►
From YouTube: CNCF Storage WG GMT2017-12-13
Description
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
C
C
C
B
C
Thanks,
okay
I
want
to
reiterate
that
the
expectations
that
we
have
around
collaboration
around
discussion,
around
professionalism
around
not
personally
attacking
people
around
using
the
right
tone
and
and
being
generally
pleasant
and
enjoyable
to
to
to
have
be
part
of
these
calls
and
be
chatting
and
and
working
with
with
everyone.
And
if
I'm
not
seeing
that
I'm
gonna
pull
you
aside
and
and
and
ask
that
we
have
a
discussion
and
potentially
ask
that
you,
you
leave
the
working
group.
So
with
that
said,
any
questions
that
folks
have.
C
B
It's
not
cool
all
right,
so
my
name
is
sue.
Ku
and
I've
been
at
YouTube
for
about
eleven
years
now
and
I've
worked
on
various
scalability
and
infrastructure
projects,
and
my
last
project
has
been
with
tests
which
I'm
going
to
talk
about
so
this
this
presentation,
I
kind
of
put
it
together
to
address
many
of
the
questions
that
CN
CF
has
for
due
diligence.
You
didn't
feel
like
it
flows
well,
but
it
should
cover
all
the
information
that
you
need.
So
hopefully
it
won't
be
too
rough.
B
So
we
test
our
mission
statement
has
been
to
be
the
best
performing
and
scalable
my
sequel
solution
in
the
cloud.
Actually,
our
our
real
mission
statement
is
to
be
a
new
sequel
storage
solution,
but
there
are
a
few
items
that
we
are
missing,
that
we'll
get
to
as
we
addressed
issues
before
I
jump
into
what
we
test
is
I
wanted
to
cover
quickly.
My
sequel,
because
not
many
people
are
from
here
with
what
with
its
significance.
B
So
my
sequel
is
basically
an
AR
DBMS,
so
people
usually
choose
an
AR
DBMS
over
a
traditional,
no
sequel
system
like
is
usually
to
because
they
want
secondary
indexes,
joins
or
transactions
which
basically
make
the
application
simpler
to
write
without
these
application
has
to
take
on
these
burdens,
and
my
sequel
itself
is
by
the
by
DBM
genes.
Ranking
is
the
second
most
popular
storage
engine
in
the
world,
so
it
still
has
a
huge,
pretty
huge
following
the
one
of
the
main
disadvantages
of
my
sequel.
Is
that
it's
a
single
instance
server?
B
B
It's
actually
not
a
big
data
server
by
itself,
because
it's
usually
awesome
for
something
about
from
100
GB
to
1
terabyte
size
per
server
beyond
that.
If
you
start
adding,
if
your
database
gets
bigger
than
that,
and
you
start
to
run
into
various
issues
with
my
sequel
and
in
terms
of
cloud
support,
there's
only
these
hosted
solutions
like
RDS
and
cloud
sequel,
my
sequel
itself
doesn't
have
much
help
in
terms
of
use.
I
want
to
just
take
stock,
my
sequel
run
it
in
the
cloud.
B
There
is
not
much
help
and
we
test
actually
is
trying
to
help
solve
that
problem.
So
what
is
we
test?
It
basically
solves
three
problems
for
my
sequel.
It
helps
you
manage
a
large
number
of
my
sequel
instances
when
the
number
of
number
of,
as
you
scale,
when
the
number
of
instances
gets
really
really
big.
There
is
not
much
help
from
my
sequel
about
managing
them.
B
So
with
SLC
with
that,
if
you
actually
are
on
bare
metal
and
want
to
move
to
the
cloud
you
can
use
with
s,
to
move
your
MySQL
instance
to
the
cloud
and
the
last,
but
not
the
least,
is
that
it's
actually
gives
you
new
sequence,
tile
scalability.
When
your
database
runs
out
of
steam,
you
can
just
shard
and
continue
to
scale.
So
those
are
three
main
features
that
we
test
gives
you
in
terms
of
categorization.
B
So,
prior
to
this
whole
new
sequel,
wave,
that's
coming
up,
people
had
to
choose
between
Ava
and
our
DBMS
or
a
no
sequel,
and
is
so
basically
you
either
choose
scalability
or
you
choose
data
integrity
features,
but
you
can't
have
both,
and
that
has
been
a
problem,
because
many
of
the
complex
applications
do
want
both.
So
hopefully,
the
new
sequel
wave
will
solve
that
problem.
So
the
there
are
actually
many
new
sequel
systems.
There's
now
tid
B.
There
is
also
a
situs
I
need
to
actually
update
the
slide.
B
Some
people
actually
were
questioning
whether
with
this
is
or
should
be,
a
new
sequel
system,
and
there
are
many
definitions-
it's
still
gelling,
so
what
I've
done
is
I've
put
with
a
slightly
below
the
bar
saying
that
we
are
almost
there,
but
not
quite
there
yet,
but
it
does
play
in
that
space.
