►
From YouTube: 2020-05-27 :: Ceph Science User Group Meeting
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
Hey
well
welcome
to
the
meeting,
I
hope
everybody's
doing
well.
If
you
haven't
been
here
before
it's
just
an
informal
talk
about
stuff
and
science
research,
big
clusters,
you
know
get
much
anything
related,
very
informal
I
try
to
keep
things
lively,
but
I
kind
of
rely
on
anybody
else
to
contribute.
If
you're,
not
all
going
on
the
pad,
you
can
go,
take
a
look
there
and
add
any
topics.
A
A
B
B
B
C
B
But
in
terms
of
outages,
not
not
too
much,
there
seems
like,
usually
if,
if
it's
not
like
an
enormous
outage,
and
you
have
corruption
or
any
sort
of
problems
like
that,
they
just
recommend
that
you
recreate
the
OSD
and
they
don't
seem
as
interested
on
the
list
anymore,
at
trying
to
really
dig
really
deeply
into
helping
you
recover.
It.
A
B
A
D
This
was
more
than
a
month
ago,
and
we
have
we
have
like
nine
or
ten
actives
I.
Think
nine
active
MDS
is
in
a
set
of
s
cluster
and
we
had
one
MDS
that
that
just
crashed
because
the
Machine
crashed
it
just
rebooted,
so
the
sand
by
it
came
in,
but
the
the
standby
was
going
out
of
memory.
And
so
then
we
we
learned
that
there's
like.
Basically,
when
you
have
multiple
multi
active
MTS's
clients
can
clients
can
like
respond.
D
So
what
we
had
to
do
was
was
decrease
the
cache
size
of
all
the
MTS's,
so
that,
like
this
flushed
all
the
the
client
caches
and
in
terms
of
what
the
clients
thought
that
they
had
cached
what
the
clients
about
the
NDIS
had
cache,
and
this
allowed
this
standby
MDS
to
to
booth.
So
this
was
a
bit
turkey
at
a
bit
scary
and
there's
some
in
I.
Think
in
octopus.
D
Were
down
for
around
an
hour,
and
we
have
no
idea
what
to
do.
We
had
to
convey
to
reach
out
to
a
developer
to
who
suggested
decrease
in
the
cache
sizes.
So
now
we
don't,
we
used
to
run
with
large
caches
on
all
of
these
MTS's
like
these
are
a
32
gigabyte,
GM's
and
we're
on
with
16
gigabyte
MDS
memory
size.
But
now
we
run
with
4
gigabytes
of
4
gigabytes,
like
per
MDS.
D
B
D
F
But
being
a
bit
of
a
problem
with
luminous
and
multi-part
objects
being
uploaded
while
alive
and
Emily,
we
keep
em
young
with
buckets
with
objects,
and
we
can't
delete
I,
think
we're
going
to
have
to
either
find
an
updated
version
or
disable
dynamic,
recharging.
I
think
I
hid
rumor
and
fixed
in
the
new
one.
Isn't
it
don't
think
those
made
it.
D
F
A
F
F
H
I
G
J
F
G
If
I
can
afford
before
the
interesting
problem
with
the
the
balancer,
which
seems
to
be
fixed
in
Nautilus,
had
the
balancer
running
and
I
was
yeah,
it's
it
depends
at
some
point,
seems
to
get
into
an
infinite
loop
and
takes
up
a
hundred
status,
you
of
a
CPU,
and
we
did
a
bit
of
yeah.
We
did
a
bit
of
tracing.
Just
transfer
figure
see
if
you
can
figure
out
what
is
going
on.
Venice
Italy,
it's
getting
stuck
in
EST
map
and
I
know:
they've
been
a
couple
reports.
G
Being
people
reporting
this
that
yeah
that
suddenly
the
de
belem
sir
just
I
was
telling
sorry
the
if
this
if
manager
starts
yeah,
basically
just
not
servicing
Prometheus
and
that
kind
of
stuff,
and
that
that's
the
behavior
that
we
saw
and
but
when
kind
of
digging
down
into
it
areas
it
was
the
balance
so
because
you've
it.
