►
From YouTube: Grafana Tempo Community Call 2022-09-08
Description
Join our next Tempo community call: https://docs.google.com/document/d/1yGsI6ywU-PxZBjmq3p3vAXr9g5yBXSDk4NU8LGo8qeY/edit
Discussed this month:
- Parquet/TraceQL progress
- Helm chart changes/Upcoming operator
- Azure fixes!
A
All
right,
so
this
is
the
september
tempo
community
call.
We
have
pretty
good
crowd
here,
appreciate
everybody
showing
up.
We
have
a
lot
of
smaller
stuff
to
talk
about
today.
Just
kind
of
progression
on
parquet
progression
on
traceql
got
some
news.
There,
I'm
going
to
farm
a
fair
amount
of
this
out
because,
like
I
said,
I've
been
in
meetings
all
morning
having
a
chance
to
review
any
of
this.
A
I
guess
I
can
talk
a
bit
about
parquet,
because
that's
kind
of
been
my
main
push
and
I
will
ask
others
to
talk
about
some
of
these
other
things
so
just
be
like
super
ready
if
you're
on
the
tempo
team,
even
if
you're,
not
maybe
I'll,
call
on
you
anyway,
let's
start
with
the
parquet
and
grafana
cloud
and
other
internal
clusters.
So
we
are
currently
running
parquet
in
all
of
our
production
clusters.
A
It's
mostly
stable,
it's
less
stable
in
our
ops
cluster,
where
we
have
a
lot
more
traffic,
but
it
we're
kind
of
struggling
a
bit
with
different
components,
but
for
the
most
part,
things
are
going
well
and
we
think,
maybe
in
two
weeks
we
can
really
kind
of
ratchet
down
a
few
settings
that
are
going
to
make
major
changes
to
cloud.
A
So
right
now
we
are
creating
parquet
blocks,
but
our
settings
are
still
kind
of
tuned
to
v2
in
cloud,
and
the
result
of
that
is
that
querying
is
not
going
to
be
that
much
quicker
than
v2.
For
the
moment
as
soon
as
we
get
kind
of
this
like
critical
mass
of
blocks
in
our
production
clusters,
we
can
change
a
lot
of
our
settings
over
to
a
kind
of
more
parquet
friendly
settings.
In
fact
the
tempo.
A
What
is
it
tempo,
1.5
blog
post
head
those
settings
I'll
go
ahead
and
put
a
link
to
that
in
the
docs?
I'm
sorry
in
this
agenda
doc
here
so
there's
just
a
handful
of
settings
that
make
querying
way
better
on
parquet,
and
we
have
some
recommendations
in
in
that
on
that
blog
post
there
likely.
A
So
parquet
is
moving
forward.
It's
kind
of
a
two
step
forward,
one
step
back
thing.
If,
I'm
being
totally
honest,
we
make
some
progress,
we
add
some
stability,
we
get
some
speed,
then
there's
you
know
a
little
problem
comes
up
and
we
have
to
kind
of
like
reset
a
bit
and
get
it
fixed
up,
but
things
continue
to
move
forward
and
we
are
looking
to
cut
2.0,
hopefully
by
the
end
of
the
year,
hopefully
november
or
so,
with
with
the
parquet
as
a
production
ready
back
end.
A
It
is
kind
of
ambitious,
we'll
just
see
how
things
go
and
we'll,
of
course,
always
keep
you
all
up
to
date
on
this
call
about
how
or
how
that's
looking
right.
Now
it's
right
now,
things
are
going
slower
than
we'd
like
and
we'll
just
keep
talking
about
it
over
the
next
few
months
as
we
go
forward,
but
we
are
running
it
in
production.
So
that
is,
you
know
something
I
suppose
tracy
yeah.
