things
that
are
kind
of
the
deferred
processing
that
happens,
but
because
those
things
are
deferred,
then
the
response
time
can
be
fast.
B
So
if
you
don't
know
about
the
seed
of
architecture,
it's
worth
read:
there's
a
nice
paper
on
it
wikipedia
scott
to
make,
but
let's
go
back
to
this
messaging
everywhere,
an
asynchronous
microservice
architecture.
So
that's
what
we
have
and
again
here
where
we
are
so
that's
the
fabric,
but
what
about
IOT?
B
What
an
interesting
thing
we're
going
to
dive
down
the
rabbit
hole
again.
I
think
about
is
when
you
have
that
many
connected
devices.
What
could
what
can
happen?
We've
actually
seen
well
take
a
look
at
this
picture
and
what's
what's
wrong
with
that
picture
and
I'll
tell
you,
this
is
right
before
the
bridge
fell
down
and
the
reason
the
bridge
felt
down
is
because
of
resonance,
that
is
the
energy
coming
and
hitting
the
bridge
from
the
wind
caused
it
to
resonate
on,
what's
called
an
eigen
frequency
and
it
resonated
to
destruction.
B
We've
actually
seen
this
a
couple
of
times
in
the
financial
world,
our
colleagues
that
work
on
financial
systems
are
often
way
ahead
of
us
in
terms
of
the
velocity
at
which
their
transactions
occur
and
with
automated
trading
they
get
get
resonance
and
it'll
be
interesting
for
me
to
see
because
I've
studied
lattice
dynamics
and
we're
now
getting
up
big
enough
to
where
lattice
dynamics
may
come
to
play.
Okay,.
A
B
B
Think
of
the
things
of
interest,
if
I
think
of
our
users,
what
is
what
are
their
things
of
interest
and
actually,
if
I,
think
of
the
new
york
times,
what
are
our
things
of
interest
or
we
are
obviously
we're
interested
in
their
phones,
we're
in
student
or
watches?
Are
we
interested
in
their
wearables?
Are
we
interested
in?
There
are
refrigerator
when
it
gets
a
display
or
we
interested
in
a
whole
bunch
of
things
and
sort
of?
Where
is
that
overlap?
B
A
B
B
Maybe
my
my
colleagues
at
the
New
York
Times
have
been
around
so
long,
they're
young,
like
you,
and
maybe
they
don't
see
it,
but
my
job
is
to
sort
of
anticipate
and
create
a
platform
so
that
they
can
bring
their
creative
ideas
to
it
and
use
that
platform.
Have
it
ready.
We
already
know
how
to
auto
scale.
B
We
know
how
to
do
lots
of
things,
but
one
thing
we
haven't
really
quite
figured
out
is
how
we
can
extend
the
fabric
and
go
from
a
universe
where
we've
got
a
billion
devices
to
a
universe
where
we've
got
50
billion
devices
is
in
a
period
of
five
years,
where
we're
interested
in
some
subset
of
us.
So
if
you
want
to
want
to
scale
this
architecture
by
ten
or
a
hundred
times,
then
what
do
we
do,
and
my
approach
based
on
my
experience,
is
to
make
it
simpler,
make
it
simpler
make
it
flatter.
B
But
where
do
we
look
to
do
that
now?
There's
a
protocol
which
you
probably
haven't
heard
of
MQTT:
it's
an
IOT
protocol,
it's
simple,
but
it
has
quality
service
associated
with
it,
its
standard
Oasis
made.
It
is
standard
last
year
and
it's
mature.
That
is,
it's
been
around
long
enough
that
it's
actually
in
use
a
lot
of
places,
but
this
who
uses
it
anybody
her
a
facebook,
Messenger
I
gave
her
a
facebook
Messenger.
B
If
you
look
you'll
see
that
in
2012
there
was
a
lot
of
noise
about
am
using
MQTT
and
then
it
went
dark,
but
that's
the
backbone
telemetry
there's
lots
of
people
if
you
scratch
rounder
to
using
MQTT,
and
there
are
a
lot
of
open
source,
implementations,
servers,
clients
and
so
on.
