►
Description
Join us for Kubernetes Forums Seoul, Sydney, 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
Intro: Vitess - Dan Kozlowski & Sugu Sougoumarane, PlanetScale
Vitess has been a member of CNCF since early 2018, and is continuing to gain adoption as the Cloud-Native database solution. This talk will focus on the basic features of vitess that make it cloud friendly. We'll also cover how it addresses indefinite scalability.
To Learn More: https://sched.co/GrcK
A
What
I've
done
is
I
have
stood
up
a
nap
in
front
of
the
Vitas
cluster
that
all
of
you
can
get
on
to
and
start
use
you
people
to
generate
some
load
while
I
do
all
these
maintenance
operations
and
we'll
see
if
it
survives
I
when
I
was
testing
it
I
did
see
a
few
transient
errors,
so
you
may
see
those,
but
eventually
things
will
go
fine,
I
think
so
we
will
see
how
that
goes.
So
that's
basically
the
intent
of
this
presentation.
A
Just
a
but
I,
will
give
you
a
five-minute
intro
of
what
we
test
is
for
those
of
you
who
don't
know
what
it
is.
I
am
the
co-creator
of
it.
Tests
I've
been
working
on
it
since
2010
when
we
started
the
project
at
YouTube.
Now,
actually,
after
we
test
moved
to
C&C,
f,
I
have
now
started
a
company
to
provide
commercial
support
for
with
test,
which
is
a
planet
scale.
A
So
what
is
with
us
in
very
simple
terms?
It's
basically
a
middleware
for
my
sequel
for
sharding,
but
it
can
basically
scale
massively
you
can.
Essentially,
there
is
no
known
limit
to
the
to
how
far
you
can
take
with
us.
At
least
there
are
companies
that
have
pushed
it
in
all
kinds
of
directions.
Millions
of
QPS
and
we
Tess-
has
so
far
kept
meeting
their
needs,
and
the
other
thing
is,
it
is
also
highly
available.
A
You
can
comfortably
get
five
nines
of
availability
would
be
tests
and,
more
importantly,
we
test
was
built
to
live
in
the
Google
Cloud
work
and
eventually
that's
those
same
abilities
were
ported
for
kubernetes
when
kubernetes
was
first
released
if
two
years
ago,
in
some
respect
with
this,
was
kind
of
a
pioneer
of
being
able
to
show
the
ability
to
run
a
stateful
set
under
kubernetes.
As
a
matter
of
fact,
the
the
oldest
with
tess
instance
has
been
running
on
kubernetes
for
over
two
years.
A
Now,
as
far
as
I
can
tell,
and
that
at
pretty
high
QPS
very
high
load
and
has
been
a
company
called
stitch
labs,
they
were
the
first
ones
to
do.
This,
I
think
it
was
2016
when
they
went
into
production
and
they've
been
still
they
are
still
on
kubernetes.
As
far
as
I
can
tell.
So
that's
a
quick
background
of
what
we
test
is.
These
are
some
of
the
stats.
A
These
are
companies
that
are
actually
betting
their
business
and
relying
on
the
fact
that
we
tests
will
not
go
down
on
them,
be
like
these
are
the
level
of
reliance
on
which,
if
we
test
went
down
on
them,
the
company
goes
out
of
business
kind
of
reliance.
So
that's
a
pretty
heavy
stamp
of
approval,
for
us
slack,
for
example,
has
about
20%
migrated,
and
now
they
are
accelerating
their
migration,
so
they
will
be
getting
200%
very
soon.
A
How
many
of
you
we're
at
cube
con
Shanghai,
none
of
you,
okay,
oh
you
were
in
cube
con
Shanghai,
JD
dot-com
is
actually
a
huge
company
in
in
China.
They
are
kind
of
the
Amazon
of
China
and
they
have.
They
are
running
they've
been
running
with
us
at
scale
at
massive
scale,
over
10,000
VT
tablets,
and
they
were
kind
of
the
highlight
of
cube
con
when
when
it
was
take
taken
place
in
Shanghai
and
all
of
you
know
square
and
a
very
interesting
tidbit
is.
This
is
a
very
recent
tweet
by
mr.
Jack
himself.
A
As
you
can
see,
the
squares
cash
app
made
it
to
the
top
charts
in
iOS
and
Android,
and
the
cash
app
now
fully
runs
on
with
us
when
they
first,
they
first
migrated
their
Uncharted
database
to
Vitesse
and
then
charted
with
us.
