►
From YouTube: 2016-07-07 Kubernetes SIG Scaling - Weekly Meeting
Description
Public meeting recording of the Kubernetes Scalability SIG.
Check comments for meeting chat log.
B
And
we
will
be
posting
these
video
public,
so
don't
say
anything
that
you
wouldn't
want
to
show
up
on
the
front
page
of
the
newspaper.
That's
pretty
high
standard
all
right.
So,
let's
start
out
agenda
items.
Is
there
anything
that
folks
want
to
bring
up
and
talk
about
now,
I
know:
Bob.
There
was
some
talk
of
Samsung
showing
off
a
sharing
scale
data.
Are
you
guys
ready
to
demo
that.
C
Yes,
in
a
limited
form,
so
I
basically
what
I
wanted
to
show
is
two
ways
that
we
can
save
reduce
data
and
it's
not
a
complete
demo
at
this
point.
I'd
like
to
actually
give
this
at
a
meet
up
or
something
else
for
actually
have
slides,
but
I
wanted
to
kind
of
show
everybody
the
two
options,
we're
thinking
about
and
then
get
some
feedback
in
terms
of
what
other
people
are
doing
in
what
they
think.
B
Would
you
say
about
oh
I
said
I
said
the
floor.
I
think
the
floor
is
David's.
Isn't
it
yeah
I
just
want
to
make
sure
that
there's
other
stuff
that
people
really
want
to
get
to
today
that
we
can
we
can
get
to
it,
but
going
once
going
twice
all
right,
David,
please
take
it
away.
Okay,.
C
C
So,
given
that
permit,
this
doesn't
have
a
way
to
do
this.
Now
we
looked
at
two
possible
ways
to
do
it,
so
I'm
going
to
show
you
the
first
way
so
I'm
doing
this
locally.
This
is
not
going
to
be
on
a
kubri
Eddie's
machine,
but
the
commands
are
analogous
so
they're,
just
some
basic
Locker
commands
so
to
start
out,
I
have
a
default,
daugher
machine
and
so
odd.
I
can't
quite
see
how
do
I
hide
myself
there.
We
go.
Ok,
so.
C
C
And
I'm
just
scraping
myself,
so
nothing
fancy.
The
other
things
to
note
are
that
I've
set
a
the
storage
directory
to
be
from
the
root
partition,
Prometheus
and
then
I've
set
the
local
retention
time.
So
this
is
really
important
because
Prometheus,
because
they
do
not
intend
to
be
a
long-term
storage
facility,
they
periodically
will
Cole
data.
The
default
is
approximately
two
weeks
or
15
days.
C
So
clearly,
if
we,
if
we
back
up
Prometheus
data-
and
we
want
to
view
it
a
month
from
now
or
a
year
from
now-
which
is
not
unreasonable
as
soon
as
you
start
prometheus,
it
will
call
all
of
the
old
data.
So
we
purposely
start
Prometheus
with
a
long
retention
time
and
in
this
case,
I
chosen
a
hundred
years.
B
So
is
there
any?
Is
there
any
talk
in
the
Prometheus
community
about
sort
of
having
Prometheus
run
in
sort
of
a
view,
only
mode
versus
versus
sort
of
a
collect
and
because
I
mean
that's
kind
of
what
you
want
to
do.
Is
you
want
to
be
able
to
point
it
at
some
data
and
have
it
be
read-only
on
that
data?
C
C
C
So
we
want
to
do
is
we
want
to
go
ahead
and
collect
this
data
now
I'm
skipping
one
small
step
which
doesn't
technically
matter,
but
what
this
means
is.
So
what
am
I
doing
now
is
I'm
going
to
back
up
all
of
the
data,
and
so
what
I've
done
is
I've
executed
at
our
command
inside
of
the
docker
container
and
I've
just
torn
up
everything
in
the
database
directory
into
a
backup
directory?
Now
you
can
see
from
the
way
I
mounted
this.
The
backup
directory
is
actually
another
directory
on
my
host
machine
right.
C
C
The
purpose
of
this
local
memory
chunks
parameter
here,
okay
d,
by
setting
it
lower,
you
can
cause
it
to
flush
your
data
to
disk
more
quickly,
but
there's
currently
no
synchronous
way
to
ensure
that
all
of
the
data
is
on
disk.