Most
people
that
want
to
scale
indefinitely
an
RDP.
My
style
system
tend
to
converge
to
be
test.
B
You
know
in
terms
of
history,
we
this
project
is
pretty
old.
We
started
actually
in
2010.
This
was
a
time
when
YouTube
was
actually
having
some
severe
scalability
problems.
Actually
the
number
of
voltages
were
growing
and
they
are
pretty
much
getting
out
of
control.
So
myself
and
my
colleague
Mike
Solomon,
we
took
ourselves
out
of
the
day-to-day
operations
and
decided
to
see
if
you
can
come
up
with
something
that
will
let
us
sleep
ahead
of
all
these
problems,
so
that
is
basically
how
a
test
was
born.
B
It
we
mostly
used
it
for
ourselves.
Actually,
we
used
to
develop
in
the
open
source
world
and
import
the
data
and
use
it
for
ourselves.
Nobody
else
actually
was
participating,
but
over
time
the
number
of
features
we
added
kind
of
became
Rhian,
civ
and
Flipkart
was
also
facing
scalability
issues,
and
they
somehow
found
us
and
contacted
us
and
said:
hey:
do
you
think
we
can
use
this?
We
said
sure,
give
it
a
try.
B
It
will
help
you
to
the
extent
we
can
and
then
a
year
later
they
went
into
production
which
was
pretty
exciting
and
then,
after
that,
that
credibility
attracted
more
users
and
it
has
been
pretty
much
growing
organically
since
then,
so
we
haven't
been
actually
there's
not
much
hype
around
with
us.
So
most
of
the
Vita's
growth
has
been
very
quiet
and
secret
almost,
but
it
is
it's
grown
pretty
pretty
fast.
B
Now
we
now
have
like
a
slack
channel
that
has
about
200
members,
and
so,
if
you
look
at
our
community,
so
these
are
the
companies
that
are
currently
using
or
adopting
with
this.
The
group
that
you
see
on
the
left
are
actually
people
that
are
already
in
production,
and
the
group
you
see
on
the
right
are
people
that
are
either
in
the
pipeline
or
evaluating.
B
Actually,
there
are
more
companies
than
this.
We
only
have
permission
to
use
these
logos,
which
is
why
I've
used
them.
There
are
other
companies
that
have
actually
some
of
them
are
actually
quite
big
that
have
gone
into
production,
but
they
haven't
given
us
permission
to
use
them
as
a
reference.
Yet
so
it's
pretty
exciting
this
part,
and
there
are
some
testimonials,
but
both
from
slack
and
squares
that
I
collected,
because
people
wanted
to
hear
from
them
about
whether
they
how
they
support
C
and
C
F.
You
can
actually
see
the
full
text.
B
I
will
link
the
document.
These
are
suggest
some
expire
accepts.
So
what?
What
are
the
things
that
we
test
does
for
you
in
the
area
of
making
my
sequel
better?
It
allows
you
to
run
my
sequel
and
multiple
data.
Centers
takes
care
of
fail,
overs,
load,
balancing
and
one
problem
that
my
sequel
has
is.
It
cannot
handle
a
large
number
of
connections
and/or.
B
A
large
number
of
or
if
there
is
a
huge
spike
in
requests,
it
just
falls
apart,
so
it
does
actually
protects
my
sequel
by
using
connect,
pools
that
limit
and
throttle
the
traffic
and
often
sometimes
expensive,
queries
hit
to
my
sequel
and
take
the
entire
database
down.
So
my
sequel
is
is
something
that
cannot
where
one
query
can
actually
affect
everything
else.
So
when
these
things
happen
with
us
intervenes
and
kills
queries
that
are
taking
too
long
or
transactions
that
are
lingering.
B
So
actually
this
we
have
been
developing
these
features
at
YouTube,
basically
based
on
issues
that
we
have
been
seeing
in
production,
and
these
are
actually
things
that
anybody
that's
using.
My
sequel
will
relate
to,
because
these
are
problems
that
all
of
these
people
face,
and
it
is
now
so
well
tuned
that
people
that
go
into
production
with
Vitesse
find
that
their
my
sequence
generally
runs
better
than
it
used
to
run
before.
B
Actually,
the
these
are
actually
not
cloud
features
per
se,
but
the
entire
retests
software
is
written
for
the
clouds,
so
there
is
actually
subtle
cloud
related
behaviors
in
all
of
these
features.
So
it's
so
that's
one.
That's
actually
my
next
slide,
there's
a
there
are
so
because
the
the
way,
actually
we
we
wrote
it
is,
we
test-
was
built
to
run
in
pork.
So
what
we
used
to
do
is
even
though
we
built
it
in
the
open
source
we
had
to
actually
import
it
and
actually
adopt
it
to
run
in
Google's
work
cloud.
B
So
we
had
to
use
lock,
servers,
discovery
mechanism,
health
check,
everything
that
Google
required
for
software
to
run
in
its
cloud.
