►
From YouTube: So You Want to Geo-Partition Your Data... - Deepthi Sigireddi & Jiten Vaidya, Planetscale, Inc.
Description
Join us for Kubernetes Forums Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
So You Want to Geo-Partition Your Data... - Deepthi Sigireddi & Jiten Vaidya, Planetscale, Inc.
One of the distinguishing features of Vitess is its support for flexible sharding schemes. In this talk Deepthi and Jiten will show how this can be used to build a custom sharding scheme that respects geo-partitioning requirements. They will then demonstrate a database cluster built using this scheme that solves data residency at the database layer obviating the need for any changes at the application layer.
A
I
guess
a
lot
of
people
are
still
at
lunch,
but
we'll
get
started
today.
We
will
be
talking
about
jurisdiction,
aware
databases
with
with
us
when
we
propose
the
stock
for
the
conference.
The
way
we
titled
it
was
geo
partitioning
and
the
answer
to
geo
partitioning.
Your
data
we
are
claiming
is
with
us,
and
in
this
talk
we
will
talk
about
how
you
can
use
wit
us
to
geo
partition
and
also
show
a
demo
of
it
working.
B
My
name
is
a
Jitendra
short
Japan.
Why
they're?
The
best
way
to
pronounce
my
name
is
a
letter
in
gene
number
10
I
used
to
manage
the
asari
and
DBA
teams
at
YouTube
back
in
2010,
when
the
Vitesse
project
was
started
there
by
my
co-founder
subbu
sukumaran
and
another
early
engineer
at
YouTube
for
scaling,
YouTube's
databases
and
for
the
last
two
years,
I
have
been
the
co-founder
and
CEO
of
planet-scale,
which
is
a
company
that
we
founded
for
making
giving
commercial
support
for
Vitesse
and
making
whitter
successful
commercially.
B
What
I'm
going
to
do
is
first
start
talking
about
the
problem
that
we
are
solving,
but
before
I
do
that.
Thank
you
everybody
for
coming
here
this
afternoon
and
thank
you
Sydney.
It's
a
lovely
city.
This
is
the
first
time
that
I
have
been
here
and
it
totally
lives
up
to
its
billing.
So
far.
So
thank
you.
So
what
problem
are
we
trying
to
solve
right?
You
all
are
probably
familiar
with
the
GDP
regulation
that
you
passed
a
few
years
ago.
B
There
are
many
different
aspects
to
gdpr,
but
one
of
the
aspects-
it's
data
locality
and
what
it
roughly
translates
to
is
that
if
you
have
a
particular
belonging
to
a
particular
political
entity,
the
data
associated
with
that
user
needs
to
live
in
physical
infrastructure,
which
is
bounded
by
the
boundaries
of
that
jurisdiction.
Right
and
basically,
that's
the
problem
that
we
are
trying
to
solve
and.
B
Many
times
to
support
this
kind
of
requirement,
you
need
to
reallocate
actual
applications,
which
means
that
all
layers
of
your
application
need
to
be
aware
about
the
geographical
location
of
a
particular
data
and
then
take
appropriate
actions
for
saving
that
data
in
appropriate
places,
and
this
is
a
daunting
task.
I
mean
you,
don't
want
to
really
collect
your
existing
applications
or
suffusing.
This
logic
throughout
your
application
makes
the
velocity
with
which
you
can
deliver
new
features
go
down
right
also,
if
you
have
an
existing
monolithic
database
that
you
want
to
make
compliant
migrating.
B
The
data
from
this
existing
database
into
multiple
database
clusters,
which
are
resident
in
multiple
jurisdictions,
is
a
pretty
daunting,
operational
challenge
and
especially
without
taking
in
a
doubt,
downtime
right.
So
these
are
the
problems
that
we
show
a
plan
to
show
you
a
solution
for
today
using
the
test.
So,
let's
start
out
by
talking
a
little
bit
about
witters.
How
many
of
you
are
familiar
with?
What
Metis
is?
B
Okay,
so
roughly
about
25%
30%
folks?
So
let
me
tell
you
what
what
witness
basically
does?
How
many
of
you
are
familiar
with
that
terminal?
Is
that
the
term
shard?
Okay,
almost
everybody
knows
what
a
shard
is
right.
So
witness
is
a
sharding
middleware
that
sits
between
your
applications
and
multiple
shards
of
my
sequel
databases.