Underneath
from
what
I
hear
is
they
are
growing
so
fast
that
they
are
shouting
every
week
is
what
I'm
hearing?
A
So
yes,
all
right,
so
this
is
kind
of
the
community
that
that
we
test
has
actually
it's
a
lot
bigger
than
this.
These
are
the
people
we
have
permissions
from
to
use
their
logos.
If
you
look
at,
if
you
join
into
slack,
you
will
see
bustling
activity,
you
can
check
out
people
from
all
kinds
of
big
companies
migrating
to
Vitesse
and
not
only
migrating
from
my
sequel
to
Vitesse.
There
are
people
migrating
from
to
Vitas
Cassandra
to
Vitesse,
so
we
are
beginning
to
see
something
that
I
have
not
seen
before,
which
is
uncomfortably
exciting.
A
This
is
the
high
level,
so
the
negative,
the
rest
few
slides
I
will
gloss
them
over
very
quickly
because
I
want
to
get
to
the
demo
and
give
myself
enough
time
for
it.
This
is
the
high
level
architecture
of
it
s
you,
the
takeaway
from
this
architecture,
is
that,
on
the
left
hand,
side,
the
app
servers
connect
to
this
proxy
called
VT
gate
and
they
don't
care
about
what
is
behind
VT
gate.
A
They
just
treat
we
de
gate
as
if
it's
a
giant
database
PT
gate
speaks
the
my
sickle
protocol,
so
you
connect
to
VT
gate
as
if
you
are
connecting
to
a
my
sequel,
server
and
just
issue
queries
and
mitigate
we'll
figure
out
how
to
fetch
the
data
for
you
and
behind
the
scenes.
That
is
a
lock
server
that
is
used
internally
by
with
us
for
discovery
and
for
management.
So
this
is
a
very
fast
and
quick
overview
of
how
we
test
does
what
it
does.
A
This
is
a
kind
of
view
that
you
get
is
that,
once
you
have
written
your
app
server
to
use
with
tests,
then
you
mostly
don't
have
to
change
it.
While
you
things
do
well,
you
do
things
underneath
to
the
database,
for
example,
you
add
replicas
and
with
tests
will
automatically
load
balance
against
them.
The
demo
that
I'm
going
to
show
is
going
to
show
he's
going
to
go
through
these
workflows,
hopefully
we'll
get
to
the
end
before
time
runs
out.
A
So
the
first
thing,
usually
you
do
when
you
run
into
scalability
issues,
is
split
your
database
into
smaller
parts.
You
basically
take
a
few
tables,
move
them
out
into
a
separate
database.
So
that's
the
first
thing
that
you
get
help
with
we
test
will
do
that
migration
for
you
without
affecting
the
application,
as
I
said
before,
and
then,
but
this
type
of
approach
goes
only
so
far,
because
eventually
some
of
your
tables
are
going
to
get
too
big
and
you
have
to
shard
them.
A
A
As
an
engineer
has
straight-
and
you
will
you
will
you
will
see
it
when
you
see
the
UI,
that
it
was
not
a
design
person
alright,
so
let's
get
to
the
demo
so
before
I
jump
into
the
demo,
I
will
show
you
what
I'm
going
to
do.
So
if
you
go
to
the
Vita's
website
and
go
to
examples,
slash
Elm,
you
will
see
a
bunch
of
EML
files.
They
are
all
numbered.
So
if
you
launch
mini
cube
and
execute
these
yeah
Mel
files,
one
after
another,
that's
essentially
what
I'm
going
to
do.
A
The
demo
and
these
files
actually
give
you
the
full
sequence
that
I'm
going
to
show
here,
but
I
have
changed.
So
these
EML
files
are
tuned
to
run
on
mini
cube.
I
have
made
some
modifications
to
that
far
to
those
files
so
that
they
can
run
real
kubernetes
instance
like
what
I,
how
you
export
the
service
and
how
you
use
the
resources
and
stuff,
so
they
are
more
expanded
and
the
instructions
for
this
are
in
the
Vita
store.
A
I
your
website,
the
brand-new
website
that
we
have
built
for
3.0
and
you
can
follow
these
instructions
right
here
under
tutorials,
ok,
so
for
for
the
sake
of
time,
I
have
already
done
the
first
launch.
