►
From YouTube: February 2020 OpenZFS Leadership Meeting
Description
At this month's meeting we discussed: Thread priorities on Linux; Feature flag activation bug; Conferences
Details and meeting notes: https://docs.google.com/document/d/1w2jv2XVYFmBVvG1EGf-9A5HBVsjAYoLIFZAnWHhV-BM/edit#
A
B
C
We
have
old,
say
musts,
probably
like
on
the
Loomis
or
like
when
all
camera
parents
have
high
bridges
in
the
user
space
it's
possible
to
set
a
priority
to
lower,
but
kernel
threads
never
can
never
work
as
the
timeshare,
so
they
practically
say
fixed
priority.
So
it
won't
even
may
exempt
same
as
user
space.
D
B
B
A
Yes,
me
in
general,
I
think,
like
I'm,
not
I'm,
not
inherently
opposed
to
the
performance
being
different
on
different
platforms
like
different
operating
systems
do
things
differently,
but
it
does
seem
like
even
just
looking
at
Linux
in
isolation,
the
there's
at
least
some
of
these
threads
that
we
are
create
that
ZFS
is
creating
that
are
effectively
not
very
low
priority.
Like
you
know,
user
threads
can
easily
take
precedence
over
the
threads
that
we're
creating
it
and
the
I
get
I
think
it's
there's
two
classes
right.
A
One
is
threads
that
we're
creating
that
are
like
doing
work,
that
the
user
explicitly
requested.
Brand-N,
like
the
same
thing
where
it's
like
you
could
have
the
user
processes
thread,
is
that
some
priority
we're
creating
another
thread
to
do
the
work
for
it
like
the
send
prefetching
or
that's
going
up
and
doing
stuff.
That's
part
of
the
send
operation
that
you
asked
us
to
do.
Those
are
actually
running
at
lower
priority.
A
Then
the
user
thread
is
running
at
which
seems
like
that
kind
of
seems
wrong
on
its
face,
then
separately,
we
have
like
the
background
threads
that
are
kind
of
doing
work.
The
whole
system,
like
the
CIO
threads
and
the
pink
thread
or
threads
down
and
those
like
I,
think
you
could
kind
of
make
arguments
either
way
like
hey.
We
should
be
super
polite.
We
should
not
be
hi
parody
this
background
stuff,
because
otherwise
we
can
just
like
Agra
or
all
of
you
know
the
user
CPU.
A
B
When
it
comes
to
things
like
the
the
send
threads
that
we're
spinning
up
sort
of
to
serve
a
user
thread
or
a
user
prompted
command,
it
would
seem
like
a
reasonable
default
would
be
to
try
to
just
copy
whatever
priority.
The
send
threat
has
like
the
the
main,
send
threat
and
just
to
use
that
as
the
priority
for
all
of
its
child
threads.
A
D
Don't
know
what
the
right
priority
would
be
offhand,
I
I
think
would
be
nice
to
maybe
have
it
behaved
reasonably
the
same
on
all
platforms,
so
maybe
it's
worth
investigating
putting
some
kind
of
generic
interface
in
front
of
these
to
get
the
behavior
we
want
I,
don't
know
I
I,
think
we've
had
issues
up
in
the
past
before
tackling
different
performance
problems
and
you
tuned
for
one
case.
You
end
up
breaking
another
right.
D
D
B
I
mean
my
instinct
for
stuff,
like
the
background
threads
is
that
they
could
be,
even
if
we
don't
want
to
try
to
mirror
the
behavior
that
we
have
on
illumos,
where
we
are
some
higher
priority
than
all
user
land
threads.
We
could
try
to
have
a
behavior
where
we're
at
least
not
like
the
lowest
priority.
You
know
we're
a
few
notches
above
there,
so
that
that
way
like
default
threads
would
still
you
know,
have
priority.
But
then,
if
somebody
was
trying
to
be
super
nice,
we
would
get
to
go
out
of
them.
A
A
I
think
that
it
does
kind
of
make
sense
to
be
at
least
as
high
as,
if
not
higher
priority
than
user
threads,
so
that,
if
somebody
like
say
Sony
who's,
not
doing
DFS
stuff
they're,
just
you
know
compressing
their
video
or
whatever
and
they're
like
oh
I'm,
gonna
run
at
the
highest
party.