It
gives
your
application
the
view
it's
talking
to
a
single
monolithic
database,
so
that
and
your
application
doesn't
have
to
be
aware
that
it's
talking
to
a
Charlotte
database,
the
way
it
does
it
is
because
of
two
things.
B
One
is
that
it
has
a
full
my
sequel,
parser
built
into
it,
and
second,
is
it
supports
the
my
sequel
binary
protocol,
so
it
presents
itself
as
my
sequel
database
to
your
application.
So
let's
take
a
quick
walkthrough
through
this
architecture.
Diagram
right.
So
everything
on
the
left-hand
side
of
the
horizontal
dashed
line
is
your
application.
B
There
are
app
app
servers.
Maybe
there
are
bad
jobs
etc.
Accessing
your
database,
everything
to
the
right
hand,
side
is
with
us,
and
my
sequel,
Questers
VT
gate
is
the
first
component
that
your
application
connects
with
VT
gate
is
a
stateless
proxy.
Has
a
full
SQL
parts
are
built
into
it
and
also
supports
the
SQL
binary
protocol
so
presents
itself
as
a
as
a
my
sequel.
Server
to
your
application.
Vd
gate
gets
the
query
parses.
It
looks
at
the
where
clause
and
figures
out
which
of
the
shards.
B
The
query
should
go
to
write
a
we
test.
Cluster
can
support
multiple
key
spaces.
A
key
space
is
basically
equivalent
to
a
database.
What
is
shown
here
is
an
N
sharded
key
space.
Each
shall
at
its
heart,
is
basically
my
sequel,
cholesterol
that
you
know
and
love
with
one
master
and
multiple
replicas
witness
makes
the
distinction
between
replicas
and
what
we
call
analytics
replicas
analytics
replicas.
We
don't
care
about.
The
replication
lag
members
of
the
analytics.
B
It
can
talk
tcp/ip
sockets
also,
but
you
typically
run
both
of
them
adjacent
to
each
other.
So
VT
gate
looks
that
it
gets
a
query
sends
it
to
a
parses.
It
figures
out
which
char
sends
it
through
via
T
tablet,
via
tablet,
parses
the
query
again
and
there's
a
whole
bunch
of
things
to
protect
the
underlying
my
security.
B
For
example,
it
adds
limit
clauses
on
your
my
sequel
query
if
it
thinks
that
this
query
is
going
to
return
more
than
10,000
rows,
10,000
is,
of
course
configurable,
but
basically,
if
it's
it's
pretty
easy
for
somebody
to
take
down
your
database
by
writing
a
query
of
death
that
pretty
much
becomes
impossible
under
the
test.
Because
of
all
these
protections
it
also
puts
gives
a
timeout
for
every
transaction
timeout
for
every
query.
It
does
connection
pooling
between
the
witty
tablet
and
my
security.
B
It
also
does
some
other
cool
things
like
hot
row,
protection
for
raids
and
for
rights.
So,
as
I
said,
you're
my
sequel
reruns
a
lot
better
under
the
care
of
witty
tablet,
then
it
would
otherwise.
The
topology
server
can
either
be
an
HDD
cluster,
a
zookeeper
cluster
or
a
cow
or
a
console
cluster.
That's
where
all
the
cluster
topology
is
is
saved,
so
how
many
key
space
are
there?
How
many
chargin
is
each
key
space?
B
How
many
mark,
how
many
replicas
in
each
char
and
what
are
the
IP
addresses
and
port
numbers
for
each
of
those
components?
So
cluster
topologies
same
there
and
a
second
and
more
interesting
information
that
is
saved
there.
Save
there
is
how
a
particular
key
space
is
Charlotte.
So
that's
the
two
type
of
information
that's
saved
in
topology
server,
VTC
TLD
is
the
control
plane
for
Vitesse.
So
this
is
sort
of
the
high-level
architecture
of
Vitesse.
What
this
does
just
to
give
you
the
context.
B
One
thing
that
is
not
shown
here
ourselves:
we
will
talk
a
lot
more
in
a
lot
more
detail
about
cells
at
YouTube,
the
biggest
key
space
that
we
had
had:
256
shards
each
shard
had
one
master
master,
of
course,
and
between
80
to
100
replicas,
distributed
across
20
data
centers,
so
between
25,000
and
13
nodes
as
part
of
a
single
key
space
application
with
giving
the
application
review.