So
it's
and
not
the
balance,
to
put
it
back
in
and
everything.
G
That's
why
working
again
as
soon
as
I
tried
to
run
run-
and
this
is
actually
just
yeah
just
to
doing
an
op
doing
and
trying
to
do
and
doing
optimization.
We
then
started
trying
it
with
OCE
OSD
map
tool
and
every
cement
floor
had
the
same
behavior,
but
then
yeah.
We
tried
it
in
Nautilus
and
it
was
working,
fine
and
Nautilus.
So
there
was
definitely
something
that's
been.
A
K
A
D
D
Because
what
I
was
gonna
say
is
that
if
you
use
like
if
the
OSD
is
that
with
blue
store
compression
mode
aggressive,
then
I
think
everything
that
the
disc
right
to
then
everything
that
the
disc
right
into
blue
sort
of
compressed,
including
all
the
maps
and
all
of
this
kind
of
stuff,
and
that's
what
was
leading
to
that's.
What
led
to
some
big
crash
like
that
I
think
I
mentioned
last
time.
D
We
we
had
this
call
that,
like
it,
exposed
a
bug
in
ELLs
and
poor,
but
if
you
just,
if
you
leave
the,
if
you
leave
the
OSD
compression
mode
set
to
passive,
then
it
will
only
compress
the
it
will
only
compress
the
pool
data
so
then
like
when
you
compress
just
the
object.
It's
just
like
compressing
large
chunks
of
memory
thoughts,
my
ID,
it
might
be
better
if
you
come.
If
you
compress
the
okay,
you
put
in
something.
K
D
K
L
D
J
The
end
we
had
to
turn
off
pre
extending
of
the
world
and
then
cycle
through
each
of
the
OSC's
taking
it
down,
bring
it
back
up,
seeing
if
it
would
come
back
up
or
whether
it
failed
due
to
the
detecting
corruption,
and
if
that
happened
and
I
had
to
destroy
the
OSD
rebuild,
it
effect,
fill
make
sure
it
came
up.
Okay,
then,
then,
move
to
the
next
OSD
and
go
through
the
whole
lot.
J
Real
painful,
but
at
least
me
in
that
way,
by
choosing
that
route,
we
chose
the
slightly
riskier
option
of
the
saving.
Extending,
whilst
that's
running
I,
understand
that
there's
a
small
risk
of
something
going
there
there,
but
I
took
our
girls
advice
and
we
tried
it
and
just
city
probably
and
nothing
bad
happens,
but
I
don't
know
what
could
have
happened,
but
essentially
it
does
leave
you
in
the
horrible
state.
If
you
get
a
power
fails,
something
like
that,
then
you'll
lose
basically.
You've
lost
large
amount
of
them.
If
you've
got
that
corruption
that
man.
J
However,
so
running
at
the
moment
with
having
done
that
I'm
running
at
the
moment,
with
the
extending
turned
off
off
and
periodically,
we
started
it's
the
odd
OSD
just
out
of
curiosity
to
make
sure
that
raspberry
start
okay
and
haven't
had
any
more
problems
but
I.
Let
somebody
else
be
the
guinea
pig
for
the
year's
next.
K
K
Hardware
related
issues
and
OSD
can't
save
his
current
super
block
values,
so
it
keeps
updated
his
value
store,
but
the
device
isn't
updated
after
result
of
the
process.
As
the
book
looks
crushed
and
there
is
quite
simple
repair,
my
thought
and
and
I
pasted
an
instruction
for
that
body
F.
We
will
have
part
of
this
issue.
I
noticed
on
the
ERC
one.
K
K
B
K
We
had
copies,
but
each
happens
on
Paris
then
hosts
when
we
upgraded
when
we
are
during
upgrade.
We
worried
about
that.
Another
100
hosts
will
have
the
same
issue
and
recreating
a
lot
of
voice.