As
you
can
see.
Let
me
scroll
back
is
the
text
visible,
should
I
make
it
a
little
bigger
I,
think
I
will
make
it
slightly
bigger
alright.
A
So
if
you
scroll
back
up
what
I
did
was
I
did
a
helm,
install
of
the
first
script,
which
is
101,
which
is
script
101.
So
what
the
script
has
done
is
actually
brought
up.
My
cluster,
which
I
can
show
you
here,
you
can
see
a
few
parts
that
have
come
up
and
you'll
also
see
about
ten
parts
for
the
Commerce
application.
Actually,
it's
better
to
show
them
here
and
I
get
jobs
in
parts
you
see
there
are
ten
application
parts
they
are
stateless.
A
They
just
redirect
convert
your
requests
into
queries
and
there
are
three
global
EDD
instances.
There
is
one
at
Selene
cents
per
zone,
there's
a
net
CB
operator
which
I
use,
and
there
are
those
components
that
I
showed
you
in
the
architecture
diagram.
The
three
VT
tablets
are
the
Commerce
sweetie
tablets.
So
if
you
go
to
the
Vitesse
console,
it
might
have
locked
me
out.
Let's
see
if
it
allows
me.
A
Yes,
it
didn't
lock
me
out.
Okay,
so
you
can
see
that
there
is
a
Commerce
database.
If
you
look
at
the
schema
for
the
database,
there's
order,
customer
and
product
order
is
called
C
order,
because
order
is
a
database
keyword,
SQL
keyword,
so
it
does
not
like
such
names.
I
mean
you
can
make
it
use
such
name
so,
but
it's
very
inconvenient
because
you
have
to
keep
escaping
it.
I
will
show
you
the
gamal
file
for
this.
A
Here's
the
amel
file,
so
the
so
I'm
using
helm,
charts
to
show
this
demo
and
what
the
help
charts
have
done
is
actually
abstracted
all
the
pods
and
stuff
away.
But
then,
given
you
the
ability
to
talk
the
Vitesse
language
you
can
see
here,
I
have
something
called
a
topology:
I
have
cells,
so
these
are
elements
of
width.
S
and
I
have
created
one
key
space,
which
is
my
commerce
key
space
and
I
say
there
is
only
one
shot,
because
this
is
an
uncharted
key
space.
A
There
are
a
few
replicas
that
I've
clear
that
I
have
specified
and
my
schema
I
have
given
it
right
in
line
here.
You
can
see
the
three
tables
product
customer
and
see
order.
I've
also
defined
a
V
schema.
A
V
schema
is
also
a
Vita
schema
or
another
way
to
say
it
is
basically
your
sharding
schema.
It
tells
with
us
how
the
database
is
charted
underneath
and
that's
what
we
test
use
to
abstract
the
details
of
riding
away
from
the
application.
A
A
A
Please
don't
try
to
buffer
overflow
me
it'll,
probably
break,
but
as
long
as
you
give
reasonable
inputs,
it
should
be
here
fine
I'm,
going
to
start
off
by
entering
myself
as
customer
I
am
customer
ID
number
one
and
feel
free
to
start
creating
what
I
have
done
is
these
tables
will
show
the
last
20
rows
that
were
created
because
I
was
I,
don't
know
how
aggressive
you
are
going
to
get
about
inserting
rows
because
you
could
do
enter
until
interment
create
lots
of
rows.
So
oh
looks
like
some
people
have
started
creating
their
customer.
A
So
ideally,
a
good
application
will
enforce
relationships,
which
means
that
if
you
enter
a
skew
it
will
make
it
has
to
make
sure
that
there
is
a
product
with
that
SKU
in
it.
This
one
does
not
because
I
didn't
have
that
time
to
do
that,
but-
and
it's
not
really
necessary
for
this
demo,
but
as
I
go
through
the
demo,
you
are
welcome
to
keep
hitting
this
application
and
generating
load.
So
the
next
thing
I'm
going
to
do
is
split
this
database
into
two,
because
we
want
to
do
the
recharging.
A
A
A
The
first
upgrade
command
that
you
issue
is
201,
which
is
the
next
one.
So
the
first
step
I'm
going
to
do
is
say
that
I'm
going
to
create
a
crease
key
space
called
customer.
So
it's
an
empty
ski
space.