We
had
to
do
it
for
with
this,
so
you
know
maybe
were
actually
cloud
ready
before
even
kubernetes
was
born,
so
when
kubernetes
came
about,
it
was
actually
very
trivial
to
make
it
work
under
there,
because
all
the
plugins
are
in
place
actually
running
in
a
cloud
is
not
just
checking
these
boxes,
because
there
is
a
specific
way
you
have
to
think
about.
B
Software
like,
for
example,
cannot
like
rely
on
config
files
and
expect
them
to
be
sprayed
around
to
change
the
way
software
behaves.
You
have
to
use
the
log
server
mechanism
to
publish
changes
you
have
to
rely
on.
Basically,
the
file
system
is
not
something
that
you
always
have
access
to.
You
cannot
log
into
a
box
and
look
at
what's
going
on
all
the
time
and
and
also
when
you
use
a
log
server.
You
have
to
be
careful
about
not
overloading
it,
because
it's
not
something
that
you
can
afford
to
spam.
B
Those
are
actually
low,
QPS
systems
that
you
shouldn't
abuse.
So
all
these
things
we
actually
learnt
them
the
hard
way
and
actually
changed.
We
test
to
actually
treat
these
external
systems
correctly
so
that
it
works
well.
So
the
best
feature
for
which
I
have
with
us
is
definitely
the
sharding
and
that's
the
one
that
attracts
most
users.
Basically,
once
you
once
you
outgrow
one
single
instance.
What
we
test
lets
you
do
is
actually
shard
your
database
and
continue
scaling
indefinitely.
B
So
the
way
it
does,
it
is
basically
lets
you
view
the
sharded
database
as
if
it
was
a
single
instance.
So
when
you
send
a
query,
it
figures
out
where
and
how
it
should
route
that
query
or
split
it
in
two
parts
and
send
it
to
different
parts
and
then
collect
the
results,
and
this
database
like
features
in
the
sharding.
We
have
cross
shard
indexes.
My
sequel
users
rely
a
lot
on
auto
increment
and
we
have
sequences
that
behave
just
like
those.
But
then
it
works
for
cross
short,
shorter.
B
The
other
one
is
actually
the
pluggable
sharding
scheme.
This
is
actually
something
that
my
sequel
users
prefer
I,
think
it's
something
that
most
users
should
prefer,
because
I
don't
think
systems
are
there
yet
where
they
can
be
good
at
figuring
out
how
to
share
your
data.
So
if
the
system
made
a
decision
about
how
data
should
be
charted,
it
is
most
likely
not
going
to
be
the
most
optimal
way
and
hopefully
machine
learning
will
help
us
solve
this
problem.
But
for
now
this
decision
is
best
made
by
a
human.
B
So
therefore,
we
not
only
make
you
decide
how
you
want
short
your
data.
You
also
decide
the
sharding
scheme
that
you
want
to
use,
which
is
something
that
many,
my
sequel
users
want,
because
not
everybody
likes
a
hash-based
sharding
scheme.
Somebody
wants
to
use
mod
base,
somebody
wants
to
use
an
md5
or
somebody
just
wants
a
sequence.
So
that's
one
thing
that
we
test
is
very
good
good
at
the
most
important
thing
is
that
the
sharding
scheme
and
the
odd
index
feature
actually
worked
hand-in-hand
together.
B
You
can
even
reshot
your
database
after
you
plug
in
your
charting
scheme
and
when
you
reach
out
a
database,
we
test
can
do
it
with
almost
no
downtime.
The
application
will
almost
not
know.
What's
going
on,
there's
probably
about
like
a
few
seconds
of
downtime
when
you
switch
over
the
Masters,
which
actually
the
proxies
will
just
buffer
those
requests
and
the
application
will
see
almost
no
downtime.
And
finally,
because
you
are
shorted,
you
are
probably
going
to
have
transactions
that
span
across
charts.
So
for
that
Vitas
has
support
for
two
PC
transactions.
B
There
are
limitations.
These
are
actually
all
addressable
limitations,
but
we
just
haven't
prioritized
them.
Yet.
The
one
thing
is
two
PC
is
actually
a
new
feature
that
we
developed
it's
the
first
version
and
it
is
expensive
and
but
then
somebody
that
really
really
wants
it
can
have
it,
but
they
have
to
pay
the
price.
It
essentially
results
in
about
50%
increase
in
data
throughput
requirements.
So
it
does
reduce
your
serving
capacity.
But
then,
if
you
can
shard,
you
can
just
go
wider.
B
C
B
Really
asked
for
it:
they
don't
seem
to
think
it's
such
so
so
important,
so
it
may
be
video.
We
haven't
encountered
users
that
need
this
yet
so,
but
this
is
something
that
is,
that
is
in
our
radar
and
when
so,
when
you
issue
a
sequel
and
if
all
of
that
sequel
can
be
served
by
one
one
shard,
then
we
test
is
very
good
at
figuring
it
out
that
sends
that
it
is
equal
to
one
shard.