This
will
take
so
much
time,
so
we
decided
to
research
why
it
happens
and
how
to
repair
that
in
a
fast
way.
I
D
K
Exactly
because
st,
when
trying
to
start
the
first
thing
is
checking
the
checksum
of
OS,
the
super
block,
and
that
was
mismatch
between
burials
the
block,
checksum
and
stored
one
in
key
value
store.
But
there
is
a
guitar
checksum
option
with
an
Albert
it
by
OSD,
tries
to
read
very,
very
old
team
up.
A
couple
store
again
because
it
doesn't
exist
was
simplest
way,
was
to
inject
Earth
OSD
maps
and
when
it
goes
up.
G
If
I
can
ask
them
and
did
you
because
every
time
I've
done
a
gentle
rewrite,
it
means
a
whole
bunch
of
data.
You
know:
there's
a
crash
map
recalculation,
so
it's
open
today
to
moving
around
the
cluster
which
doesn't
need
to
did
you
observe
same
room?
Same
result,
I,
say
similar
kinds
of
problems
or
yeah.
A
G
So
it
is,
you
know
it's
and
you
kind
of
need
to
yeah.
So
so
we've
we've
had
gentle
rewear
to
you.
Also,
we
just
try
to
I
mean
yeah.
We've
also
try
to
just
drop
hard
way
in
and
and
see
what
safe
does,
and
what
you
noticed
was
that
because
IP
Jesus
OPEC,
when
it
does
the
recalc
and
it
doesn't
obviously
it's
not
it's
not
clicking
at
the
size
of
the
pts,
and
we
immediately
decided
over
filling
out
these
on
the
on
the
new
hardware.
So
that's
we've
definitely
thought
that
was
nobody
to
it.
G
It
seems
like
that
when
we
base
what
to
do
is
he
I
mean
yeah?
This
is
something
at
Dan
talked
about
a
while
back
where
you,
you
said
no
rebalance
on
the
cluster,
and
then
you,
you
add,
you
change
the
crash
map
and
then
you
use
up
map
to
add
map
or
the
PG
SPECT
where
there
were,
and
this
this
all
seems
to
work.
The
classic
goes
back
to
health,
okay,
but
then,
whenever
I
run
the
rebalance
again,
it
tries
to
undo
what
I've
done.
G
That's
basically
just
moving
all
that
data
again
and
I
just
thought.
This
is
a
we
about
to
yeah
I.
Want
you
to
start
trying
to
use
this
in
in
production,
so
I've
written
a
whole
bunch
of
scripts.
That
does
this
automatically
for
me
and
it's
pretty
cool
I
can
I
can
go
and
I
can
I
said
very
balanced,
I
have
new
hardware
and
then
I
within
three
or
four
minutes.
G
And
you
can
you
can
scroll
through
it?
You
can
see,
there's,
there's
PG
temp
items
and
this
PGE
up
map
items
and
so
a
whole
bunch
of
stuff,
and
it's
it's
it's
basically
when
it
does,
when
the
balancer
does
its
work.
It
goes
and
has
a
look
at
that,
and
it
basically
has
to
follow
the
rules
that
are
set
there
and
it
seems
almost
impossible
to
undo
those
bits
of
information.
So
it's
like
almost
like.
G
Even
you
know
using
the
the
app
map
you
still
need
to
do
you
do
those
bits
of
data
that
you
didn't
want
to
move
but
as
say,
I
thought.
This
is
a
perfect
opportunity
just
find
out
if
anybody
else's
used
at
map
and
we've
had
some
experiences
and
also
just
to
check
if
other
people
doing
it
in
any
different
ways.
G
F
G
A
G
G
At
at
some
point,
I
will
I
will
then
I
just
didn't
want
to
yeah
once
you
kind
of
understand,
what's
what's
going
on
first
before
I,
I
am
sure
them
because
yeah
I've,
they
have
had
a
look
at
various
options.
I
mean
this
because
there's
PG
FF
allows
you
peih-gee
erik
and
peih-gee
items
in
the
beginning.