It's
just
a
placeholder
if
you
actually
went
and
saw
into
the
dashboard
customer
has
shown
up
here
as
a
key
space,
but
it
has
no
shards.
It
has
nothing.
The
next
thing
I'm
going
to
do
is
create
one
shard
for
that
customer
and
also
make
it
do
some
aliasing.
A
So
the
what
we
are
going
to
do
is
the
customer.
Key
space
is
going
to
be
created.
We
are
going
to
add
tables
the
customer
and
order
tables
to
that
key
space.
But
when
the
application
sends
queries,
they
are
still
going
to
go
to
commerce,
which
is
the
place
where
the
original
tables
are.
So
when
that
happens,
what
we
are
going
to
do
is
under
the
covers.
We
are
going
to
copy
the
contents
of
customer
and
order
into
the
customer
database
and
set
up
continuous
replication
from
commerce
to
customer.
A
So
once
the
replication
has
fully
caught
up,
we
can
then
say
just
cut
over
the
traffic
from
one
to
the
other,
and
hopefully
the
application
should
see.
No
change
at
all.
So
I
have
to
see
these
parts
take
a
while
to
come
up.
Let's
see
if
they
have
come
up,
they
are
coming
up,
it's
4
of
6.
It
should
go
to
6
of
6
when
it
is
done
and
in
the
meantime,
I
will
show
you
what
the
Vita's
console
shows.
It
says
that
now
customer
has
one
shard,
which
is
short.
A
0
0
usually
means
uncharted,
but
you
can
also
mark
it
as
charted,
so
it
allows
us
to
go
back
and
forth
in
terms
of
testing,
and
these
are
the
I'm
basically
having
three
one
master
and
two
replicas
as
a
setup
for
this,
let's
see
if
our
sword
has
come
up,
it's
mostly
come
up.
I
think
the
next
step
should
work.
Let's
see
so
the
next
step
is
to
say
start.
This
process
of
reach
Rd,
which
is
to
a
vertical
split
I,
will
show
you
what
the
command
is
while
it
executes.
A
If
you
look
at
the
vertical,
let
me
see
how
is
the
load
going
and
we
all
been
really
doing
load
generation,
whoa,
all
right,
59
and
244
awesome!
Ok!
So
when
I
do
the
vertical
split
you
can
I
will
show
you.
The
command
here
keeps
pace,
so
here
I
am
saying
that
perform
a
vertical
split
clone,
which
means
that
clone
the
customer
and
order
tables
from
the
customer
from
the
Commerce
into
customer.
So
this
is
the
command
and
once
I
issue
this
command.
It
should
hopefully
finish
soon.
A
Let's
see
what
it
is
done
so
that
job
ok,
it
says
what
VT
vertical
a
VT
worker
vertical
split
is
successful,
which
means
that
it
has
finished
the
clone.
It
has
done
the
copy
and
it
is
now
doing
replication.
So,
as
you
insert
rows,
all
those
rows
are
actually
being
copied
into
the
customer
database.
Now
I
am
going
to
start
doing
the
cut
over,
so
the
cut
over
is
204
vertical
migrate,
replicas,
so
the
cut
over
is
done
one
step
at
a
time.
A
So
remember
these
are
things
that
you
do
with
hundreds
of
terabytes
of
data.
So
all
these
steps
are
things
that
happen
over
a
period
of
many
days,
so
I'm
just
crunching
them
all
within
a
few
minutes.
So
at
this
point,
I'm
saying
the
the
traffic
is
being
cut
over
all
the
replicas
traffic
now
is
being
moved
over
to
customer
and
I
can
show
you
using
my
sequel.
Actually
it'll
be
a
cool
thing.
Let's
see
if
we
can
do
that,
use
customer.
A
A
A
A
All
right,
so
what
we
have
done
is
now
that
if
you
now
go,
look
at
the
database
here
in
the
dashboard,
you
will
see
that
commerce
still
has
customer
and
product,
but
you
cannot
write
to
it
anymore,
but
customer
has
also
customer
and
order
you.
But
you
can
write
to
all
your
rights
are
now
going
to
customer
and
order,
and
at
this
point
I
can
go
delete
the
tables
from
commerce
because
they
are
not
useful
anymore,
which
is
what
this
cleanup
step
does,
which
is
206.
A
A
So
these
are
logical
foreign
keys.