That
is
talking
to
one
humongous
database.
B
So
all
of
you,
YouTube's
user
and
video
metadata
lived
in
in
there,
so
I'm
going
to
now
drill
down
and
talk
about
various
other
concepts
that
we
need
specifically
for
putting
together
a
cluster
that
helps
you,
but
that
helps
us
solve
the
gdpr
data
locality
problem
right.
So
the
first
concept
that
Vitas
supports
for
this
is
cell,
so
a
cell
is
basically
a
failure,
domain,
isolated
from
failures
and
other
cells.
It's
a
group
of
servers
and
network
infrastructure
example.
It
could
be
a
full
data
center.
It
could
be
a
rack
in
a
data
center.
B
It
could
be
an
availability
zone
in
a
region.
It
could
be
a
kubernetes
cluster.
Any
of
these
things,
basically,
a
collection
of
group
servers
and
network
infrastructure,
which
is
which
can
be
treated
as
a
failure
domain
and,
as
you
can
see,
these
are
physical
things
that
can
have
physical
boundaries,
which
means
that
they
can
be
located
in
in
jurisdictions
where
you
want
them
to
be.
So
that's
what
we
are
going
to
use
the
cell
for
now.
Let's
talk
about
what
a
key
space
is
so
key
space
is
a
logical
database
in
Vitesse.
B
If
you
have
a
single
shared
key
space,
that's
identical
to
a
MySQL
database,
no
different,
but
if
there
are
multiple
shards
in
a
key
space,
multiple
multiple
bicycle
clusters,
multiple
masters,
each
with
identical,
schema
right,
that's
what
a
key
spaces
and
in
either
case
a
key
space,
whether
it's
single
shadow
or
multiple
chartres.
It
appears
as
a
single
monolithic
database
to
the
application.
The
application
can
treat
it
that
way.
Now,
what's
the
key
space
ID
are
those
of
you
who
are
familiar
with
key
value.
B
Databases
probably
have
some
way
of
thinking
about
a
key
space,
ID,
basically
a
key
space
ID.
You
can
think
of
it
as
a
street
address
of
a
row.
It
decides
where
a
particular
row
lives
and
the
way
in
which
Vitesse
key
space
ID
is
different
from
the
key
used
by
key
value
stores.
Is
that
it's
a
completely
internal
witness
concept?
It's
not
stored.
B
It's
calculated
right
and
the
application
doesn't
need
to
know
about
the
key
space
idea
at
all,
and
it's
the
key
space
idea
for
a
given
row
is
used
to
decide
which
shard
a
particular
row
should
live,
live
on.
So,
let's
jump
in
and
see
how
this
is
actualized.
This
is
actualized
using
what
something
that
we
call
V
index.
So
we
have
extended
the
relational
semantics.
Urate
relational
databases
have
schemas
a
shadow
database
that
runs
under
Vitesse
has
a
V
schema.
The
V
schema
basically
is
a
collection
of
V
index
definitions.
What
is
a
V
index?
B
B
It
needs
to
define
two
pieces
of
information.
First,
what
column
in
that
table
should
be
used
for
sharding
the
data
in
that
cab
table
and
second,
depending
on
the
data
type.
Typically
of
that
column,
what
sharding
function
should
be
used
to
calculate
the
key
space
ID,
which
will
allow
us
to
decide
where
that
role.
That
row
goes
so
two
pieces
of
information
column
and
a
shouting
function
right.
So
here
is
a
quick
example
that
let's
say
that
there
is
a
table
called
customer,
it
has
a
column
named
ID.
B
B
Now,
let's
talk
a
little
bit
about
what
a
shard
is
conceptually.
So
a
shard
is
basically
just
it
has
a
begin
and
an
end
value
for
the
key
space
range
that
that
particular
shard
serves
so
for
a
single
shattered
system.
The
begin
the
beguine
value
would
be
0
0.
The
end
value
would
be
FF
FF
FF.
Basically,
what
that
means
is
any
keyspace.
Id
is
always
going
to
live
in
that
shot
right.
B
So
there
is
a
big
in
value.
There
is
an
end
value
and
for
a
given
row
when
you
calculate
key
space
ID,
if
begin
is
less
than
or
equal
to
key
space
ID
less
than,
and
if
that
condition
is
true,
then
that
particular
row
belongs
in
that
shark.