B
B
The
enterprise
style
uses
the
one
thing
that
their
main
complaint
has
been,
that
it's
very
complex.
It's
based,
basically,
because
just
we
have
been
going
in
for
seven
years.
It
has
a
lot
of
options
and
sometimes
people
get
confused
of
what
what
works
with
what
and
so
right
now.
They
do
need
to
talk
to
us
sometimes
too,
and
they
have
probably
configuring
with
this.
So
this
is
another
big
area
that
we
have
to
focus.
Actually
these
last
two
currently
are
our
biggest
priority:
the
documentation
and
reducing
the
complexity
of
bringing
up
every
test
cluster.
B
But
the
good
news
is
that,
once
you
have
configured
with
tests,
it
runs
really
really
smoothly.
So
that's
been
the
experience
with
our
users
all
right,
so
this
is
actually
the
Vitesse
architecture.
It
shows
how
we
do
what
we
do.
The
way
we
test
is
configured
is
at
every
my
sequel
instance.
We
attach
what
we
call
as
a
VT
tablet,
which
is
actually
all
queries
go
through.
Vt
tablet
and
VT
tablet
makes
sure
that
the
queries
do
not
abuse
my
sequel.
B
There
are
lots
of
protection
like,
for
example,
if
two
identical
queries
hit,
my
sequel,
it
won't
send
them
both
to
it
it'll
send
only
one
of
them
and
then
share
the
results
to
everybody
that
has
requested
it
and
VT
tablet.
Also,
the
can
perform
housekeeping
work
like
taking
backups,
restores
and
all
the
cluster
management
work
it
can
take
care
of
and
on
this
side
on
the
proxy
level.
That
is
this
thing
called
VT
gate,
which
is
actually
stateless.
B
So
it's
something
that
can
scale
up
and
down
as
needed
when,
when
the
load
on
the
application
goes
up
and
down
and
the
application
knows
only
about
VT
gates,
it
does
not
know
about
all
the
details
underneath
so
it'll
just
send
the
query,
as
if
VT
gate
is
a
one
big
giant
database,
and
then
we
ticket
will
figure
out
how
to
get
that
query
served.
So
it
greatly
simplifies
the
view
the
application
has
about
the
cluster
underneath
and
how
do
we
tie
all
this
together
is
actually
through
a
log
server
here.
C
B
Put
the
HDD,
but
actually
with
that
supports
all
the
log
servers
it
keeps
track
of
which
tablets
are
going
up
and
down
and
inform
city
gate
of
those
changes
and
then
mitigate
accordingly
routes.
Traffic
and
VDC
TLD
is
our
dashboard
that
we
use
to
look
at.
What's
all
the
all
the
cluster
and
perform
actually
operations
like
my
parent,
English,
sharding,
etc?
B
B
D
B
Actually,
a
lot,
it's
pretty
elaborate.
We
actually,
you
can
actually
use
multiple
log
servers.
We
have
one
called
the
global
log
server,
which
is
actually
unique
across
all
data
centers,
and
we
actually
let
you
configure
one
log
server
per
data
center.
So
the
idea
is
that
when
one
data
center
goes
down,
it
does
not
take
everything
else
with
it,
and
the
global
log
server
has
very
little
information.
It
only
tells
you
where
the
master
data
base
is.
B
C
D
B
It
does
it
VT
tablet
actually
will.
When
it
comes
up.
It
will
actually
register
itself
with
that
city,
saying
that
I
am
now
here
and
I'm
a
replica
and
I'm
ready
to
serve
traffic
and
that
will
eventually
and
mitigates
our
actually
have
a
watch
set
on
this
Exedy.
So
when
it
comes
up,
they'll
receive
the
notification
and
will
start
sending
will
start
load
bar,
including
that
VT
tablet
in
the
in
the
load,
balance
right.
D
B
Actually,
we
recently
implemented
the
my
sequel
protocol
at
the
VT
gate
level.
So,
if
speed
client
could
just
connect
to
mitigate
as
if
it
was
a
my
sequel
database
and
it's
transparent,
there
is
no
difference.
Yeah
yeah,
so
yeah.
There
are
actually
some
some
differences
by
the
fact
that
things
are
shrouded
and
the
fact
that,
for
example,
the
VT
gate
here
represents
all
the
databases
here,
like
master
replicas
and
stuff.
So
the
app
server
could,
for
example,
say
I
want
to
connect
to
my
key
space,
but
I
want
to
connect
to
the
master.
B
C
B
Extensions
but
otherwise
query
wise.
It's
the
same.
You
just
said
once
you've
specified
your
database.
You
just
send
queries
like
like
normal,
and
we
also
have
some
non-standard
extensions
like
for
example,
you
could
do.