I
was
using
PG
up
items,
but
what
seems
to
happen
is
that
when
you
use
PG
FF
items
on
an
RPG
with
their
primaries
changed,
it's
it
it.
It
doesn't,
allow
you
to.
G
It,
doesn't
like
to
up
map
it
kind
of
changes
up
one
of
those
these
the
whole
time.
So
then
this
is
a
chronicle
PDF
map
which
allows
you
to
just
say.
This
is
what
the
what
the
f
map
state
must
look
like
and
that
works
that
works
great,
but
I
said
then.
This
kind
of
this
is
this
is
hysteresis
in
the
system.
You
know
that's
every
time
you
run
the
balancer,
it
goes
and
looks
at
what
you've
done
and
says.
G
D
G
D
It
just
has
it
has
a
few
like
heuristics
to
make
sure
that
you
haven't
done.
You
haven't
done
some
up
map
which
is
like
illegal,
which
is
breaking
the
failure
domain,
but
this
thing
probably
removed
like
especially
for
a
Checotah
pool
that
probably
removed
things
that
it
would
be
even
valid
up
Maps,
occasionally,
if
you
were,
if
you
were
generating
those
yourself,
but
that's
probably
yeah.
D
Just
that,
maybe
maybe
with
the
RPG
up
map
instead
of
up
my
two
items
it
it
might
call
it
might
not
use
that
same
clean
cleaning,
your
attic,
but
I
might
just
have
it
stick
or
made
like.
Maybe
you
are,
maybe
you
are
up
mapping
in
a
case.
That's
like
breaking
the
failure
domain.
Is
that
possible,
like
then
mapping
to
olefin
that
they
see
one
PT
to
the
same
rack
twice
or
something.
G
Well,
just
to
describe
what
I'm
doing
is
yeah
I
mean
I'm
I'm,
yeah
I'm,
sitting
very
balanced
on
the
cluster
and
then
and
then
changing
the
crash
crash
map.
It's
so
then
it's
it's
all
a
bunch
of
he
G's
into
Remax
backfilling
and
back
full
weight.
I
then
go
and
run
analysis
on
that
and
just
look
at
what
the
acting
setters
and
then
just
call
PG
at
map
safety
PDF
map
with
the
with
the
acting
set
so
that
it
forces
the
upset
to
be
there
acting
set.
G
So
it
does
this
whole,
like
rewind
and
I,
said
it
works
like
a
charm.
This
is
great
but
yeah
I
said
then,
when
there's
a
drag
me
balance
to
the
balance
is
going
and
and
and
so
somehow
in
the
in
the
OST
map,
it's
it's
it's
keeping
track
of
what
I'm
doing.
If
and
if
you
look
at
the
at
the
HD
map,
it
lists
some
in
PG
e
and
PDF
map.
I
can
see,
it's
is
listing,
hope
and
is
listing
the
iced
teas
that
I've
that
I've
I've,
that
I've
said
and
then
I
said.
L
G
G
D
L
D
Lady,
so
the
balancer,
we
need
to
turn
it
off
when
you
doing
those
kind
of
not
bring
out
up
map
remapping,
because
it
will
obviously
it
will
undo
what
you
do,
but
that's
sort
of
exactly
what
you
wanted
to
do
afterwards
right,
you
want
to.
You
want
to
up
map
to
where
the
data
is
now
and
then
you
want
to
gradually.
You
would
move
those
up
maps
to
put
the
data
onto
the
new
day.
G
It
is
yeah
exactly,
but
then
so
do
that
as
soon
as
around
the
balancer,
the
balancer
is
not
actually
moving
any
data
on
to
the
new
hardware.
The
balance
is
just
moving
moving
moving
P
G's
around
anyway,
but
with
the
new
hardware,
and
so
I
say
that
and
I
said
we
have.
This
is
on
our
test.
Cluster
and
I
really
butyl
eyes,
this
Proteus
cluster.