That
I
can
and
then
therefore,
like
totally
starve
out
somebody
who's,
you
know
doing
writes
you
know
that
doesn't
seem
like
a
great
behavior
yeah
yeah.
E
D
I
was
gonna,
say,
yeah
I,
think
that
makes
sense.
I,
don't
think
we
want
to
start
the
DFS
ribs
in
general,
at
least
like
Matt
was
saying
we
have
a
couple:
caddy
Razia
right
of
threads
that
have
important
work.
That
needs
to
get
done
now
and
then
some
of
the
lower
priority
background
stuff
that
you
know
we
don't
care
what
eventually,
when
it
gets
to
it.
Like
a
I,
don't
know,
I'll
be
the
background
half,
but.
A
There's
not
many
of
those
threads
that
we
don't
care
about.
The
performance,
though,
like
you
know,
even
the
sink,
even
though
it's
like
totally
asynchronous.
Nobody
is
like
you
usually
know
it
is
explicitly
waiting
on
it,
but
the
performance
of
that
thread
directly
impacts
like
your
overall
write
bandwidth
that
we
can
do
so.
You.
A
E
F
E
A
B
B
D
A
F
On
yeah
so
well
testing
the
changes
for
Zed
standard.
We
found
that
if
you
set
the
compression
type
on
a
data
set
to
Zed
standard,
but
then
don't
write
any
blocks,
it
doesn't
activate
the
feature
flag.
So
then,
if
you
export
that
pool
and
then
import
it
on
another
machine,
that's
running
like
0.8
or
whatever
and
doesn't
know
what
that
said.
It
is,
then
you
hit
assertions
that
you
know
the
compression
type
is
out
of
range
of
the
enum
and.
F
B
Go
ahead.
Well,
as
you
can
say
from
your
initial
description.
Definitely
it
sounds
to
me
like
the
sort
of
most
objectively
or
the
most
correct.
Behavior
is
the
sort
of
reference
tracking
when
the
property
has
been
set
and
when
objects
have
been
sent
with
that
as
their
compression
type,
but
I
also
I,
believe
you,
when
you
say
that
that's
more
complicated
to
do
I'm,
just
saying
that
that
sounds
like
sort
of
the
the
one
that
most
accurately
reflects
the
way
that
I
would
expect
the
system
to
work.
B
A
Yeah
I
agree
with
what
Paul
said
and
I
I
think
we
should
look
at
that
solution.
Some
more
did
you
like
I,
don't
think
it's
inherently
intractable
because,
like
when
we're
changing
the
property,
we're
deleting
the
dates
that
or
whatever
we're
in
Sinking
like
we
have
to
go
to
sticking
contacts
to
do
that
and
syncing
context
is
where
we
would
need
to
be
to
increment
decrement
get
your
ref
count
as
well.
F
A
F
A
A
A
A
A
Think
that
we
should
you
know
somebody
should
go,
spend
the
time
to
investigate
these
code
paths
and
I
think
that
the
the
two
ways
that
we
thought
about
it
before,
where
one
is
we
just
bump
the
ref
count
kind
of
using
a
similar
mechanism
for
all
of
the
file
systems
where
it
gets
where
the
property
takes
effect.
When
you
do
the
DFS
set
and
the
other
way
that
we
had
talked
about
last
time
was
maybe
we
have
a.
A
We
say
that
the
ref
count
is
like
the
sum
of
all
the
data
sets
that
have
used
this
new
texting
algorithm,
which
is
when
Turing
now
plus
anywhere
that
we
have
explicitly
set
the
property.
So,
in
other
words
like
when
you
do
this,
if
I
set
inherits
down,
we
would
pump
the
ref
count
by
one
and
then,
when
you
change
the
when
you
change
the
property
again
later,
we
would
decrement
that
one
I
think
that's
like
a
little
bit
more
complicated.
A
So
preferably,
we
could
do
it
with
the
former
mechanism
of
just
saying
like
instead
of
waiting
until
we
write
the
block
first
lock
to
my
boot,
we
go
and
explicitly
like
you
know,
enable
it
for
this
data
set
on
everything
that
inherits
the
property.