So
I'm
not
sure
we'll
end
up
with
that,
but
this
is
where
we're
looking
now
some
features
of
MQTT,
it's
got
retained,
storage
and
a
session
transcends
connection.
Wait.
What
is
what
are
those
things
mean?
B
These
are
things
we
had
to
code
ourselves
into
our
fabric.
A
retained
storage
means,
if
you
send
a
message
and
somebody's
temporarily
disconnected
you'll,
keep
that
message
for
them
and
deliver
when
they
connect.
Okay
session
transcends
connection.
That
means
pretty
much
what
you
think
it
means
it
means
that
you
can
have
a
device
that.
B
B
So
I
can
I
can
avoid
some
pushback
and
and
scope
creep
from
my
clients
where
they
want
to
have
a
little
bit
because
not
allowed
sorry,
it's
not
a
standard.
So
now
we're
going
to
dive
into
what
is
a
message,
we're
sending
messages
everywhere
and
you
can
kind
of
think
well
things
and
events
and
those
are
consumed
by
other
things.
B
The
key
is
the
topic:
we're
going
to
talk
a
lot
about
topics
as
I
continue
and
the
values
of
blob.
We're
not
going
to
talk
about
it
any
more
than
that
and
in
fact
our
platform
really
doesn't
care
the
payload,
the
blob,
the
value
is
opaque.
As
far
as
we're
concerned,
we
only
look
at
the
topic
and,
if
necessary,
we
embed
metadata
the
topic
so
we'll
ignore
the
blob.
B
But
what's
a
topic
and
just
to
highlight-
and
this
is
often
a
point
of
confusion
in
messaging
systems-
the
topic
there
really
kind
of
two
flavors
there's
the
atomic
topic-
we're
pretty
much.
You
subscribe
to
the
top.
If
you
publish
to
the
topic-
and
it's
not
further
subdivided,
but
then
topics
can
also
be
represented
as
trees
and
with
a
separator
foo
bar
baz
is
a
is
an
example
of
a
published
topic.
I
would
publish
that
topic,
foo
bar
baz,
and
I
could
subscribe
to
it
using
either
the
same
values
or
wild
cards.
B
So
I've
got
a
wild
card,
then
I
can
use
a
plus
to
represent
any
level.
In
other
words,
I'll
show
you
some
examples
and
I
can
put
a
hash
at
the
end.
This
is
the
most
typical
way
in
which
topic
wild
cards
are
expressed.
There
are
other
ways:
rabbitmq
is
a
different
way
but
uses
different
symbols,
but
the
effect
is
the
same.
So
if
I've
got
this
topic,
foo
bar
baz
and
then
I
get
that
I
can
match
using
these
match
statements,
okay,
food
plus
doesn't
match,
and
the
reason
is
there's
no
third
term.
B
Okay,
they're,
okay
and
then
okay.
Well,
so,
let's
think
about
Cassandra
and
I'll
just
make
a
statement
here
and
then
I'll
try
to
attend
I'll,
explain
it
Cassandra
how
it
accesses
data
I!
Think
of
it
as
it
hashes
to
a
tree.
Ok
and
ok,
I'm
going
to
jump
in
the
rabbit
hole.
First,
if
you
haven't
done
the
course
is
on
at
the
data
sex
Academy,
you
should
do
them
they're,
really
good
and
they're
free
and
why
not
get
certified
I?
B
B
But
what's
the
partition
really
looked
like
it's
a
map
of
maps,
it's
ordered
by
the
cluster
keys
right
or
you
can
figure
it
as
grouped
by
the
cluster
keys
and
in
fact
it
can
be
considered
a
tree.
Okay,
you
can
think
of
it
as
a
tree.
That's
been
traversed
and
that
ends
up
being
useful
because
we
already
saw
that
you
know
a
topic
is
like
a
tree
and
if
a
partition
is
like
a
tree,
then
maybe
we
can
look
for
a
way
to
relate
those
two
concepts.