You
could
issue,
show
statements
that
return
information
about
charts
like,
for
example,
you
can
say
show
me
all
the
key
spaces
here
or
show
me
all
the
shards
inside
the
key
space,
so
a
few
extensions
which
people
can
optionally
use,
but
typically,
if
you
just
point
your
app
to
vita
gate,
it
should
work
asses.
B
B
It
lets
you
use,
bind
variables
at
your
app
player
itself,
which
is
which
is
actually
which
makes
things
a
little
more
efficient
and
not
only
that
it
actually
makes
the
this
VT
gets
fully
stateless,
which
means
that
you
could
start
a
transaction
with
one
mitigate,
and
if
that
VD
kate
goes
down
in
the
middle,
you
can
actually
complete
that
transaction
using
another
one.
So
it's
truly
stateless
and
load-balanced,
so
that's
a,
but
we
have
noticed
that
most
people
don't
care
much
about
this
feature.
So
most
of
them
just
use
the
my
secure
protocol.
B
Alright,
so
this
the
the
whole
idea
of
this
thing
is
we
actually
worked
very
hard
at
reducing
the
number
of
moving
parts
here.
These
are
things
that
we
felt
are
absolutely
necessary.
Anything
less
would
be
a
problem
or
anything
more
would
be
a
problem.
So
this
a
lot
of
thought
has
gone
into
this
architecture
and
we
actually-
and
this
is
also
being
tuned
over
many
years-
so
I
think
I
think
we
are
very
happy
with
what
we
have
here.
C
B
We
are
probably
going
to
stay
with
it
for
a
very
long
time
in
terms
of
versatility.
This
is
where,
with
this
really
shines,
it
can
pretty
much
learn
any
place
anywhere
like
right
now.
It
learns
like
YouTube
grunts
with
us
on
Borg
stitch
labs
runs
it
on
gke,
better
cloud
uses,
mesos
and
GCE
HubSpot
uses
AWS
and
kubernetes,
and
others
like
booking
and
Flipkart.
They
basically
run
on
bare
metal,
there's,
actually
a
bunch
of
uses
that
run
on
bare
metal.
It
can
also
work
with
any
loc
server.
B
We
when
we
import,
we
tests,
we
basically
swap
in
a
chubby.
We
originally
developed
it
with
zookeeper,
but
now
we
also
supported
CD
and
console,
and
all
these
are
used
in
production
by
some
one
or
the
other
and
in
the
same
way,
in
the
open
source.
We
use
G
RPC,
but
when
we
import
it,
we
swap
it
out
for
stubby,
and
there
are
plugins
that
people
have
provided
for
monitoring
with
prometheus
in
flux,
TB,
this
collecti
and
obviously
pokemon
internally.
So.
B
Stuff
that
we
spent
a
lot
of
time
in
our
software
design
Vitas
is
fully
pluggable.
So
basically,
it's
not
just
these
things
that
our
pluggable
there
is
actually
a
whole
bunch
of
things
that
we
do
when
we
import
with
us
into
Google,
we
bring
in,
we
add
throttle,
errs,
we
add
monitoring
health
checks,
ACLs
tapper,
and
actually
the
list
is
pretty
pretty
big.
This
is
these
are
some
other
big
things
that
that
I
wanted
to
mention
and
the
most
importantly,
this
plug
ability
does
not
affect
efficiency.
B
Typically,
if
you
deploy
with
tests
on
top
of
my
sequel,
the
what
people
have
reported
is
about
under
two
millisecond
overhead
added
to
what
would
have
otherwise
been
a
direct
round
trip
with
my
sequel,
which
most
of
them
find
acceptable
and
over
time
we
have
we
have
actually
when
we
started,
we
were
not
very
disciplined,
but
then
we
brought
in
some
discipline
about
coding.
So
we
now
our
coding
standards
are
pretty
high
and
stringent
readability
is
the
most
important
thing.
B
When,
when
somebody
writes
software,
we
spend
a
lot
of
time
making
sure
that
it's
structured
such
a
way
that
anybody
else
that
comes
in
and
reads
the
code
can
understand
what's
going
on,
and
this
is
actually
attracted
more
contributors
and
they've
all
been
very
happy
with
witnesses
code.
We
follow
all
go
coding
standards.
B
We
require
unit
tests
to
be
written
with
strict
coverage
requirements,
neither
is
needed,
end
to
end
tests
have
to
be
written
and
they
should
not
affect
performance
and
we
integrate
with
Travis
code
climate
and
notify,
which
tells
us
which
runs
all
the
unit
tests
or
all
the
tests.
A
unit
is
an
end-to-end
test
before
we
actually
merge
code
in
so.
B
Our
master
branch
is
something
that
we
import
from
regularly
and
deploy
on
YouTube
and
many
of
our
users
do
the
same
thing.
So
we
require
the
master
guides
to
be
absolutely
stable
all
the
time
there
are
only
one
or
two
users
that
use
an
official
release
of
witness
which
is
currently
at
2.1,
and
we
are
about
to
release
3.0
in
terms
of
I.
B
A
Thank
you
for
the
for
the
presentation.