B
Oh
yeah
marty
jumping
here
for
parkland
we're
we're
we're
making
a
lot
of
learnings
as
we
go
along
like
there's
so
many
tunables
and
different
things
you
can
do
with
parquet,
like
the
group
size
max
block
size,
different
encodings,
compression
algorithms,
so
we're
trying
a
lot
of
different
things
like
we're
still
finding
like
what
is
working
best.
B
I
would
say
that
the
defaults
in
tempo-
probably
right
now,
are
not
the
best
like
for
sure
so
we'll
def
we'll
definitely
be
coming
back
and
improving
some
of
those
yeah,
but
I
was
wondering,
is
anyone
using
parquet?
Has
anyone
like
had
a
chance
to
turn
that
on
and
have
anything
to
share
good
or
bad.
A
I
know
that
alton
has
because
he's
filed
like
three
bugs,
because
the
single
binary
is
panicking
for
him,
but
we'll
we'll
figure
that
out
the
distributed
mode,
doesn't
panic,
but-
and
then
we
have
this
so
one
thing
the
parquet
library
does
is
use
a
lot
of
unsafe
pointers
for
speed,
and
so
we
have
some
garbage
collection
issues
and
we
keep
going
back
and
forth
on
the
stability
of
the
single
binary.
A
A
On
that
note,
oh
here
we
should
show
this.
This
is
we're
currently
seeing
one
plus
terabytes
per
second,
I'm
gonna,
say
effective,
terabytes,
effective,
effective,
eb
per
second
scanned
with
park
a
and
what
I
mean
by
effective
is
basically
you
know.
The
b2
blocks
would
have
to
pull
the
all
data
all
the
time
right
and
it
would
just
on
marshall
proto
and
then
assert
your
search
conditions
and
keep
or
like
keep
a
trace
or
drop
a
trace.
A
So
if
we
continue
to
use
those
metrics
like
just
the
total
size
of
the
block,
we're
scanning
over
a
terabyte
a
second
right
now
in
our
ops
cluster,
so
queries
are
in
our
way
better
spot.
They
go
much
faster
and
we're
pulling
far
less
data
and
we're
just
continuing
to
like
tune
and
improve
this
over
and
over
again
and
one
time
I
did
get
around
four
terabytes
a
second.
I
don't
know
it
was
some
kind
of
fluke,
but
don't
tell
anyone
that,
because
it's
I
I
can't
reproduce
it.
A
But
one
time
I
got
like
around
four
terabytes,
a
second
and
I
think,
it'd
be
awesome
to
see
that
more
consistently
like
four
plus
terabytes,
a
second
on
queries.
So,
while
we
are,
you
know
struggling
a
bit
with
this
and
we're
continuing
to
move
forward.
Just
the
promise
of
the
back
end
is,
you
know,
continues
to
be
there
and
we
have
a
lot
of
hopes,
moving
forward
for
both
traceql
and
just
performance
in
general.
A
B
Hey
yeah,
so
we
started
implementing
traceql.
I
think.
Maybe
we
talked
about
this
last
time
or
maybe
we
said
we
were
going
to
start
well,
we
have
started,
I
don't
know,
maybe
it's
halfway
like
it's
going.
Well,
I
think.
Actually
what
we
were
thinking
here
is,
I
think
we
invited
andre,
hey
and
you
have
maybe
some
really
cool
ui
work.
B
That's
in
progress
too,
so
we're
kind
of
working
in
parallel
between
the
front
end
of
the
back
end,
and
I
thought
there
was
some
really
cool
stuff
going
on
there,
like
with
the
traceq
editor
and
things
like
that.
C
Cool
so
yeah.
A
C
All
of
this
is
still
pretty
much
experimental.
It's
really
not
stable,
bugs
may
appear,
but
yeah
there's
been
some
progress
on
actually
building
the
the
editor.
How
to
complete
syntax,
highlighting
which
I
think
is,
is
pretty
cool
and
should
help
everyone
understand
risky
well
and
write
some
some
really
good
queries,
so
that
the
main
thing
is
that,
as
soon
as
we
start
to
build
a
spam.