B
B
Okay,
any
current
messages
that
come
flowing
along
the
pipeline
you'll
get,
but
also
does
query
so
impute
dt
acts
like
a
database
and
gives
you
the
last
set
of
matches
for
your
subscription
and
your
subscription
can
include
wildcards
okay,
so
there
we
are
with
a
problem,
because
we
need
to
do
this
query.
We
want
to
do
it
in
a
way
that
scales.
We
want
our
whole
infrastructure
to
scale
and
we've
got
a
handle,
an
MQTT
query
and
the
feature
that
does
that
is
the
retain
storage.
B
B
A
B
B
So
those
the
methods
of
doing
and
secondary
index
keys
are
similar
and
result
in
other
composable
strategies
that,
as
I,
can
use
them
together
to
access
my
data
and
so
we're
kind
of
looking
for
the
same
thing
want
our
craft
square,
your
query
strategies
to
be
composable.
We
of
course
also
want
them
to
be
efficient
and
predictable.
B
A
B
But
everything
you
do
in
computing
was
in
the
mind
the
bedskirt
I
sixth
row
in
by
1980,
and
so
let's
talk
more
about
the
strategy.
Well,
you
also
said
start
with
the
most
risky
difficult
piece
first
and
if,
if
you
can't
kill
it
okay,
well,
we
couldn't
kill
it
so
strategy.
0
is
not
enough
because
we
want
to
be
able
to
handle
those
wild
cards.
B
B
B
Actually,
that's
an
area
where
I'd
like
to
get
in
and
maybe
contribute
a
bit
to
Cassandra,
because
I've
worked
with
databases,
all
my
work
in
life
and
so
some
of
the
strategies
from
postgres
or
mature
sequel
databases
could
probably
really
help
with
this
secondary
index
with
the
sender
secondary
index
performance
problems
that
we
have
in
current
releases,
so
we'll
scratch
that
strategy.
But
my
third
strategy
actually
ended
up
being
very
useful
and
that
is
to
look
around
and
find
some
open
source
software,
and
I
highly
recommend
you
check
out
strategy
on
the
scene.
B
They've
got
a
presentation
tomorrow
at
240
and
you
can
grab
this
man
right
over
here
on,
dress
and
he'll.
Tell
you
all
about
it,
but
it
basically
was
so
simple,
just
drop
in
the
jar
and
you're
off
and
running.
So
you
get
new
capabilities,
so
it
was
a
lot
easier
than
some
of
the
other
things
that
I
had
consider.
B
So
what
about
performance?
Where
are
we
on
and
okay?
So
no
wild
card
excellent
performance
right,
because
I
can
just
actually
my
topic
and
be
a
partition.
Key
and
I
can
just
go
directly
to
it.
Excellent.
I
read
excellent
on
right:
cluster
keys,
actually
excellent
on
read
single
partition
read
and
it's
okay
on
right.
The
number
of
inversions
it
does
depends
on
how
much
data
you're
storing.
B
So
if
you
keep
your
hierarchy
shallow
your
tree
shallow,
then
this
can
be
a
very
high
performance
solution,
forget
about
secondary
index
keys
and
estradiol
leucine,
it's
okay
on
reading
and
okay,
I'm
right
and
as
far
as
we've
been
able
to
determine
so
far.
So
where
do
we
go
with?
This
will
remember
these
strategies
are
composable.
B
So
if
I've
got
a
query,
it
looks
like
this.
How
am
I
going
to
how
am
I
going
to
break
that
down,
and
so
one
of
my
approaches
which
would
be
with
is
to
just
use
the
scene.
It's
moderate
complexity.
Basically,
we
generate
a
filter
with
a
regex
in
it,
but
by
itself,
and
we
hit
all
nodes
for
read
just
a
single
node
to
write
how
about
no
wild
cards
plus
the
scene
so
think
of
us
going
from
left
to
right.