I
think
that's
something
that
the
unity
of
see
is
gonna.
Ask
just
to
play
devil's
advocate
with
you
and
yes,
something
to
prepare
for.
Is
you
know
what
what
do
you
like?
Why
do?
Why
do
you
want
the
test
to
to
go
to
the
foundation
like
what
do
you
see
as
the
trajectory
free
to
the
test
if
it's
accepted
as
project
yeah.
B
Actually,
the
the
main
thing
that
we
are
looking
for.
Actually
this
was
the
this
is
the
reason
why
YouTube
is
pushing
is
very
insistent
what
this
happening
is
that
we
test
is
being
seen
as
a
project
owned
by
YouTube,
and
we
want
to
change
this
to
be
a
project
that
is
owned
by
the
community
by
the
fact
that
all
these
people
depend
on
it
and
they
contribute.
B
B
A
B
I,
it
will
so
basically
YouTube
is
committed
to
so.
The
way
meters
has
grown
is
the
features
that
YouTube
uses
so
actually
I,
don't
know.
If
you
have
heard
YouTube
ease,
has
made
a
decision
to
actually
migrate
to
spanner,
but
it's
a
multi-year
project
so
but
then
YouTube
is
still
committed
to
supporting
with
tests
and
therefore
it
wants
to
make
sure
that
it
grows
the
community
and
establishes
leadership.
B
B
Yeah
yeah,
so
architectural
early,
they
are
very
different
because
we
test
relies
on
an
RD,
be
me
on
a
relational
database
underneath
so
because
of
that
that
that
makes
it
a
very
different
product
product
compared
to
spanner.
However,
we
do
get
pitched
we
do
get
compared
with
the
spanner
cockroach
tid.
Be
all
these
new
sequel
databases.
B
Vitas
is
better
at
handling
SQL
as
a
language,
mainly
because
it
relies
on
and
real
lar
to
be
a
mess
underneath.
So
in
that
respect,
for
example,
it
supports
DM
ELLs,
which
allows
you
to
change
data
using
SQL
statements,
which
system
will
expand
our
and
cockroach
DB.
Don't
support
the
place
where
it
falls.
Short
of
final
is
something
that
I
mentioned
in
the
limitations.
B
Is
that
spanner
uses
an
LS
m
system
and
it
can
actually
obtain
a
consistent
cross
short
view
of
the
data
which
we
test
cannot
do
right
now,
because
for
that
you
need
to
actually
the
database.
The
underlying
database
has
to
support
that
feature
which
my
sequel
doesn't
support
yet,
but
otherwise,
if
you
are
just
coming
top
down
for
you,
there's
a
there
is
very
heavy
feature:
parity
between
spanner
and
Vitas.
They
both
scale
equally
well.
B
There
are
some
differences
in
how
we
how
we
like,
for
example,
spanner,
has
chorim's
and
they
have
cross
data
center
currents.
So
they
give
you
better
durability
for
your
transactions.
But
then
you
also
pay
the
price
in
terms
of
latency
for
it.
But
as
we
test
relies
on
a
different
mechanism
called
semi
sync
replication,
which
relies
on
within
data
center
core
own
style,
consistency
which
is
actually
faster.
D
B
It
keeps
coming
all
the
time
because
if
you
look
at
the
architecture,
nothing
prevents
us
from
actually
replacing
this
pair
with
the
Postgres
with
a
VT
tablet
that
can
handle
post
Chris.
Yes,
that
is
definitely
something
that
is
on
our
radar.
But
the
thing,
though,
is
right:
now
we
have
our
hands
full,
supporting
the
my
sequel
community.
So
once
that
pressure
goes
down
like
every
week,
we
have
a
new
user
that
come
up
and
says:
oh
I
need
help
here
and
stuff.
So
it's
been
pretty
busy
lately
just
supporting
what
we
have
now.
B
Because
if
you,
if
you
actually
looked
at
the
VT
gate
code,
all
it
does
is
look
at
SQL
by
the
SQL
standards
and
figures
out
how
that
SQL
needs
to
be
broken
out
into
smaller
parts
and
sent
to
different
instances
underneath.
So,
if
you,
if
you
actually
you
have,
what
you
would
need
to
write,
is
actually
a
new
VT
tablet
component
for
Postgres
that
can
manage
post
chris
here
on
this
side.
So
once
you
have
that
from
the
VT
gate
perspective,
it
is
just
very
routing.
The
query:
that's
right!
Thank
you.
E
C
B
B
They
said
that
both
both
the
hub
spot
and
seach
labs
also
uses
use
still
of
jesus
gke.
They
said
that
they
did
have
to
create
a
few
scripts
to
deploy
with
s,
so
so
that
part.
Some
is
something
that
we
can
definitely
do
more
work
on,
so
within
YouTube.
Actually,
it's
a
pretty
big
instance.
I
think
we
have
like
tens
of
thousands
of
nodes
and
pretty
hike
ups
as
a
set
up
like
about
20
data,
centers.