C
We
are,
we
have
all
of
the
completion
options
available
scopes.
Resource
fan,
the
intrinsic
values
are
possible,
like
the
ratio
name
status
and
then
all
of
the
tags
that
we
retrieve
from
the
the
tempo
api-
and
I
think
this
works
pretty
well.
It
auto
completes
operators,
the
the
possible
values
again
retrieved
from
the
tempo
api,
and
this
should
make
writing
queries
a
lot
easier.
C
D
C
But
yeah,
I
think
this.
This
is
working
pretty
well
for
for
just
an
experimental
status.
C
But
yeah
for
the
editor.
This
is
our
progress.
It's
again
getting
getting
stuff
from
the
the
the
tempo
api,
but
also
for
the
response.
What
we're
trying
to
do
is
build
something
that
shows
us
the
traces
and
then
which
spans
we
we
matched
using
the
query.
C
So
we
we
want
to
show
a
quick
look
at
the
few
spans
like
two
three
four
maximum
inside
that
that
trace
where
we
can
see
which
attributes
matched
so
that
the
user
can
have
a
quick
look
at
and
how
the
query
performed
and
and
what
the
results
are.
This
will
all
be
linkable.
Of
course,
this
is
just
using
some
mock
data,
some
random
data
that
we
generate,
but
it
will
be
some
some
real
data
very
soon.
B
Hey
yeah,
so
I
think
this
is
part
of
the
traceql
language
design.
So
I
was
bringing
up
a
link
here
for
kind
of
like
a
intro
to
the
language
and
all
of
those,
the
dot,
the
dots
that
pre
precede
the
attribute
name
and
the
query,
or
I
mean
to
actually
search
both
places.
So
it's
searching
the
span
level
and
the
resources
in
that
the
span,
the
if
they're
specified
to
both
the
spay
level
takes
precedence.
B
So
it's
searching
everywhere,
but
you
can
actually,
with
those
parts
of
language,
do
resource
dot
and
then
it
only
it's
scoped
only
to
look
at
the
resource
or
the
span
so
like
by
default.
It
should
be
as
easy
to
use
as
like
kind
of
like
the
tag
search
that
we
have,
but
you
can
also
dig
a
lot
be
more
precise
for
that
kind
of
stuff,
and
then
I
was
gonna
post
a
link
here
for
kind
of
like
the
stock
here.
So
this
is
an
intro
to
the
language.
B
It
doesn't
have
like
the
full
capabilities
and
everything
that
that
we
have
but,
like.
I
think
it's
a
good
good
intro
to
read
and
kind
of
catch
up
on
these
concepts.
E
B
Oh
yeah,
another
thing
here:
can
you
pop
up
the
operator?
We
also
have
like
a
lot
more
expression
there
so
currently
for
tags.
Sorry,
for
the
equal
sign,
yeah
we'll
be
able
to
do
integer,
math
string
regex,
so
you
can
handle
case
sensitive
your
insensitivity
things
like
that.
It's
going
to
be
a
lot
more
powerful,
too.
A
F
A
Is
going
to
be
kind
of
enabled
here
is
that
the
amount
of
work
you
ask
of
your
backend
is
going
to
be
kind
of
transparent
to
you
like
it
makes
sense
to
the
user,
and
you
can
train
your
users
and
yourself
to
write
efficient
queries.
If
you
write
a
query
that
says
dot,
name
or
say
dot
tag,
then
it's
going
to
look
in
both
resource
and
spam.
It's
going
to
pull
a
lot
more
data.
B
Yeah,
I
think
that's
really
cool
part
of
the
attributes.
I
guess
we
did
touch
on.
That
was
our
idea.
There
was
to
actually
bring
back
the
parts
of
the
span
that
are
relevant
to
the
query,
so
there's
still
some
things
to
figure
out
there,
but
basically,
if
you're
filtering
on
an
attribute
while
we've
scanned
through
and
have
found
that
attribute
we'll
go
ahead
and
return
it
to
the
ui.