When
we
do
it,
we
need
you
to
go.
You
know
in
here.
Yes,.
E
F
That
seems
more
reasonable,
because
that's
the
other
way
the
problem
might
be
if
you
set
it
that
activates
the
feature.
So
when
the
first
block
is
born,
it
doesn't
increment
the
ref
count,
because
the
feature
is
already
activated
and
then
you
changed
the
checksum
algorithm
to
different
value,
and
that
would
decrement
it,
but
there'd
actually
still
be
blocks
they're
using
it
and
they
wouldn't
be
ref
counted
correctly.
Yeah.
A
B
Quick
question
just
about
one
case
and
whether
it
covers
it,
if
you
were
to
set
the
property
on
a
data,
set,
create
a
file
with
no
blocks
just
like
touch
a
file
and
then
unset
the
property.
Would
we
have
an
erroneous
traffic
out
there,
because
I
think
the
file
would
still
have
the
compression
algorithm
set
to
the
Z
standard,
but
there
wouldn't
be
any
blocks
compressed
with
it.
A
A
A
A
Don't
think
that
should
be
the
solution
to
the
this
specific
problem,
because
I
think
it's
going
to
be
like
a
bunch
of
more
complicated,
and
you
know,
as
we
saw
the
problem,
shows
up
even
when
you
the
data
set.
Is
it
really
using
it?
Could
you
flag
it's
just
that
the
property
is
set,
so
you
having
the
property
set
and
doing
like
you
know,
ZFS
lists
or
you
just
get
all
we
get
confused
about
this
property
value
that
we
don't
know
about.
A
So
we
can
definitely
work
out
like
chipping
away
at
that
kind
of
stuff,
like
saying
oh,
like
in
new
versions
of
ZFS
like
we
can,
just
if
we
see
a
property
value
for
maybe
for
specific
properties
or
whatever
that
we
don't
understand,
then
we're
just
gonna
say
you
know
unknown,
and
then
you
get
the
default
like
I
think
that
might
be
a
reasonable
thing
to
do.
But
we
should
probably
consider
that
kind
of
orthogonal
to
these
other.
Two,
the
like
very
real
and
concrete
problem
that
we
have
now,
which
is
like
this
bug.
A
A
We
kind
of
got
that
connection
through
the
most
folks
who
worked
at
giant
and
they
were
listing
also
the
most
infrastructure,
like
you
know,
the
most
website
and
whatnot
I'm
giant
enjoying
it
was.
You
know,
providing
that
service
without
any
charge,
so
that
was
nice,
but
most
of
our
kind
of
contacts
within
that
community
of
are
working
at
doing
any
work
so
kind
of
for
a
long
term.
You
know:
there's
no
problem,
there's
no
immediate
problem,
they're
still
hosting
it.
A
There
are
asking
us
to
move
off
for
anything,
but
just
in
terms
of
making
sure
that,
like
eventually
they
may
change
their
minds
or
they
might.
You
know,
forget
about
us
there,
whatever
so
we're
looking
at
moving
into
University
of
Washington,
which
also
hosts
a
bunch
of
other
like
common
and
open
source
projects,
websites
and
other
you
know,
static,
hosting
infrastructure
and
stuff.
A
A
D
I
should
say:
I
have
one
thing,
just
a
quick
thing:
if
I
can
call
for
reviews,
there's
two
features
outstanding
that
are
still
looking
for
some
additional
reviewers
if
they
want
there's
the
encryption
performance
work
that
was
in
97
49,
it's
in
pretty
good
shape
and
there's
a
persistent
LT
Ark
and
the
FS
and
likes
issue
95
82
both
are
really
close
for
a
final
reviews.
So
if
you
have
time,
take
a
look,
the.
A
A
Oh
yeah
I
know
we
discussed
that,
maybe
one
or
two
months
ago,
out
of
this
meeting
yeah
and
do
you
know,
did
we
send
out
email
on
the
mailing
list
about
changing
the
default
I,
don't
believe
so
all
right?
Let's
do
that
submit
you
know
on
the
open,
ZFS
mailing
list
just
that
books
are
aware
and
can
give
any
feedback
about.
Oh
I.