G
G
That's
why
I
thought
I'd
just
check
if
anybody
else
has
used
up
map
and
had
some
experiences
or
because
I'm,
because
it's
yeah,
because
right
nice
that
I
can
now
you
know
when
I
am
I,
think
go
and
do
an
analysis
and
I
can
see
where
it's
trying
to
remove
data
to
I
bet.
He
lists
the
rack
and
the
host
that
it's
and
that's
trying
to
move
data
to
just
I
can
get
energy
about
what
and
then
the
side
of
the
PG
as
well.
G
Just
gonna
do
a
about
what
data
is
moving
around
and
I
said.
Then
that's
what
that's
when
I
started,
seeing
kind
of
like
things
are
not
that's
not
doing
what
I
thought
that
there
and
I
said
then
are
in
the
balance
again.
I
look
at
the
analysis,
and
it
shows
me
that
yeah,
it's
I'll,
try
to
move
any
data
onto
the
new
hardware
at
all.
Yet
but
I
say
I
mean
if
other
people
I've
been
using
this
and
it's
it's.
G
B
G
A
F
Putting
on
production
entry
course,
and
one
things
I
want
to
do
is
enable
the
telemetry
and
I
was
just
wondering,
because
we'll
have
to
sign
up
our
security
team
in
flow
and
that's
be
straight
forwards
or
be
nice.
To
that
I
mean
looking
at
the
the
data
that
shared
I
can't
see
will
be
a
problem
bite
if
interested
it
annual
has
any
kind
of
interesting
stories
of
either
success
or
otherwise.
With
that.
B
Did
a
similar
thing
to
Kevin
I,
just
kind
of
looked
and
it
looked
pretty
innocuous,
so
I
didn't
really
ask
for
anybody's
permission
and
did
it,
but
I'm
also
on
a
small,
empty
I
mean
the
main
problem
I've
had
with
telemetry.
Is
that
when
it
falls
over
like
when
their
ELQ
staff
goes
down,
it
puts
your
whole
cluster
into
warring
mode.
That
is
hard
to
clear.
A
B
D
B
D
F
D
Issue
was
that
when
the
telemetry
server
goes
down,
there
was
no.
There
was
no
exception
handling
for
that
in
the
in
the
telemetry
module
and
it
was
happened
a
couple
of
times,
I
think
yeah.
It's
happened
a
couple
of
times,
I
think
they
put
in
something
but
I,
don't
know
exactly
the
state
of
that
yeah.
B
They've
merged
certain
fixes,
but
it
seems
like
it's
kind
of
been
like
whack-a-mole:
they
have,
they
haven't,
really
nailed
it
yet,
but
when
it
does
happen,
I
mean
it's.
It's
harmless
like
it's!
It's
never
gonna!
It's
isolated
enough
in
a
module
where
it's
it's
not
going
to
cause
of
production
out.
It's
just
annoying
to
get
paged
and
have
it
complained
so
loudly
so
they've
been
reducing.
They've
also
been
reducing
those
failures
from
a
heir
to
a
warning
or
something
like
that.
B
B
Only
have
plans
to
go
next
month
so
whenever,
whenever
15
to
3
is
out,
if
nobody's
up
great
a
15
to
2,
whatever
that's
out
I'm
gonna
give
this
spin
and
I
have
a
pretty
mixed
workload
between
a
pretty
large.
You
know
about
two
and
a
half
petabytes
in
object,
storage
or
browse
gateway
and
then
about
half
of
petabyte
in
death
offense.
B
I
Did
sound
right
here
so
I
have
a
1
point:
1,
petabyte
and
350
OSD
stressed,
which
is
an
octopus
unfortunate
thing
to
do
now,
but
in
F
an
issue
so
far,
it's
quite
heavily
used
for
the
CFS
and
for
the
job
processing,
so
data
processing,
so
typically
10
5
to
10
gigabytes
per
second
all
the
time.
So
no
issues
of
our
and
actually
it
works
much
better
than
now
the
rest
but
I'll
check.
We
have
this
to
the
corruption
bugs
anyway,
since
the.