So
you
can
see
it.
I
think
that
doesn't
make
sense
for
things
like
equals.
B
You
know
method
equals
get
right,
there's
only
one
value,
but
if
you
have
something
like
a
regex
or
a
greater
than
then
you
have
a
range
of
values
and
we're
bringing
that
back.
So
it's
useful
to
see
what
was
the
actual
value
that
it
matched
yeah.
There's,
definitely
definitely
a
ton
of
stuff
here
and
it's
pretty
exciting.
B
A
Cool,
let's
move
on
to
additional
work.
We've
recently
moved
forward
to
go
119..
I
think
we're
mostly
on
top
of
that,
it's
been
out
for
what
how
long
month
or
two
not
too
long
or
am
I
way
behind
couple
months.
I
want
to
introduce
one
of
our
newer
team
members,
jenny
she's,
going
to
tell
us
about
this.
She
super
doesn't
want
to
because
you
know
she's
kind
of
new.
That's
what
she's
welcome
to
make
me
talk
about,
go
on
team
119.!
F
Hi,
I'm
jenny
so
one.
I
think
that
the
code
was
not
that
particularly
different,
but
golang
ci
gave
us
a
lot
of
lensing
problem,
so
they
also
needed
to
upgrade
their
thing
to
1.19
and
I
don't
think
they
predicted
that
there
were
going
to
be
a
lot
of
issues
with
that.
So
their
first
release
of
one
night
support
was
kind
of
terrible
and
then
we're
using
their
latest
version
of
like
1.49,
which
now
they
got
rid
of.
F
They
got
rid
of
a
few
of
those
checks
because
they
were
deprecated
and
the
the
original
contributors
were
just
not
responding
anymore.
So
I
guess
they're
just
getting
rid
of
those
checks
to
get
those
gloves
down.
But
it's
it's
fine.
Now
everything
pass
on
the
ci
job,
so
I
think
we're
good
to.
A
Appreciate
that
work,
I
asked
jenny
to
upgrade
us
to
go
on
19,
expecting
it
to
be
like
a
find,
replace
118
to
119,
and
it
turned
into
like
a
week
of
work
of
like
tracking
down
a
million
weird
go,
go
like
ci
linter
issues
and
ones
that
weren't
even
really
issues.
So,
thanks
to
her
for
getting
us
moved
forward,
119.
F
A
A
D
Yeah,
so
for
the
last
few
months
you
know
the
other
database
teams
have
been
sort
of
working
on
helm
and
it's
been
more
important
to
the
community
seems
like
a
lot
of
people
deploy
with
helm,
even
though
we
internally
have
been
json
users
for
a
long
time,
and
so
the
tempo
team
is
a
little
slow
to
join
the
bandwagon,
but
we're
getting
there.
So,
for
the
last
few
weeks
I
was
taking
a
look
at
a
little
bit
of
a
refactor
to
try
to
align
with
some
of
the
approaches
in
the
helm
chart.
D
This
is
mostly
you
know
around
code
reusability
and
a
little
bit
of
patterns
that
have
been
developed
by
the
other
teams.
So
seeing
that
I
raised
an
issue
with
the
community.
So
if
you,
if
you
have
comments,
feel
free
to
look
at
the
issues
and
then
I've
got
a
big
gnarly
pr
that
does
the
refactor
and
so
that's
currently
working
locally.
All
testing
is
going
very
well
there's
a
couple
changes.
D
I
need
to
go
back
to
and
address
some
feedback
in
the
pr,
so
that'll
probably
be
coming
in
the
next
couple
weeks
and
also
we
included
jet
install.
So
this
will
be
a
way
that
enterprise
customers
can
take
advantage
of
the
installation
and
all
of
the
foundation
of
the
helm
chart
that
already
exists
today.