G
And
I
just
wanted
to
be
I
asked
about
that
I
think
yesterday,
and
just
because
I
want
to
be
sure
if
that
does
happen,
to
keep
it
consistent
since
we're
getting
real
closely
with
the
stuff
I
talked
about
in
November.
Getting
that
finished
up
and
making
that
available.
So
we
don't
want
to
start
having
people
create
a
whole
bunch
of
pools
on
CCM
and
then
all
of
a
sudden.
Here
then
it's
like,
oh
well,
now
everything's
GCM,
and
if
you
want
to
switch
well,
you
basically
have
to
you
know
recreate
your
pool.
G
G
Cool
well
good
I
did
have
a
quick
just
general
question.
Hopefully,
if
this
isn't
big
bro
I
just
wondered
just
you
know,
you
know
you
know
Fossum.
You
just
finished
up
here,
though,
over
the
weekend
I
think
this
weekend
and
was
coming
up.
I
know
scale
is
coming
up
here
in
March
and
well
I,
don't
think,
there's
anything
a
lumo,
specific
obvious
there's
a
lot
of
Linux
stuff
and
I
also
saw
some
ESD
stuff.
G
A
That's
a
great
question:
I
I
know:
I,
don't
think
that
deltax
has
a
new
plans
and
you
do
those.
For
instance,
we
Jordan
I,
spoke
at
scale,
I,
think
one
or
two
years
ago
doesn't
one
you
may
have
just
been
one
year
ago,
yeah
so
last
year,
George
nice
look
at
scale
and
gave
a
well-received
talk.
That
was
a
good
kind
of
an
overview
of
ZFS
and
the
presenter.
The
sorry,
the
organizers,
were
very,
very
receptive
to
your
having
more
ZFS
box
there
in
the
future,
but
I.
A
If
anybody
submitted
anything
for
this
year,
I
think,
if
you're,
if,
if
folks
are
going,
then
maybe
see
if
there's
like
a
possibility
of
doing
a
boss
or
meet
up
there,
I
think
you
probably
find
a
lot
of
people
that
wanted
to
hear
about.
What's
new
with
ZFS,
you
know
anyone
who's
involved
in
the
community.
A
A
Yeah
I
mean
sounds
like
nobody
here
is
like
planning
anything
for
scale
or
fosston,
but
you
know
I'd
love
to
I'd
love
to
see
us
do
more
talks
they're
both
just
getting
the
word
out,
engage
you
with
folks
from
you
know
the
broader
community
beyond
you
know,
open,
ZFS
and
I.
Think
the
BSD
communities
are
very
well
informed
about
ZFS,
but
you're.
A
Spreading
that
knowledge
to
the
broader,
suffer
industry
and
I
think
Linux
Linux
world
specifically,
would
would
be
really
advantageous
for
the
project
as
a
whole,
and
we
I
think
that's
something
that
del
phix
would
like
to
do.
But
we
that's
just
not
our
focus
like
for
this
year
and
it's
also
something
that
I'd
like
to
see
or
the
open
ZFS
organization
be
able
to
do.
You
know
we,
in
the
medium
to
long
term,
to
like
have
you
to
build
up
funds
that
we
could
send
people
to?
A
You
know
Spiga
conferences
that
don't
have
the
funds
to
do
so.
You
know
from
the
company
or
individually
and
also
so
like
you'll.
Potentially,
you
have
have
or
participate
in
your
booths
and
stuff
like
that.
I
know
that,
like,
for
example,
the
FreeBSD
community
has
done
a
pretty
good
job
of
like
having
having
a
booth,
that
Linux
conferences
and
other
general
technology
conferences
to
kind
of
get
the
word
out
about.
A
F
Buzzed
him
for
the
last
couple
of
years,
the
FreeBSD
and
other
most
tables
have
been
side
by
side
and
giving
out
ZFS
stickers
and
so
on.
Cool
I
wasn't
able
to
go
this
year,
but
last
year
at
FOSDEM,
I
Dave
Zetas
fests
have
the
art.
Caching
works
talk
on
the
in
the
main
track.
In
addition
to
my
usual
talk
about
ZFS
in
the
BSD
dev
room,
but
this
year
I
wasn't
able
to
go.