These
are
to
be
enforced
by
the
application.
So
we
test
assists
you
but
doesn't
enforce
it.
The
there
is
a
big
debate
about
foreign
keys,
because
many
people
shun
it.
There
are
some
good
aspects
of
foreign
key,
some
bad
aspects
of
foreign
key.
We
tests
exposes
the
good
ones
at
least
what
we
believe
are
the
good
ones,
which
is
the
efficiency
of
reusing
indexes
when
there
is
a
foreign
key
relationship,
but
enforcing
a
foreign
key
potentially
is
very
expensive.
Therefore
we
test
doesn't
reinforce
them.
A
A
Yeah,
oh
ok,
ok,
I
won't
reveal
the
number
because
you
will
hit
it
ok,
so
the
next
step
is
so.
This
is
kind
of
an
easy
part.
This
is
something
that
you
could
brute-force
manually
moving
tables
from
one
database
to
another.
You
can
kind
of
brute
force
it
using
some
kind
of
my
sequel
filter
replication,
but
the
next
step
is
to
rashard
the
customer
database
split
those
tables
into
smaller
parts.
A
So
for
that
what
we
are
going
to
do
the
first
step,
the
first
it's
a
step
itself
is
a
problem,
because
if
you
saw
those
tables,
they
have
my
sequel
order
increments.
How
is
auto
increment
going
to
work
in
sharded
environment
where
tables
are
spread
across
multiple
shots,
so
it
doesn't
work.
So
what
we
did
in
width
s
was
provide
an
equivalent
feature
using
a
sequence
table
that
issues
monotonically
increasing
numbers.
So
we
are
going
to
do
that
step,
which
is
basically
migrated
from
mice
equals
Auto
increment
to
using
Vitesse's.
A
Thousand,
oh,
it
is
thousand,
let's
see,
let's
see
now
thousand
is
very,
it
went
somebody
insert
something:
10,000,
okay,
so
yeah
IIIi
was
worried
that
you
will
hit
over
a
thousand,
so
I
made
it
ten
thousand
so
yeah,
so
I
started
the
sequence
off
at
ten
thousand
so
that
there
are
no
overlaps,
and
now,
if
you
actually,
you
can
see
that
it
is
now
counting
up
from
ten
thousand.
But
these
rows
are.
A
These
rows
currently
are
still
going
to
an
uncharted
database,
but
we
are
going
to
split
it
and
it
will
continue
to
behave
the
same
way
so
the
next
step.
Now
that
is
there
is
an
interesting
thing
that
to
that
I
want
to
show
in
the
amyl
here
is
now
the
sharded
V
schema
is
not
a
simple
JSON
anymore.
It
is
a
lot
more
complex
because
we
are
now
expressing
a
very
complex
deployment
scheme.
To
with
this,
we
are
saying
that
these
tables
are
sharded.
A
This
customer
table
is
charted
with
us
does
not
believe
in
the
word
sharding
key,
even
though
we
use
it
and
we
like
to
call
it
the
primary
windex,
because
we
are
a
database
system.
It's
kind
of
like
a
primary
key,
but
it
is
a
primary
sharding
key
and
but
then
the
primary
index
is
capable.
You
are
the
best
part
of
the
primary
of
of
index
is
that
you
can
provide
custom
functions
for
it,
and
the
other
cool
thing
that
we
have
done
is
in
the
customer
table.
I
have
said
the
customer.
A
Id
is
your
primary
index
in
the
order
table,
it's
a
gain
customer
ID,
which
is
the
primary
index.
Even
though
order
ID
is
the
primary
key
for
order.
The
customer
ID
is
the
sharding
key
for
order.
So
what
what
that
gives
me
is
to
co-locate
rows
of
a
customer's
orders
with
the
customer
itself,
which
means
that
when
I,
do
a
joint
between
customer
and
order,
it'll
be
a
single-shot
query
that
entire
join
will
be
sent
down
to
one
shot.
A
If
you
have
transactions
that
affect
both
customer
and
order,
they
will
also
be
single-shot
transactions
and
therefore
efficiently
executed.
So
this
is
something
that
I
will
try
to
show
you
an
example
data
later
about
whatever
how
you,
whatever
you
created,
ended
up
in
different
charts
and
and
with
rows
that
are
related
still
living
together.
A
Alright,
so
this
is
now
apply.