Each
shard
physically
has
one
master,
has
multiple
replicas
and
those
replicas
can
live
across
multiple
cells?
B
If
you
want-
or
they
can
be
in
a
single
cell
shouting
functions,
which
we
talked
about
a
little
bit
when
I
talked
about
windex
those
there
are,
we
give
you
six
predefined,
shouting
functions
which
are
described
here,
but
one
of
the
coolest
things
about
Vitesse
is
that
you
can
write
about
40
to
50
lines
of
go
code
and
write
a
shouting
function
of
your
own,
which
satisfies
the
interface
for
for
shouting
functions.
And
then
you
have.
B
B
Typically,
you
end
up
using
either
hash,
which
is
a
3
null
key
hash
of
numeric
columns,
or
you
end
up
using
the
unicode
lose
md5
and
on
on
basically
vark
air
or
water
binary
columns
right
that.
That's
what
you
typically
end
up
using,
but
people
have
used.
You
know,
I
mean
you
can
even
just
have
a
JSON
map
that
maps
a
value
or
the
column
value
to
a
to
a
key
space
ID
for
satisfying.
B
Typically,
when
people
migrate,
if
they
have
their
own
way
of
shouting,
that
can
be
expressed
through
a
JSON
table
and
so
on.
So
yes,
many
different
ways
of
doing
this.
So
that's
the
numeric
static
map.
So
here
we
are
going
to
use
a
custom
shouting
function.
So
in
general,
this
the
solution
is
going
to
look
like
this.
What
we
are
going
to
going
to
do
is
that
we
are
going
to
specify
the
column
in
our
a
table
that
we
will
use
for
specifying,
which
is
a
marker
of
a
location.
B
In
this
case
we
will
use
a
country
column,
but
the
the
V
index
or
the
solution
that
we
have
come
up
with
sort
of
things
of
the
problem
in
two
different
at
two
different
levels.
One
is
jurisdictions
like
NAFTA
or
EU,
or
a
country
could
be
jurisdictions,
and
the
second
is
some
marker
of
location
or
for
a
particular
user
or
a
particular
piece
of
data
right
and
a
mapping
between
the
jurisdiction
and
various
the
lower-level
values.
So
the
lower
value
value
that
we
will
use
here
is
a
country,
but
it
could
be
a
zip
code.
A
So
this
is
the
cluster
that
we
have
pre
created
in
the
in
GCP
gke.
There
are
four
kubernetes
clusters,
one
in
each.
We
have
Asia
Australia
Europe
and
the
United
States
and
using
those
clusters
we
are
already
running
with
us.
With
these
two
key
spaces.
We
have
a
user
DB
key
space
which
is
sharded
and
it
has
four
shots
and
we
have
an
auxilary
key
space
called
look
up
which
is
uncharted,
which
is
what
we
will
use
to
do,
the
routing
of
the
data.
A
Let's
take
a
look
at
what
our
schema
looks
like
so
in
the
user
DB
key
space
we
have
a
single
table
called
customer,
and
the
customer
table
has
an
ID,
a
name
national
ID
and
a
country.
So
national
ID
is
our
sample
column
for
representing
data
that's
personally
identifiable,
and
that
is
subject
to
the
various
data
locality,
loss
and
country
is
the
field
that
we
are
going
to
use
to
decide
how
the
data
should
be
stored.
This
doesn't
necessarily
have
to
be
country.
B
So
the
point
that
I
want
to
make
is
that
these
are
simple
insert
statements
with
no
regard
to.
There
is
no
hint
there
about
where
the
data
should
leave.
Your
application
hasn't
changed
at
all.
Your
application
does
not
need
to
concern
itself
with
regions;
it
does
not
need
to
concern
itself
with
where
these
cells
are
located,
etc.
These
are
simples
insert
statements
that
there
are
inserting
data
and.
A
All
we
are
inserting
here
into
the
table
are
the
name
and
the
national
ID
and
the
country.
Now
that
we've
inserted
the
data,
let's
go
and
look
at
what
our
witters
control
panel
tells
us.
So
earlier
we
had
0
rows
in
this
key
space.
Now
that
we
have
inserted
the
data,
let's
go
and
look
at
what
that
looks
like.
So
we
want
to
look
at
the
user
DB
key
space
and
in
shard
40,
which
is
located
in
East
Asia.