B
J
I
Exactly
the
same
for
us
may
I
just
mention
one
thing,
because
I've
seen
several
issues
when
you
have
really
heavy
right
in
the
staff
better
service
and
if
you're
doing
maintenance
or,
let's
say
recovery
of
disks
or
if
you're
rebooting
the
notes
or
upgrading
them,
then
it
can
happen
quite
a
lot
of
time.
If
you
don't
have
enough
back
pills
per
OSD,
let's
say
the
degraded.
The
number
of
degraded
blocks
control
in
time,
because
the
some
placement
groups
which
are
degraded
can
stay
degraded
because
they
are
waiting
for
the
slots.
I
Alright-
and
this
might
be
automated
somehow
it's
not
important
when
you
do
rebalancing
or
miss
placement
doesn't
matter,
but
for
for
the
grading
part
is
really
critical.
So
let's
say
we
have
got
quite
heels
to
do
this
whenever
we
do
maintenance
to
increase
it
to
16
threads
per
SD
or
even
32,
to
get
it
faster,
faster
recovery
as
soon
as
possible.
But
this
might
be
automated.
J
Found
that
when
I
was
doing
the
repairs
on
the
mall
and
rebuilding
the
corrupted
ones,
that
yeah
certainly
widely
with
the
default
setting
of
one
that
was
painfully
slow
but
think
I'll
crank
it
up
to
an
eight,
and
it
was
much
much
better
than
doing
one
two
time
and
it
did
correctly
do
the
degraded
ones.
First,.
D
J
No,
not
quite
you
can
set
it
to
false,
but
there
is.
There
is
a
small
chance
that
something
will
go
horribly
wrong.
There
I
don't
quite
understand
what
that
is.
I
know,
I
was
concerned
about
it,
but
he
said
it
was
a
very
small
chance,
so
we
took
I
injected
it
on
the
fly
into
one
OSD
and
ran
that
a
little
while
and
I
yeah
I
tried
another
one
and
in
the
end,
I
just
put
it
in
globally.
J
Didn't
do
no,
but
no
effects
sped
effects.
That
I
saw
so
then
it
was
just
a
matter
of.
If
you
leave
it,
we
don't
leave
it
to
run
for
a
while,
while
it
set
to
false
then
eventually
as
it
reuses
and
recycles
blocks
out
of
the
world,
I
get
written
with
correct
talent
data,
so
one
option
is
just
to
you
know,
reset
it
and
then
just
leave
the
things
for
a
while.
J
J
Came
back
so
there
was
a
certain
amount
going
on
I.
Don't
know
how
much
the
type
of
media
that
the
world
was
on
effects
that
we
had
as
run
in
D&E,
because
the
bug
was
the
synchronous,
I/o
library,
processing,
iOS
out
of
order
and
those
over
lacked.
So
you
could,
because
it
was
pre
extending
we'll,
try
and
extended
it
further.
Then
you
get
an
overlapping
one
coming
in
later
for
the
bit
yet
extended
into,
but
they
got
done
in
the
wrong
order,
see
you
even
at
the
zeroes
there.
J
J
But
in
the
good
news
is
that
having
gone
through
each
one
once
we
recreated
it
not
important
to
have
these
extend
sitting
global
at
that
point,
so
you
don't
ever
run
and
if
you're
creating
a
new
one,
you
don't
want
it
to
be
running
at
all
with
extending
in
they
ran
okay.
After
that,
we're
still
running
with
a
pulse,
obviously
and
I
assume
that
there
was
a
minor
performance
hit,
but
it
hasn't
been
too
noticeable.
J
I
J
A
C
And
octopus,
it's
a
little
bit
better
because
they
use
Putman
and
I,
find
the
containers
just
to
be
cleaner,
so
I'm,
a
social
site
I'm,
not
really
a
safe
user
I
just
like
play
with
it
in
the
lab,
but
I've
done
it
a
few
times
with
the
containers
and
just
absolutely
hated
it,
and
then
I
did
it
is
October
this
and
I'm
like
okay.