So
in
the
real
world.
C
What
you
would
normally
do
is,
you
know,
pause
for
five
seconds,
just
like
in
the
old
days,
ext3
just
to
wait
for
the
five
seconds
sink
right,
and
this
is
one
of
the
limitations
of
this
method
of
collecting
and
we'll
see
later,
a
different
method
of
collecting
which
doesn't
have
that
limitation,
but
then
has
other
limitations.
So
so
we
backed
up
this
data
and
let's.
C
B
C
However,
there
aren't
many
tools
that
know
how
to
natively
manipulate
that
database
and
in
particular
you
wouldn't
want
to,
because
if
you
had
Prometheus
modifying
that
the
database,
which
would
be
the
normal
case,
you
would
end
up
with
correct
yeah
few
people
potentially
reading
and
writing.
I
just
wanted
to
point
out
that
there
is
a
really
good
article
here
that
talks
about
the
way
the
disk.
C
C
C
C
B
E
D
C
C
Now
we
have
a
ridiculous
instance
again
again:
I
apologize,
I've
been
modifying
this
since
five
this
morning,
so
just
a
few
hookups,
so
we
have
a
ruthless
server
running
again.
The
darker
machine,
IP
is
not
change,
so
let's
go
ahead
and
see
if
we
can
access
it.
I
reloaded
the
metrics
page
and
sure
enough.
There
is
data
there.
So
let's
go
ahead
and
in
let's,
let's
do
a
query,
so
I
did
a
query
now.
C
So,
let's
see,
how
can
we
find
out.
C
Well
so
so
let
me
finish
off
with
the
with
what
I
demonstrate
so
far
I've
backed
up
prometheus
I
restarted
it
using
a
database
from
yesterday
and
I
now
have
the
difficulty
that
I
can't
actually
easily
query
it
because
I
don't
remember
what
the
date
was
in
a
unix
timestamp
yesterday.
Let
me
see
if
I
have
anything
in
my
notes.
Maybe
this.
C
C
Method
is
everything
is
in
prometheus.
It's
compact,
which
I
do
not
think,
is
a
particular
advantage,
because
this
is
the
reason.
That's
not
an
advantage.
Is
that
when
you
actually
do
analysis
with
this
theta
you're
going
to
be
dealing
with
large
data
sets
and
in
all
likelihood,
you're
going
to
end
up
using
MapReduce
jobs
or
some
other
tool,
I
mean
you
could
depend
on
or
fauna.
If
all
you
want
to
show
is
a
dashboard,
but
I.
C
Think
part
of
the
value
of
sharing
is
to
be
able
to
do
sort
of
the
comparative
analysis
that
we
saw
with
the
core
west
scalability
paper,
one
of
the
things
you'll
notice
about
the
scale
of
paper
they
did
there
is
that
none
of
the
graphs
were
dashboards.
Their
graphs
were
all
generated
using
something
like
matplotlib
Boca
or
even
it
could
have
been
our.
But
it's
it's
used
the
used
graphing
libraries
that
are
amenable
for
Clinton
publication,
so
the
second.
So
the
current
method
that
I've
shown
shown
is
that
you
can
stand
up.
C
A
committee
of
server
pointed
at
an
old
database.
It
will
start,
and
you
will
have
that
data,
and
then
you
can
point
your
opponent
front
end
to
it
and
you
can
have
all
of
your
pretty
dashboards
in
you
know.
Do
whatever
you
want
with
it.
However,
not
everyone
may
run
from
it
to
us,
so
there's
a
second
way
that
we
could
collect
data
which
doesn't
have
some
of
the
limitations
of
the
previous
one.
So
the
previous
one,
we
see
the
raw
database
itself,
which
is
again
it's
permitted
specific.
C
C
Going
to
see
what
I've
done
wrong
here,
my
too
far
into
the
future.
C
Well,
I'm
not
sure
at
any
rate,
what
what
you
saw
earlier
and
I'm
not
able
to
show
right
now
is
you
can
query
the
Prometheus
database
while
it's
running
so
instead
of
actually
capturing
the
database,
you
can
query
it
and
get
a
JSON
result.