So
that's
pretty
cool.
If
you
have
feedback,
like,
I
said,
feel
free
to
use
it,
you
know
drop
it
drop
it
on
the
either
the
issue
that
I
filed
or
on
the
on
the
pr
itself.
D
A
D
Yeah,
I
agree,
I
think
I
think
we
should
balance
that
and
I'm
kind
of
relying
on
feedback
from
the
community
about
about
that
approach.
I
you
know
I'm
definitely
building
on
the
work
of
the
community.
This
is
not
a.
This
is
not
a
rewrite
in
any
way,
obviously
like
during
a
refactor,
certain
portions
of
the
code
are
going
to
get
smushed,
but
but
yeah.
This
is
definitely
built
on
the
work
that
the
community
has
done
over
the
years.
D
Want
to
you
know,
encourage
contributions
in
the
future
and
that
kind
of
thing
as
well,
but
as
the
company
as
grafana
gets
more
involved
with
the
helm
charts,
it's
gonna
make
sense
for
us
to
align
in
some
ways
and,
and
so
you
know,
individual
changes
between
the
charts.
You.
C
D
Might
be
looked
at
a
little
bit
differently
in
the
future,
but
yeah
definitely
definitely
a
concern,
and
we
do
have
some
breaking
changes
in
this
in
this
change
coming,
so
those
are
documented
be
on
the
lookout
if
you're
just
a
you
know,
bump
the
version
and
run
it
kind
of
person.
You
might
want
to
review
the
change
log
for
this
next
upgrade.
D
And
then
yeah
there's
no
questions
about
that.
I'll.
Just
give
a
brief
mention
about
the
operator.
So
speaking
of
community,
this
is
a
community
driven
project.
I
think
it's
building
on
the
work
that
the
loki
team
has
already
accepted
into
the
into
the
project,
and
so
this
is
sort
of
modeled
after
that,
but
we've
got
an
operator
there's
a
tempo
operator
channel
in
the
community
slack.
D
I
would
encourage
you
to
join
and
give
feedback
if
you
have
it,
there's
a
design,
doc
and
a
review,
and
all
of
that,
so
all
that's
starting
to
take
shape
and,
like
I
said,
it's
kind
of
modeled
after
what
happened
in
in
the
loki
project.
So
we're
probably
going
to
take
a
lot
of
nuggets
from
that
that
adventure.
A
So,
like
zach
said,
we
don't
use
helm
and
we
don't
use
operators
either
at
grafana,
and
we
really
rely
on
the
community
a
lot
to
make
these
things
a
good
tool
like
a
worthwhile
tool
to
use
so
appreciate.
All
of
the
input
on
the
helm
chart.
Please
jump
in
that
pr
that
zach
has
and
give
some
feedback,
and
then,
if
an
operator
is
something
that
interests
you,
we
can
all
make
this
better.
The
best
way
to
make
this
better
is
by
chatting
in
that
channel,
giving
us
some
feedback
making
use
of
it.
A
These
kind
of
things
we
do
really
deeply
rely
on
feedback
from
the
community
to
make
these
particular
components
work
again,
because
we
are
not
consumers
of
them
ourselves,
but
yeah.
We
are
very
happy
to
see
some
interest
here.
Operators
are
a
cool
feature
of
kubernetes
and
hopefully
that
project
will
move
forward
and
we'll
see
something
see
something
cool
in
the
next
few
months.
A
All
right.
On
that
note,
I
want
to
introduce
siraj
who
showed
up
on
our
team
and
solved
a
problem.
We'd
had
and
ignored
for
about
a
year
and
a
sure,
and
had
only
two
or
three
issues
in
the
reap
of
on
so
he's,
going
to
give
us
some
some
tech
details
on
dns
lookup
and
how
he
fixed
an
azure
error,
like
literally
we've
been
ignoring
for
six
months
and
occasionally
going
back
and
trying
to
fix
and
band-aiding,
and
he
he
got
into
the
nitty-gritty.