This
is
heading
in
the
right
direction.
B
B
You
know
like
if
I'm
trying
to
do
something
special
like
I
have
a
house
that
has
you
know
sixty
spinning
discs,
and
it
has
you
know
six
nvme
drives
to
you
know:
do
the
journal
cashing
or
something
like
that
like?
How
is
it
going
to?
Is
it
going
to
be
smart
enough
to
actually
correctly
use
all
that
hardware
like
I
I,
don't
see
how
that's
gonna
happen.
H
Yeah
from
our
perspective,
we
have
quite
a
interesting
East
Coast.
We
use
Seth
Morris,
like
an
object
store
first
buffer,
that
we
display
deploy
onto
ran,
not
HPC
cluster,
the
idea
of
her
dock
rising
as
well.
On
top
of
that,
we
want
to
be
as
close
to
like
bare
metal
as
possible.
So
it's
a
bit
disconcerting
on
one
side,
boys
for
our
main
cluster,
we're
not
lost.
B
H
So
we
have
it
so
when
someone
wants
to
so,
we
have
currently
go
into
semi
production
there.
When
someone
wants
to
spin
up
a
job
that
would
go
on
the
cluster,
so
instead
using
our
file
system,
you
deploy
the
object
store
as
they
create
their
job.
They
can
then
run
it
inside
the
object
store
and
then
they
can
cut
that
out
somewhere
and
then
the
object
store
then
closes
that
insert
just
brace
yourself
down
again,
so
they
can
have
object.
Storage
on
the
fly,
that's
fast!
That's
how
we're
using
you
know
the
moon.
C
C
H
So
it
can't
use
a
proper
Ram
disk
in
sense
because
I'm
staff
uses
of
young,
so
you
can.
We
have
our
own
in-house
kernel,
module
that
presents
Ram
as
a
Bob
device
currently
and
then
there's
also
the
links
teapot,
which
is
BRD.
So
what
we
do
is
on
the
so
when
the
user
asks
awesome
around
now
for
the
system,
they
didn't
see
how
much
one
they
want
and
then
it's
let
up
into
Orestes
on
the
host
so
they're
using
dynamically,
so
they
can
request
as
much
as
they
want.
F
A
B
A
Hope
not
a
lot
of
us
like
the
staples
every
time.
Something
you
know
goes
crazy.
My
kubernetes
cluster,
it's
just
the
biggest
pain-in-the-ass
to
troubleshoot,
usually
yeah
I,
fear.
I
know
the
stuff
won't
be
including
that
is,
but
you
still
talk
or
apartment
or
that
layer
of
ambiguity
that
it
gives
you
and
the
troubleshooting
at
least
the
simplicity.
It
could
be
nice.
B
B
B
Just
this
kind
of
an
addendum
for
the
stadium,
I
put
a
note
about
you
know
we're
going
to
move
to
probably
relate
concurrently
with
doing
some
of
our
upgrades
just
because
it
seems
like
there's
new
features
coming
out
that
are
not
going
to
be
supported
around
7:00,
like
we
don't
run
a
dashboard
today.
Actually
we'd
like
to
do
that.
We
have
like
a
few
Griffin
dashboards
that
we
have
set
up
on
our
own,
but
I'm
just
kind
of
interest
in
seeing
what's
it
within
the
dashboard.
B
M
A
Please
desk
I'm
not,
but
if
it
has
crossed
my
mind
many
times,
given
the
ease
of
installation
from
dos
Santos
other
people's
to
give
something
a
bit.
L
A
M
A
M
Yeah,
it's
I
read
about
we're
not
being
to
notice
and
I
prevent
that.
Then
we
read
on
this
part
in
the
inset
of
the
spirit,
we're
not
long-running
clients,
and
so
it
is
handling
a
better
better.
It's
also
this
we're
getting
the
slow
requests
or
things
where
users
are
running
sort
of
locking
themselves
sort
of
not
voting.
We
leave
him
out
for
all
the
right
locks.