It
should
be
applied
I'm
not
going
to
check
I'm
going
to
assume
optimistically
that
it
is
done.
I
am
now
going
to
fire
up
the
shards.
This
takes
a
while,
because
six
parts
need
to
come
up
and
those
part
need
to
create
my
sequel
instances
and
initialize
them.
So,
let's
see
you
will
see
a
lot
of
units
that
are
so
what
so,
what
I
am
now
doing
is
actually
the
way
you
reach
shard
in
with
us.
Is
you
don't
touch
the
original
shot?
A
The
original
shot
remains,
as
as
it
is,
you
create
brand
new
shots
which
are
your
target
shots
into
which
Vitesse
is
going
to
move
data
and
the
way
it
does
it
is
by
splitting.
So
when
is
it's
again
two-step
process,
when
you
do
a
split
clone
is
basically
copies,
the
data
divides
them
into
two
or
or
three
or
n,
any
any
number
that
you
want
and
then
copies
different
portions
to
the
different
targets
once
they
are
all
copied.
A
It
sets
up
replication
from
source
to
all
those
destinations,
but
in
a
filtered
fashion
it
will
only
send
the
row
that
was
affected
to
the
shard
that
it
belongs
to
and
again-
and
this
is
again
very
similar
to
vertical
slip
split,
but
it
happens
for
recharging.
Also
with
us
can
do
merge
shards.
If
you
think
that
your
shards
have
become
too
thin,
you
can
also
merge
them
back.
A
So,
let's
see
if
they
are
running
it
just
looks
like
it's
almost
ready
four
of
six,
and
let
me
see
if
I
can
show
you
the
Yammer,
for
that
you
can
see
the
the
Yamal
keeps
growing
and
shrinking
depending
on
what
we
do.
So
the
cool
thing
about
this
help
chart
is,
you
can
create
jobs
and
then,
in
the
next
yeah
Mel
you
delete
that
job.
You
create
a
new
one.
A
It'll
also
cause
the
old
job
to
go
away,
so
it
keeps
your
kubernetes
instance
pretty
clean,
so
you
can
see
that
some
of
the
jobs
that
I
launched
earlier
are
not
there
anymore,
like
those
vertical
splits
and
horizontal
splits.
They
don't
exist
anymore.
Let's
see
where,
if
we
stand,
okay,
we
are
all
ready
to
do
the
sharding.
How
are
we
doing
on
the
load
test?
This
is
when
you
should
pound
the
database.
A
Okay,
I
am
ready
to
start
the
shouting.
The
first
step
is
actually
non-intrusive,
because
I'm
just
going
to
do
the
split
alone,
it's
going
to
copy
all
the
data
in
the
background,
while
while
it
is
there,
so
there
is
nothing
affected,
there
is
no
cut
over
or
anything.
So,
let's
see
how
long
that
process
takes
no
it's
already
done
and
which,
let's
then
start
migrating.
A
Let's
do
3
or
4
replicas
traffic
is
migrated
and
let's
see
if
it
is
done,
it
is
quickly
done
alright.
So
now
is
now.
If
you
migrate,
you
may
see
some
errors
because
we
are
now
going
to
move
from
short
0
to
shards,
80
and
80.
So
if
you
ask
why
80
and
why
not
50
it's
because
we
are
hexadecimal
8-
is
the
middle
for
hexadecimal
1,
2,
3,
okay,
go!
A
A
A
Oops
whoa
1672
rose.
Somebody
placed
a
lot
of
orders,
so
this
is
actually
was
a
single
shot
query.
So
we
test
knows
that
these
these
customer
and
order
rose,
lived
together
and
sent
that
query
to
just
that
one
shot
so,
and
you
can
just
so
that's
the
cool
thing
about
Vitas
is
you
can
just
bring
up
with
us
and
then
bring
up
my
sequel
console
connect
to
it
and
start
issuing
random
queries
like
this
cool
all
right,
and
that
then,
after
this
there
is,
there
are
a
few
more
steps
that
you
performed
to.
A
A
A
It's
basically,
how
wide
can
you
shard
so
typically
I
have
seen
that
a
single
my
sequel
instance
can
ingest
about
three
thousand
inserts
insert
statements
per
second,
if
you,
if
you
actually
group
those
inserts
into
single
statements,
it
does
go
much
faster,
but
since
there
is
no
limit
on
the
number
of
shards
there
is,
there
is
essentially
no
limit.