C
So
an
alternative
way
of
sharing
data
would
be
to
just
save
the
JSON,
and
what
you
would
do
is
for
each
of
the
metrics
on
this
metrics
page
or
any
of
the
ones
that
you
wanted
to
share.
You
would
do
a
query
during
which
covers
the
duration
of
the
test,
and
you
would
just
curl
back
query
into
a
file
and
then
share
those
flat
text
files
with
others.
C
Now,
one
of
the
advantages
of
doing
this
is
that
you
then
have
the
data
in
a
format
that
you
can
import
into
any
of
the
tools
I
mentioned
earlier.
Things
like
our
boca,
matplotlib,
etc,
and
so
you
can
do
more
advanced
analysis
that
you
can't
do
using
simply
grow
fauna.
The
disadvantage
of
this
is
you
lose
for
fondest
querying
capabilities,
because
you
only
have
the
ability
to
use
the
query
language
at
the
time
that
you
capture
the
results
for
sharing.
C
C
Itself
and
then
sharing
that
which
then
has
the
complexity
of
people
having
to
start
Prometheus
to
see
the
data,
but
you
get
the
advantage
of
having
a
query
language,
vs
sporting,
just
simple
JSON
for
sharing,
which
gives
it
the
advantage
of
more
tools
can
understand
JSON.
But
it's
a
lot
well
the
size
again,
I
already
doesn't
matter.
You
lose
the
ability
to
use
the
primitives
query
language,
which
I
think
is
a
pretty
significant
loss.
B
Then
it
would
have
any
opinions,
I
mean,
so
the
ideal
here
would
be
something
that
mixes
these,
where
we
have
something
that
is
usable
by
both
promesas
Prometheus
itself,
so
that
we
can
sort
of
retain
the
Korean
language.
But
then
it's
also
so
you
know
well
documented
with
good
tool
set
so
that
you
can
be
used
in
other
contexts.
Would
you
say
that's
correct,
I
agree
I'm,
just
I
mean
like
that
would
be.
That
would
be
the
sort
of
you
know
if
we
could
wish
for
a
pony
type
of
thing.
Well,.
C
So
I
think
the
amount
of
Jason
we're
talking
about
saving
is
on
the
order
of
you
know.
Tens
of
megabytes,
which
is
even
if
we
were
talking
about
hundreds
of
megabytes,
I,
wouldn't
I,
wouldn't
think
that's
too
much
storage
as
well
as
will
be
cheap
again.
One
of
the
reasons
that
I
think
that
actually
storing
the
JSON
is
better.
Is
that
I
think
that
at
some
point,
we'd
like
to
be
able
to
do
more
than
simply
create
dashboards,
I?
Think
nor
der
to
do
more
fine-grained
analysis.
C
E
So
I
may
have
missed
this
during
your
walk
through
David,
but
the
second
method
allow
for
easier
normalization
of
timestamps.
It
certainly
felt
like
you
were
having
to
fumble
between
today
and
yesterday
and
tomorrow
for
timestamp
queries
and
back
in
the
I'm
on
a
pony
world,
I'd
love
to
see
a
time
series
that
was
like
normalized
to
you
know.
Zero
is
the
first
time
stamp
on
words
that
way,
conceivably,
we
could
at
least
visually
compare
different
test
runs
pretty
easily.
They
started
from
the
same
checkpoint
or
something
right.
C
So
that's
a
good
point:
I'm,
not
aware
of
a
way
I,
so
I
know
the
Prometheus
does
store
the
timestamp
and
the
time
stamp
it
stores
is
is
now
so
it
doesn't
ever
start
at
zero.
You
can
have
it
seek
to
a
certain
time
and
consider
that
zero
for
the
purposes
of
graphing,
but
the
data
itself
as
far
as
I'm
aware,
is
always
stored
as
the
current
timestamp
as
the
node
sees
it.
C
That's
something
that
if
you
were
using
like
a
Python
program
or
something
to
you,
read
the
JSON
you
could
you
could
account
for
that
and
modify
it
fairly
easily
and
I
think
Ravana.
You
could
also
seek
to
a
particular
time,
but
you're
right,
you
would
need
to
know
the
time.
I
think
that
part
of
the
difficulty
with
this
presentation
is
that,
because
I'm
not
showing
ravana,
I
did
have
to
have
a
little
bit
more
knowledge
of
the
time.