A
We
have
four
rows
now
and
if
we
look
at
shard
42
ad,
which
is
located
in
Australia,
we
also
have
four
rows
of
data,
and
we
could
do
this
for
the
remaining
two
shards
and
we
will
see
the
same
thing.
But
let's
now
take
a
look
at
this
data
in
the
database
directly,
instead
of
going
through
a
UI,
so
I'm
going
to
connect
to
the
key
space
or
the
database
using
VT
gate,
and
this
is
giving
me
a
unified
view
of
the
four
different
shards
for
different
physical
databases
that
we
have
so
I'm.
A
Just
going
to
say
show
me
everything
from
the
customer
table
and
what
I
get
here
is
a
scatter
gathered
query.
The
query
was
routed
to
all
four
shards
and
the
results
were
gathered
back
together
by
VT
gate
and
presented
here,
and
this
is
a
regular,
my
sequel
client
that
I'm
using
because
VT
gate
implements
the
binary
protocol.
Now,
let's
look
at
the
individual
shards,
so
the
syntax,
for
that
would
be
I
would
say,
use
this
database.
A
A
And
let's
see
what
we
have
in
that
database,
so
shard
4280,
which
is
our
australian
shard,
only
has
data
that
belongs
to
Australia
and
New
Zealand.
It
doesn't
have
data
from
any
other
countries
which
don't
belong
to
the
region.
The
other
thing
I
want
to
show
you
is
what
is
present
in
the
mapping
table.
So
how
did
we
actually
do
the
mapping
of
the
data
from
a
country
to
a
region?
A
A
A
A
A
C
A
In
general,
you
cannot
change
the
columns
based
on
which
you
have
done
the
sharding,
but
recently
we
have
built
a
new
feature.
A
new
framework
in
witters
called
read
application
using
which
you
can
change
the
value
of
the
sharding
column,
but
you
are
effectively
doing
a
reading.
So
it
is
not
a
simple
update,
it's
a
more
complex
workflow,
but
you
can
relocate
the
data
from
one
region
to
another.
You
can.
B
The
second
question
was
transactions
transactions
right,
so
what
witness
is
really
good
at
is
co-locating
data
that
needs
to
be
together
in
a
single
shard.
So,
for
example,
we
have
that
the
customer
table
which
is
charted
using
ID.
Let's
say
that
you
have
orders
table
that
has
its
own
ID
column,
but
it
also
probably
has
a
another
column
called
customer
ID
identifying
the
customer
who
is
placing
that
order.
So
we
witness
allows
you
to
say
shard
my
cus,
my
orders
table
not
using
its
ID
column
but
using
its
customer
ID
column.
B
So
the
typical
example
that
I
give
is
that
you
have
a
payment
app
and
you
have
your
shower
and
using
users
and
one
user
give
gives
other
user
money
created
from
one
road
day
between
two
another
row.
As
a
part
of
single
transaction,
those
two
users
can
either
be
in
the
same
shard
or
can
be
in
two
different
shots.
Let's
take
the
example
when
they're
in
two
different
shots
right,
so
what
we
allow
you
to
do
is
we
allow
you
to
use
two
PC
transactions?
We
have
implemented
two-phase
protocol
in
witness.
B
Where
do
we
do
a
prepare
and
charmin
preparing
shot
to
commit
in
Charlotte
one
convert
in
shark
two,
and
we
write
the
redo
logs
in
a
sidecar
database
so
that
before
the
after,
let's
say
that
between
commit
and
chard
one
and
commit
in
shard.
If
shark
two
goes
down
and
you
need
to
fail
over
to
another
replica,
the
redo
logs
are
in
the
sidecar
database,
which
have
been
replicated
into
the
replicas
and
then
after
the
new
master
comes
up.
We
can
decide
either
to
rollback
that
transaction
or
to
let
it
succeed.
B
D
B
D
D
A
What
that
does
is
you
can
choose
how
many
bytes
of
the
key
space
ID
you
want
to
allocate
to
the
region
and
for
each
country
we've
given
a
bite
value
and
the
number
of
bytes
that
were
chosen
in
this
case,
one
byte
for
the
region.
We
basically
write
this
value
into
that
byte
of
the
key
space
ID
and
the
key
space
rank.
The
key
range,
the
full
key
range
from
0
to
FF,
is
divided
into
four
equal
parts,
so
this
byte
value
then
determines
which
fourth,
it
belongs
to
and
the
way
we
chose.