B
B
We're
going
to
go
forward
this
time,
it's
kind
of
what
am
I,
what
kinds
of
things
am
I
thinking
about?
Where
are
we
going
to
go?
And
you
know
the
things
all
exist
out
in
the
wild
zone
and
they
come
and
go
come
and
go,
but
we
have
a
highly
available
zone
that
we've
created
in
AWS
and
with
our
own
internal
infrastructure
and
so
we're
looking
at
moving
more
things
to
the
edge
of
the
network
to
the
blue
boundary
there,
and
so
we
create
think
think
about
things
like
creating
user
proxy.
B
B
This
is
my
colleague,
elizabeth
and
I
hope
to
get
into
there
sometime
if
we
jump
down
a
rabbit
hole,
and
I
can
give
you
guys
some
advice,
I
think
it
would
be
this.
It
would
be
go
where
the
excitement
is.
That's
worked
for
me
for
my
whole
working
life,
don't
be
afraid
to
jump
down
the
rabbit
holes
and
try
things
out.
B
It's
a
great
time
to
be
raising
money
and
bringing
new
ideas.
Yeah
I've
been
through
a
number
of
revolutions.
I
made
a
lot
of
money,
I've
lost
a
lot
of
money
in
the
late
90s
and
no
turn
of
the
century,
and
we're
kind
of
in
that
same
sort
of
time.
Now
look
at
the
Internet
of
Things
and
it's
everybody's
got
a
protocol.
There's
a
lot
of
them
co-op
a
whole
bunch
of
stuff
happening
down
at
the
network
level.
B
B
B
So
I've
taken
the
for
example,
the
whole
publishing
stream
that
comes
out
of
New
York
Times,
the
Harry
Potter
novel
and
I
push
it
into
an
MQTT
broker
and
then
I
have
then
I
that
it's
very
easy
to
just
sort
of
tack
on
little
applications
that
read
from
that
broker
and
invert
things
and
do
stuff.
So
I
may
make
that
data
available
by
URL
by
topic
by
tag
all
kinds
of
different
ways
in
this
wild
card
format,
and
then
a
user
I
can
do
this.
B
B
So
the
key
for
me
is
I
built
a
platform
that
allows
the
developers
and
users
to
use
that
wild-card
capability
to
get
what
appears
to
them
to
be
almost
instantaneous
access,
because
it's
right
to
the
right
to
the
partition
of
interest,
pulls
it
right
out
so
in
and
it's
just
a
simple
structure.
So
that's
that's
one
way
to
think
about
the
use
of
Cassandra.
Another
way
to
think
about
it
is:
oh
I
got
this
global
thing
there.
You
know
it's
this
by
the
way
is
peer-to-peer
network
we
don't
just
publish
and
subscribe.
B
B
Other
ones
are
just
sensor:
data
I
mean
you
can
there's
all
all
the
stuff
you
can
grab
off
a
iphone
or
any
phone.
You
know,
location
oriente,
call
that
all
this
stuff
is
that
useful.
Is
that
useful
for
us
for
the
New,
York,
Times
and
I
would
say
right
now
now
right
now,
we're
not
using
that,
but
two
years
from
now
I
think
we
will
be
using
it.
So
we'll
be
looking
for
that
mutuality
of
things
of
interest
in
location
may
become.
One
of
those
you
know
when
you
open
the
door
to
your
house.
B
Is
that
when
you
want
to
see
your
breaking
news,
for
example,
I
mean
is
how
important
is
geolocation
to
us.
It's
getting
more
and
more
important
tag
everything
so
I
can
think
of
lots
of
kind
of
applications,
but
I
know
that
my
colleagues
back
the
New
York
Times
can
think
of
many
more
that
I
don't
and
they're
they're,
driven
by
analytics
and
I'm,
driven
by
sort
of
looking
in
the
rabbit
hole
and
seeing
what
I
can
imagine
for
the
future,
but
making
a
general
platform.