Stamping.
C
So
that's
something
that
if
you
look
at
the-
and
I
know
this
doesn't
fix
it,
so
your
points
well
taken,
I
think,
is
something
to
consider.
But
if
you
look
at
my
original
backups,
I
wasn't
including
the
timestamp
and
I
only
started,
including
the
timestamp
this
morning,
because
I
realized
I
couldn't
do
the
math
in
my
head
fast
enough.
Okay,.
E
E
It's
but
there
was
certainly
a
lot
of
chatter
in
the
prometheus
back-channel.
They
continue
to
maintain
a
hard-line
that
prometheus
is
not
meant
for
long-term
storage,
it's
not
meant
for
archival,
and
it's
not
meant
for
backfilling.
So
their
proposal
for
things
of
this
nature
is
you
need
to
have
prometheus
dumped
off
to
external
databases
as
quickly
as
possible,
something
like
in
flux,
DB
or
something
like
open,
TS
DB,
and
it
could
be
that
that's
sort
of
the
longer
term
version
of
option
number
two
I
think.
E
Maybe
the
hope
here
it
was
that
option
number
one
if
he
could
best
fake
prometheus
out
with
the
right
amount
of
theta
was
going
to
be
lowest
cost
best
effort
get
us
eighty
percent
of
the
way
there
but
seems
like
we
need
a
little
more
polish
at
the
moment.
C
So
I
specifically
actually
didn't
so
you're
right.
That's
as
I
mentioned.
Maybe
not
prometheus
is
not
going
to
be
a
long
term
storage
platform
ever.
It
was
definitely
just
to
be
clear
with
soup.
E
E
Just
over
and
over
and
over
again
that
was
you
know
many
months
ago,
so
things
may
have
changed,
but
certainly
for
the
collection
aspect,
I
think
prometheus
dumping
into
its
own
files
as
quickly
as
possible
seems
like
that
still
scale
as
well.
But
the
export
story
seems
a
little
painful,
so.
C
I
purposes,
so
I
purposely
didn't
bring
this
stuff
up,
because
I
think
it's
a
more
contentious
and
potentially
long-running
discussion
that
we
have
there
is
talk
of
using
influx.
Db
is
a
long-term
storage
solution
for
prometheus
there's,
also
talk
of
potentially
using
some
other
database.
My
vote
personally
leans
towards
influx
Phoebe,
but
I
think
that
there
are
potential
problems
with
the
Rope
out
of
the
fact
that
the
clustered
nature
flux
DB
is
become
close
source
as
of
a
few
months
ago,
etcetera,
etc.
So
I
completely
agree
with
you
bit
having
a
difference.
C
A
long
term.
Storage
format
is
something
that
we
need
to
follow
up
on,
but
the
purpose
is
here
is
or
is
really
justifying
the
lowest
common
denominator
that
allows
us
to
move
forward
now
without
having
to
resolve
the
issues
of
whether
or
not
we
want
to
use
influx,
DB
or
something
else.
I
intend
to
continue
looking
at
an
influx
TV,
but
I
think
that
a
JSON
representation
is
is
clearly
the
lowest
common
denominator.
C
E
Capital
alkylating
tier
earlier
so
back
in
September,
says
Samsung
SDS
was
doing
a
Prometheus
collection
on
a
thousand
hits
and
you
found
exactly
what
you
did
that
in
flux.
Db
just
doesn't
scale
for
right.
At
that
volume,
influx
TV
point
eight
or
in
flux,
DB
point:
nine!
Take
your
pick
how
they
handle
the
wide
variety
of
columns
that
keeps
your
Prometheus
threw
at.
It
just
fell
apart.
So
the
reason
I
think
we're
talking
about
influx
DD
here,
isn't
to
write
data
into
it
as
quickly
as
possible,
but
use
Prometheus
for
that.
C
That
we
would
only
my
interest
would
largely
be
to
use
on
flex.
Db
is
a
format
for
sharing.
I
think
that
the
scalability
problems
that
people
see
within
blocks
DB
many
of
them
are
due
the
fact
that
influx
TB
has
a
format
which
takes
in
order
of
magnitude
more
space.
