►
From YouTube: SIG - Performance and scale 2023-07-06
Description
Meeting Notes:
https://docs.google.com/document/d/1d_b2o05FfBG37VwlC2Z1ZArnT9-_AEJoQTe7iKaQZ6I/edit#heading=h.tybh
A
Okay,
the
statistic
scale:
it's
July,
6
23.-
is
that
yourself
as
an
attendee
all
right,
let's
start
with
the
one
I
think
we're
almost
to
the
point.
We
can
wrap
this
up
right,
I
think
we're
take
a
look.
What
are
we
having
it's
open.
A
Oh
lee
wanted
to
add
some
other
metrics
away.
B
Yeah
so
I
think
I'm
going
to
comment
soon,
but
that
instance
type
tests
are
in
the
density
test
and
we
decided
not
to
upload
the
density
benchmarks
because
we
have
not
had
consistent
results
if
that
job
was
failing
for
a
long
time
and
the
graphs
are
not
really
meaningful
so
and
another
issue
is
that
in
our
scraping
we
only
scrape
the
first
two
jobs.
B
We
leave
the
instance
type
out,
so
we
need
to
add
another,
like
small
fix,
to
also
scrape
the
instance
type
so
and
I
don't
think
this
is
going
to
be
possible
for
V1,
but
definitely
possible
going
forward.
Okay,.
A
Okay,
I
think
I
asked
for
almost
to
take
a
look.
I
have
a
hard
turn,
so
we
need
recent
songs,
get
approval
on
so
I,
don't
know
if
we
need
that,
but
all
right
all
right.
So
let's
do
some
progress
there.
A
B
No
I
we've
discussed
this
last
time.
I
think
we
need
to
wait
for
that
thread
on
Google
Groups
keyword,
Dev
mailing
thread
where
they
are
discussing
about
the
changes.
Once
that
is
settled,
we
can
create
a
page
for
sixth
scale.
B
B
Ones
so,
apart
from
the
keyword
dock,
everything
else
will
be
moved
to
a
different
tracking
issue
right
because
essentially
we
are
saying
that
that's
not
required
for
V1.
A
Yeah,
so
it's
just
these
three
and
then
oh,
no,
that's
the
you
just
sent
me
this
right,
yeah
where's!
This
is
it
this.
B
No
I'll
put
it
here.
A
A
B
A
All
right
there
we
go
okay,
so
that
that'll
be
the
last
okay.
We
can
merge
that
one
today
and
then
I'm,
just
gonna
close
these
out
in
half
until
next
Tuesday,
okay,
yeah,
so
I
guess
as
soon
as
you
can.
Let's
get
those
comments
closed
out
and
then
I
can
pin
Roman
again
and
let's
see
if
we
can
get
this
merged
in
the
next
day
or
so.
A
A
A
All
right,
maybe
this
will
be
quick
and
easy.
Do
you
want
to
talk
about
the
oh?
Is
there
anything
else
about
V1?
Before
we
just
say
we
can
talk
about
some
flow
control.
B
No
I
think
that's
all
for
V1,
so
the
two
open
Major
items
are
that
PR
in
the
blog
post
I
think
we
have
plans
for
both
of
them,
so
that
covers
it.
A
Okay,
there,
it
goes
all
right,
let's,
let's
just
quickly
talk
about
flow
control,
so
did
so
I
think
this
is
the
test
that
you
did
right.
You
follow,
you
did
a
test
like
this
and
we
should
have
some
results
right.
B
A
B
The
restrictive
policy
was
that
it
allowed
eight
list
requests.
So,
however,
many
request
to
be
fire,
API
server
will
allow
eight
of
them
to
go
through
yeah
and
yeah.
So
with
this,
I
saw
that
API
server
platters
out.
So
initially
we
were
using
a
like.
Initially,
the
aps
server
was
using
around
nine
gigabytes
of
data
and
then
after
this
it
was
close
to
14.
So
there
is
a
spike
up
for
the
first
two
minutes
and
then
for
the
rest
of
the
time
the
API
server
memory
usage
is
plateauing,
so
it's
a
straight
line.
B
Although
the
test
fires
requests
for
30
minutes,
oh
actually,
I
was
staying
that
for
the
30
minute
time.
Sorry
I'll
keep
continuing
the.
Although
that
client
fires
the
request
for
30
minutes
the
API
server,
because
it
has
enqueued
a
lot
of
requests
and
because
client
is
configured
to
retry
upon
failure.
The
test
ends
up
stretching
for
a
long
period
of
time,
so
let's
say
an
hour
or
so
before
I
hit
like
session
timeout.
B
So
that
was
my
other
observation,
as
in
the
when
the
client
is
done,
the
test
is
not
done.
It
continues
to
process
those
requests
and
we
see
the
plattered
line
go
up
to
more
than
an
hour.
A
A
B
B
That
I
have
used,
they
don't
context
timeout.
So
let's
say
our
client
go
routine.
One
has
started
a
list
called,
and
that
list
called
let's
say
timed
out
or
returned
an
unknown
error.
Then
the
same
go
routine
same
client
internally
will
retry
and
continue
to
retry
until
it
passes.
So
eventually
that
list
call
will
give
us
like
an
end-to-end
timeout
of
let's
say
15
minute
right.
It
took
15
minutes
for
that
list
call
to
continue.
B
So,
if
you
have
similar
list
called
skewed
up,
you
can
imagine
that,
even
though
the
initial
burst
was
of
30
second
internally,
the
library
is
spending
more
time
and
that's
why
our
test
is
stretching
out,
because
the
test
is
stretching
out
the
load
on
the
API
server
is
stretching
out.
So
we
can
see
the
plateau
last
longer
than
just
30
minutes.
A
B
We
did
not
I
think
my
next
step
is
to
take
one
API
server
down
and
see
if
this
test
continues
to
pass
with
two
API
servers.
A
Like
run
off
by
that,
okay,
good,
that's
cool
all
right,
let's
see
what
we
found
from
that
yeah
that'll
be
good,
started
right,
I!
Think!
That's
all
we
had
from
last
time.
I
want
to
follow
up
on
I,
don't
know!
Is
there
anything
else,
so
we
want
to
go
through
I
think
it
was
a
full
control.
V1
yeah.
B
I
think
at
some
point
we
would
have
we
would.
It
would
be
good
to
do
like
a
post.
We
want
triage,
but
not
ready,
yet
we'll
have
to
create
a
new
tracking
issue
for
all
the
Skipper
fan
scale
related
items
that
open
that
are
open
and
find
a
new
like
post
V1
issue
for
it,
and
we
can
triage
that
in
the
next
one.