A
B
Right
now,
the
way
I
think
they
have
done
it
is
so
when
you
scale
you
actually
so
once
you
have
the
vita
setup
done
you.
Basically,
as
you
can
see,
there
are
very
few
moving
parts
here
per
se,
so
scaling
with
us
basically
involves
just
bringing
up
more
and
more
replicas
here
and
those
are
operation.
B
Does
today
correct
yeah
yeah,
so
the
reason
is
because
so
VT
gate
is
actually
auto
scale
that
scales
up
and
down
automatically
based
on
the
load
generated
by
the
app,
but
here
they
actually
have
to
provision
because
they
have
to
decide
how
much
disk
they
want
and
stuff.
So
you
are
right.
This
is
one
place.
We
could
probably
support
auto
scaling,
but
we
don't
do
that
right
now.
So
a
human
has
to
say,
oh
looks
like
our
read:
traffic
has
gone
up.
We
need
to
bring
it
at
more
instances
on
this
side
kind.
D
B
Nobody
in
youtube
has
ever
asked
to
auto
scale
these
things
that,
yes,
our
is
actually
want
a
complete
control
over
how
many
instances
this
is.
These
are
actually
planned
on
a
weekly
meeting
where
we
say
oh
looks
like
we
need
to
do.
Rashard
looks
like
we
need
to
add
more
replicas,
so
these
are
humans,
at
least
in
YouTube.
Humans
have
preferred
to
make
these
decisions.
D
E
And
then
one
other
question
events:
the
pc
tablets
configuration
so
once
you've
got
sort
of
obviously
a
very
tightly
coupled
relationship
between
the
v3
tablet
and
an
instance
of
my
sequel,
because
there's
a
beefy
tablet
handle
any
of
the
functionality
in
terms
of
actually
managing
the
lifecycle
of
that
my
sequel
instance
ie,
keeping
the
service
alive
and
or
maybe
upgrades
of
the
software
or
any
of
those
sort
of
things.
Yeah.
B
That
is
actually
meaty
tablets
job
right
now,
if
it
actually,
the
the
main
housekeeping
work
that
it
does
is
actually
taking
backups
and
performing
restores.
It
also
helps
you
with
the
schema
deployment
and,
and
then
there
is
also
coordinating
about
who
is
the
current
master.
So
those
are
the
housekeeping
work
that
we
do
tablet
does
for
my
sequel,
other
than
that.
All
it
does
is
serve
queries,
okay,.
E
B
Yes,
so
we
actually
have
you
can
actually
issue
manual,
failover,
zouri
parents
and
then
so.
The
the
one
thing,
though,
is
here,
is
that
when
a
master
failover
happens,
it's
actually,
you
have
to
actually
inform
the
VT
gates
immediately.
You
cannot
afford
to
go
through
HCB,
because
that
is
actually
the
generally
the
SLA
for
something
like
HDD
disseminating
information
is
usually
15
to
30
seconds
and
sometimes
even
longer.
So
what
the
way
we
do.
B
This
is
actually
these
VT
gates
are
connected
to
these
VT
tablets,
and
there
is
actually
a
health
check
stream
that
flows
from
all
the
VT
tablets
to
ET
gate.
So
as
soon
as
the
VT
tablet
becomes
a
master,
the
health
check
immediately
informs
the
VT
gates
and
within
seconds
all
VT
get
self
move
the
traffic
over
to
the
new
master,
so
HDD
is
mainly
used
as
passive
scary.
D
B
Stuff,
so
it
s
actually
has
a
tool
called
my
sequel
CTL.
My
stands
for
my
sequel
control,
where
we
have
simplified,
how
you
configure
my
sequel
and
actually
a
lot
more
people
than
we
expected
use
it,
because
we
thought
that
most
DBA
is
like
to
configure
their
own.
My
sequels,
but
a
large
number
of
people
actually
have
adopted
that
tool.
It
actually
makes
it
very
easy
to
spin
up
my
sequel
instances
and
the
good
thing
that
my
sequel
CTL
does.
B
The
way
we
test
works
is
like
we
run
small
instances
of
my
sequel,
so
in
a
particular
box
you
could
probably
see
20
or
30
instances
of
my
sequel
running
per
box
and
what
my
sequel
CTL
does
can
do
is
actually
isolate
all
of
them
the
default.
My
sequel
is
expected
to
be
running
like
one
instance
per
box.
So
but
then
we
have
changed
it
such
that
it
configures
it's
that
to
make
sure
that
they
are
each
isolated
from
each
other.
A
They
say
if
like
if
you
were
to
compare
this,
almost
make
a
decision
to
to
leverage
like
in
the
public
lot,
if
they're
gonna,
leverage,
Amazon's
RDS
or
in
Google's
sequel
cloud
like
what
would
the
decision
like
what
kinds
of
things
do
they
or
do
they
consider
when
they
decide
to
either
use
the
tests
or
the
public
cloud
sequel
instances?