Timotheus,
it's
not
nearly
as
compact,
and
the
reason
is,
is
because
they
did
you
analyze
everything
so
that
every
data
point
has
all
of
its
metadata
with
it
some
steep.
This
is
a
disadvantage
because
they
think
that
they
want
to
save
space.
C
In
my
experience
and
I
had
this
at
amazon,
we
spent
a
lot
of
time
trying
to
compress
all
of
our
metrics
and
what
we
found
is
it
just.
It
was
effectively
like
our
metrics,
because
we
couldn't
get
any
of
our
tools
to
work
with
the
compressed
versions,
so
we
ended
up
deciding
that
we
could
take
the
three
or
four
orders
of
magnitude
storage
increase.
In
order
to
be
able
to
have
easy
access
to
the
data,
the
source
was
cheap.
C
The
programmer
time
was
not
so
I
think
that
prometheus
historically,
there
has
been
complaints
about
its
scalability
I.
Think
many
of
them
are
probably
due
to
the
fact
that
it
requires
more
hardware
to
do
the
same
thing
with
you.
It
does
because
it's
doing
something
it's
making
the
trade
off.
So
that's
something
we
could
revisit
later,
but
take
your
pointer
and
I
I
think
there
is
when
you're
talking
about
archiving
I,
don't
think
if
you
have
a
yeast
feel
a
little
problems
because
you're
not
again,
it's
not
real
time
ingestion.
A
I
think
related
to
the
influx
DB
comment
and
our
earlier
experiments.
It
started
to
become
obvious
that
the
influx
DB
cluster
required
to
to
that
you
would
need
to
scale
up
in
order
to
measure
a
cluster
was
getting
to
be
a
significant
fraction
of
the
cluster
size.
And
could
you
scale
it?
Maybe
would
you
ever
want
to
in
a
production
environment,
spend
that
much
on
just
an
influx
DB
cluster
for
for
this
case
almost
certainly.
B
B
All
the
options
have
some
real
downsides,
whether
it
be
in
flux,
DB
with
you
know,
you
guys
hit
the
scale
with
mobility
issues,
but
then
there
is
also
the
clothes
clustering
issue
that
it
looks
like
inside
Prometheus,
as
viewed
as
is
making
flux,
pretty
much
be
a
non-starter
as
the
default
option.
They're
looking
at
doing
sort
of
generic
writing
back
ends
that
you
can
sort
of
you
know
put
anything
in
there.
There's
a
lot
of
talk
about
long-term
storage
backends
being
right
only
so
you
can't
actually
read
in
query.
B
So
it's
up
to
you
to
use
the
query
language
of
the
database
you're
writing
into
to
be
able
to
do
any
aggregations.
So
that's
a
little
unfortunate
there.
Also.
So,
even
if
we
did
have
a
long-term
storage
system
like
influx,
TB
or
or
open
to
ftp,
we
still
wouldn't
be
able
to
use
the
prometheus
front
end
to
query
that
stuff.
At
least
it
looks
like
that's
the
case.
A
Joe,
what's
your
what's
your
gut
reaction
on
if
we
started
publishing
JSON
from
you,
think
this
with
that
has
any
value
or
here's.
B
What
I
think
might
be
interesting
if
we
have
a
way
to
export
data
out
of
Prometheus
right
so
because,
because
we're
in
a
situation
where
the
Prometheus
storage
back
end
is
good
enough
for
scalability
runs
okay,
let's
ignore
sort
of
long-term
storage
for
four
clusters
in
general,
so
it's
good
enough
for
scalability
runs
we're
not
going
to
be
collecting
so
much
data
that
we
have
to
get
it
off
post
directly,
and
so
now
it's
a
matter
of
okay.
We
have
this
this
data
sitting
around.
B
If
you
want
to
for
sort
of
live
querying,
and
so
this
you
know-
and
it
looks
like
there's,
there's
talk
of
having
an
import,
API
and-
and
so
this
doesn't
seem
like
in
and
all
the
exporting
there's
an
exporting
PR
for
doing,
live
export
to
an
external
long-term
data
storage,
and
that's
hung
up
on
questions
of
retries
and
reliability
in
keeping
up
and
synchronous
versus
a
synchronous
versus
buffering
that
type
of
thing
I'm.