So
we
are
running
some
benchmark
tests
with
our
customers
and
we
are
seeing
like
200,000
QPS
of
inserts,
and
that
is
just
on
basically
on
the
hardware
that
we
have
provisioned.
A
So
the
load
data,
a
local
in
file-
is
a
feature
that
is
being
worked
on.
There
is
actually
a
pull
request,
that
is,
that
I
have
held
off
on,
because
I've
told
them
that
they
need
to
write
tests.
That
is
one
thing
that
we
really
really
insist
on
with
with
test,
because
the
the
way
we
test
works
is
we
push
code
today
in
production
tomorrow,
slack
is
going
to
import
that
and
push
it.
We
submit
code
into
github
today
on
master,
then
three
days
later
slack
pulls
that
code
and
applies
it
on
their
production.
A
So
we
are
very,
very
insistent
that
whatever
feature
you
write
will
work
and
has
all
the
tests
that
are
required,
and
it's
the
case
with
almost
every
Vitas
user
today,
YouTube
also
within
a
week
of
getting
something
getting
submitted,
gets
pulled
into
YouTube
and
goes
into
production.
So,
and
sometimes
they
say,
why
do?
Why
do
you?
Why
don't
you
have
more
releases?
We
don't
like
every
one
of
our
submits
to
master
is
a
release.
So
that's
how
you've
been
operating.
Yes,.
A
Yes,
yes,
we
test
can
run
on
any
platform.
Essentially,
it
runs
on
Prem
in
draw
platform,
it
runs
on
slack
uses
it
on
AWS,
but
not
without
not
with
kubernetes.
They
run
it
on
their
ec2
instances.
It
even
runs
on
mesas.
It
runs
on
whatever
what
else.
Yes,
it
is
very
customizable
and
it
can
run
on
any
hardware.
A
Yes,
I
recently
created
a
VT
tablet
modes
section
in
with
s
which
actually
tells
you
how
to
make
the
tests
how
to
point
with
tests
at
an
existing.
My
sequel
instance,
you
can
just
it's
a
middleware
right,
so
you
can
bring
up
with
test
and
say
here:
is
the
my
sequel
serve
traffic
and
then
you
can
go
back
and
forth?
You
can
send
traffic
to
my
sequel
directly
or
you
can
send
it
through
with
us
and
you
can
that
way,
slowly
migrate.
If
you
already
have
a
working,
my
sequel
instance.
A
Oh,
where
does
it
say?
Five?
Five?
Oh,
that's
a
lie.
That's
a
that's
with
us
claiming
to
be
five
five.
It's
a
number
that
we
have
hard-coded.
In
fact,
the
I
don't
think
we
support
five
five,
really
I.
We
support
five,
six,
five,
seven,
eight,
oh
Maria,
DB,
ten
dot,
X
all
mariadb
versions;
yes,
we
it's
probably
it's
probably
five:
five,
because
Maria
DB
ten
dot,
zero
dot.
One
was
based
on
five
five
and
we
hard-coded
it.
That's
five.
Five.
A
A
A
A
A
A
A
A
So
yes,
we
actually.
The
question
is:
how
does
it
work
with
github
Zorka
strider?
We
actually
approached
slow
me,
who
is
the
author
of
Orchestrator,
pretty
early
on
to
integrate
with
that?
So
actually
the
the
reparent
story
for
we
test
on
a
master
going
down
unplanned
fail,
a
failure
of
a
master.
He
is
actually
handled
by
the
orchestrator,
so
you
can
set
up
special
variables
in
the
orchestrator
where
Orchestrator
will
perform
the
failover
and
tell
with
a
Z
I,
have
now
repented.
Here's
your
new
master,
oh.
A
A
A
That's
a
very
good
question.
So
what
is
the
latency
overhead
of
Vitesse?
It
is
around
2
milliseconds
per
hawk,
it's
basically
two
hops,
so
it
is
one
millisecond
per
hop.
It's
a
very
conservative
estimate.
People
have
observed
lower
latencies,
but
if
you
are
planning
you
have
to
kind
of
assume
that
every
one
of
your
queries
there's
going
to
be
this
two
millisecond
additional
overhead
at
least
I
had
a
high
level
planning
it's
good.
To
keep
that
in
mind.