M
D
Have
one
I've
been
looking
into
this
recently
because
on
our
service
s
we
get
this
pretty
often,
and
so
what
I?
What
I
learned
is
that
there's
an
option
in
luminous
and
not
an
mimic,
will
also
still
a
Nautilus.
It
controls
like
the
maximum
number
of
caps
that
will
be
trimmed
from
the
from
the
LRU
cache.
D
This
might
not
be
exactly
related.
I
might
be
confusing
two
things,
but
there's
an
option
to
control
a
number
of
taps
which
determines
from
the
Owlery
cache
each
mvs
tick
and
in
a
new
luminous
and
mimic
this.
This
is
like
thirty
two
thousand
and
the
tip
is
every
five
seconds
and
in
Nautilus
it's
still
30
mm,
but
the
tick
is
every
one
second,
so
it
terms
a
lot
more
quickly.
D
So
that
yeah,
okay,
so
I'm
I'm,
probably
the
two
issues
our
confuse
are
always
like.
They
always
happen
at
the
same
time.
For
us
like
will
get
client
isn't
responding
to
to
the
client
caps
with
replications
and
then
also
the
like.
The
the
MDS
cash
can
go
can
be
too
large
to
be
larger
than
what
is.
It
spoke
to
those
three
so
on
our
cluster,
which
we
haven't
upgraded
yet
and
oddly,
we
I
just
increase
that
up
the
this
value
from
32k
to
like
400,000
or
something
and
it,
and
it
fixed
that
memory
at
issue.
D
M
So
we're
getting
these
kind
of
large
Olmec
obliques
warnings,
but
it
seemed
me
in
in
the
metadata
sore
so
seems
to
be
the
right
like
a
file
that
keeps
track
of
how
many
fives
or
open
or
something
like
that,
and
that
keeps
growing
and
I
read
a
post
on
the
forums
that
you
should
just
increase
it
and
not
to
I.
Don't
know
400
thousand
or
something
at
the
moment.
So
this
is
there
I,
guess,
there's
a
drawback
of
sending
it
too
high.
But
what
can
one
do
really
is
their
answer.
M
I
I
get
warnings
that
there
is
large.
All
map
objects
instead
also
larger
than
what's
supposed
to
be,
or
something
at
that
Sutton's
in
the
in
the
metadata
store
and
when
I
look
at
the
obvious.
This
could
be,
it
seems
to
be
the
file.
It
keeps
track
called
open
pipes
or
something
like
that.
So
it's
naturally
growing
since
I
had
like
500
clients
and
they're
doing
they
had
lots
of
files
open.
M
So
that's
where
I
guess
set
is
keeping
track
of
open
fast,
so
that
makes
sense
that
it
was
the
one
I
put
on
and
the
more
open
files
they
have
so
but
I
was
thinking
about.
Is
that
I
guess
there's
a
drawback
and
it's
related
to
recovery
if
I
had
it
too
large,
but
but
does
anyone
have
an
excellent
as
I
said
anyone
increased
it
and,
and
has
it
been
safer,
has
have
you
been
running
into
problem
problems,
I?
Think
it's
a
400k
at
low.
B
B
It
could
be
a
problem,
but
I
think
there
are
a
lot
of
instances
where
the
set
developers
are
too
conservative
about
where
they
check
these
limits
on
large
o
map
size
and
also
they
might
be
optimizing
for
smaller
clusters
than
what
some
of
us
are
running.
I
don't
know.
Yes,
I
have
increased,
it
was
not
a
problem.
That's.
M
B
A
A
A
That
could,
if
you
use
that
main
index,
that
could
be
a
nice
place
where
yeah.
If
everybody
must
have
put
down
you
know
name
email.
What
you're
using
versions
giveaway
object,
store
to
store
file
store.
You
know
sefa
fast,
whatever
cluster
size
of
like
drive
types,
general
condition.
Maybe
that
could
be
useful.
So
if
anybody
wants
to
ping
somebody,
they
can
yeah
doctor
Mao.