B
If
we
say
you
know
what
this
stuff
is
all
offline
and
it
you
know-
and
it's
all
you
know,
and
none
of
it's
happening
real-time,
it
may
be
easier
to
actually
move
Prometheus
in
some
of
these
directions.
If
we
want
to
try
and
move
Prometheus
I
mean
you
know,
it's
not
like.
We
have.
You
know
engineers
falling
out
of
trees
here.
So
you
know,
in
that
case
I
mean
like
exporting.
The
JSON
gets
us
there.
Another
interesting
thing
to
think
about
is
that
you
know
prometheus
is
written
in
go.
B
Is
it
possible
to
extract
the
prometheus
sort
of
query
language,
query
processor
and
run
that
in
a
batch
mode
across
json
data
right?
So
now
you
know,
I
could
imagine
a
command-line
tool
where
you
point
it
at
a
some
sort
of
you
know,
store
on
disk,
and
this
thing
just
runs
and
turns
through
that
loads
it
from
disk
as
necessary
and
uses
the
same
query:
languages
Prometheus.
So
it's
essentially
offline.
You
know,
post-hoc
analysis,
I,
think
you.
C
Don't
even
need
to
go
that
far,
because
so
pretty
those
has
something
called
exporter's
and
they're
normally
used
for
things
like
applications
which
aren't
instrumented
worthless
things
like
tail
tailing
log
files
right
that
they
didn't.
So
you
could
have
an
exporter
and
they
may
already
have
one
which
would
just
read
the
JSON
and
ingest
it
into
prometheus.
Okay,.
B
Does
that
time
work
with
like
more
historical
data?
Will
it
in
insert
it
into
or
is
it
just
you
know,
because
a
lot
of
these
things
they're
only
built
to
ingest
stuff,
that's
happening.
You
know
now
now
some
some
definition
of
now
right,
and
so,
if
you
use
one
of
those
exporter's
to
give
a
data
that
came
from
last
week,
will
it
will
it
it
freaked
out
I,
don't.
B
For
me,
yeah,
so
I
don't
know
so
it
sounds
like
we're
trying
to
use
Prometheus
in
sort
of
a
weird
way,
I
think
the
JSON
stuff.
You
know
if
we
can
get
this
stuff
exported
as
JSON,
then
you
know
that's.
At
least
you
know
something
that
that
is
well
understood
and
can
be
can
be
widely
used.
But
you
know,
if
we're
not
careful,
we're
going
to
end
up
rewriting
all
the
alga.
The
sort
of
you
know
aggregation
calculations
on
top
of
that
JSON
output,
which
seems
silly
right.
B
B
B
One
thing
that
I
didn't
want
to
mention
and
we
are
totally
out
of
time
I
if
prometheus
works,
like
I,
think
it
might
work,
it
might
be
possible
to
actually
do
a
little
bit
of
a
binary
search
across
the
time
scale
to
actually
take
a
data
set,
take
Prometheus
and
say
at
what
times
do
we
have
interesting
data
for
this
metrics,
and
so
the
question
there
is
is
essentially
there's
your
hitting
that
11,000
point
when
you
do
that
step.
Does
that?
Does
that
essentially
down
sample
it?
B
So
if
you
say
step
equals
2,
does
that
say:
I
want
to
actually
return
every
second
step.
Could
you
go
to
step
e
equals
100
and
do
a
wider
thing
find
out
where
you
might
have
hit
a
data
point
and
then
zoom
in
into
that?
That
might
be
a
way
to
go.
Also
in
terms
of
like
building,
because
it's
like
again
it's
like
you
know,
Prometheus
doesn't
have
the
way
to
say
you
know.
B
When
is
there
interesting
data
for
this
time
series
you
have
to
essentially
know
the
time
you're
looking
for
so
that
you
know,
building
some
to
ability
around
that
might
make
this
whole
thing
flow,
a
heck
of
a
lot
better
yeah,
but
very
interesting
stuff,
thanks
for
thanks
for
demoing
cool
and
with
that
I
think
we
are
out
of
time.
Thank
you,
everybody,
and
since
we
have
this
recorded,
will
try
and
post
it
and
see
how
all
that
works.
It's
brave
new
world
for
us
and
see
you
later
today,
thanks.