So.
B
Actually,
we
des
can
make
both
RDS
and
cloud
sequel
better,
because
and
I
was
talking
to
somebody
from
Amazon.
They
said
that
Vitas
can
even
help
with
Aurora
it's,
because
Vitas
can
do
sharding
and
none
of
these
systems
can.
So,
if
you
put,
if
you
deploy
with
s
in
front
of
RDS
or
cloud
sequel
or
Aurora,
you
solve
the
big
scalability
problem
that
they
have
where
the
write
throughput
is
limited
by
the
fact
that
only
so
much
can
be
handled
by
one
box.
So.
D
B
Actually,
this
is
how
we
presented
it
here,
but
VT
tablet
can
run
outside
of
the
box
that
my
sequel
runs
on.
It
can
just
connect
through
a
regular
port,
I,
say
and
actually
Square
uses
it
that
way
they
actually,
they
actually
have
a
huge,
my
sequel
instance.
They
are
like
they
are
coming
from
a
monolith,
my
sequel
and
they're
slowly,
shutting
out
of
it
and
they
cannot
afford
and
they
actually
the
way
they
are
set
up.
B
A
B
Yeah
so
RDS
cloud,
sequel
or
aurora,
their
limitation
is
each
instance,
is
only
one
machine,
so
you
can
choose
the
size
of
your
machine,
but
once
you've
gone
to
the
largest
machine
that
you
can
get,
that's
pretty
much.
You
are
pretty
much
stuck
there
at
that
point
and
if
you're,
if
you
suddenly,
if
your
data
size
blows
up
out
of
control
beyond
that,
then
you're
you're
in
trouble
and
we
we
we
want
to
do
actually
RDS,
fortunately
or
unfortunately,
nobody
has
asked
for
it
yet,
but
and
we've
been
busy
supporting
the
current
community.
D
D
B
B
Every
time
they
mentioned
that
we
should
there
are
places
where
they
can
use
with
us.
Actually
I
met
the
head
of
the
era
sorry
last
week
and
he
was
still
saying
the
same
thing,
but
I
think
that
is
there's
also
the
fact
that
things
are
working
now
at
Facebook
and
there
is
inertia
about
taking
the
effort
to
replace,
what's
currently
working
with
something
else,
so
so
I
think
eventually
I
think
the
way
they
would
both
Facebook
and
others.
Other
companies
are
also
going
through.
This
is
when
they
encounter
big
problems
when
they
say.
B
A
A
Alright,
one
more
so,
regarding
you
know
the
the
interaction
with
with
my
sequel
has
anyone
you
know
who's
contributed
in
my
sequel,
expressed
interest
in
absorbing
some
of
the
features
of
the
text,
because
you
know
you
described
with
the
BT
tablet
and
it
seems
like
some
of
that
I
know
throttling
or
other
logic
that's
in
place
there.
It
could
be
beneficial
to
my
sequel
on
its
own.
Any
discussions
there
for
what
pieces
makes
sense
to
stay
into
tests
and
what
pieces
may
just
need
to
be
absorbed
into
my
sequel,
yeah.
B
I,
don't
I
thought
I
talked
to
the
Oracle,
my
sequel
team,
at
least
twice
a
year
and
force
one
reason
or
another.
None
of
them
think
that
it's
important
I
don't
know
why,
but
they
for
they
are
I,
think
they
are
still
so
there
is
some
relational
trade-off
that
we
make
when
we
do
a
connection
pool,
for
example,
right
like
you
at
that
point,
eat
you
enjoy.
Each
individual
connection
is
actually
stateless,
which
is
actually
not,
which
is
actually
a
move
away
from
a
relational
concept.
So
I
I
have
a
feeling.
B
They
don't
want
to
go
that
route.
They
want
to
remain
a
pure
relational
database
that
honors
everything,
because
if
you,
for
example,
if
you,
if
you
connect
to
my
sequel
on
that
connection,
for
example,
you
can
create
session
variables
that
state
all
those
things
actually
make
every
connection
unique
and
expensive,
but
because
of
the
fact
that
they
are
committed
to
supporting
those
features,
they
cannot
move
into
something
that
what
we
test
us
in.
We
test
this
case.
What
we
say.
Those
are
actually
features
that
most
people
don't
use
in
Vitas
all
he
says.
B
So
yeah,
and
if
you
ask
me
they
should
be
part
of
my
sequel
and
the
only
reason
we
actually
tried
to
change
my
sequel
to
to
push
some
of
these
features
inside
there.
But
what
we
realized
was
that
as
soon
as
you
patch,
my
sequel,
nobody
wants
to
touch
it
because
it's
very
hard
to
build
it.
It's
people
have
probably
downloading
building.
So
we
use
we
started
with
a
5-line
patch
and
nobody
wanted
to
use
it.
So
we
quickly
moved
away
from
that
and
made
a
decision
that
we
will
work
only
with
stock.