Did
he
create
nitty
gritty
details
and
fixed
it.
E
Hello,
hey
I'm
suraj.
So
if
you
are
running
tempo
on
kubernetes
with
azure
background,
you
probably
would
have
seen
these
errors.
We
saw
these
errors
in
rh
of
clusters.
E
We
ended
up
going
deeper
and
then
we
found
out
that
this
happens
when
we
try
to
do
dns
lookup
and
in
kubernetes
dns
lookups
are
amplified.
I
attached
a
github
issue
in
the
agenda
dope
that
shares
details,
but
mostly
what
happens
was
there
is
something
called
and
dots
and
kubernetes
has
and
dots
configured
to
5
by
default.
E
So
if
you
don't
have
fully
qualified
domain
names,
kubernetes
will
try
to
do
local,
dns
lookup
and
that
results
in,
I
think
around
11
local
lookup
and
then
it
finally
is
resolved
and
that
used
to
take
fair
bit
of
time
and
we
used
to
see
the
timeouts
in
like
when
right
trying
to
to
call
as
your
blog
stories
from
competitors.
E
We
did
two
fixes
in
this
one.
Was
we
increased
like
max
connections
so
that
we
don't?
You
know
like
create
new
connections
every
time
that
made
the
problem
better,
but
it
didn't
solve
it
completely.
Another
fix
we
did
was
we
wanted
to
make.
The
azure
backend
store,
is
url
fully
qualified
by
adding
a
dot
at
the
end
of
tld,
but
turns
out
that
needs
bit
more
work.
E
I
have
an
issue
attached,
so
you
can
look
at
that
issue
and
see
why
that's
the
case,
mostly
azure
is
not
happy
when
you
pass
like
post
name
header
with
the
trailing
not
so
as
a
workaround.
We
lowered
our
n
dots
to
three
in
in
azure
clusters
and
the
max
connection
and
n
dot
change
solved
that
issue
for
us,
but
yeah.
Another
pr
that
came
out
of
this
was
all
the
tempo
services
like
if
you're
running
tempo
in
distributed
mode.
E
You
have
like
bunch
of
tempo,
related
services
right
and
those
services
also
do
the
dns
lookup
stuff,
so
we
made
them
fully
qualified
by
default.
So
if
you
are
using
tempo
on
kubernetes,
you
probably
want
to
do
that.
That
will
improve
latencies.
You
know
because
you
get
to
skip
all
the
local
dns
stuff.
E
I
have
a
follow-up
issue
on
this
one.
We
we
do
want
to
support.
You
know
like
doing
the
fully
qualified
domain
name
lookups
in
azure
and
other
backends,
also
because
from
what
we
know,
this
should
save
us.
You
know
like
dns,
lookup
time
when
we
are
doing
reads
and
write,
and
theoretically
we
do
a
lot
of
reads
and
writes
right
practically
also.
What
you
know
depends
on
the
the
cluster
load,
so
this
should
they
should
help
us
on
that
front.
E
Yeah
anyone
who's
running
kubernetes,
and
this
is
not
just
tempo
if
you're
running
any
other
applications
that
make
you
know,
api
calls
for
outside
services.
You
may
want
to.
You
know,
make
usefully
qualified
domain
names
for
them
as
well.
A
A
It
just
would
fail
some
percentage
of
the
time
and
this
fix,
I
guess
just
massively
reduced
the
amount
of
requests
we
were
making
to
dns
so
always
a
win.
I
suppose
cool
thanks
to
siraj
for
fixing
that
it's
been
something
that's
been
bugging
us
for
a
long
time.
It
was
difficult
to
prioritize
and
we
appreciate
him
jumping
in
and
helping
us
on
that
note
I'll
open
up
the
floor
a
bit.
We
have
quite
a
few
people
here
from
the
community
if
you
have
any
questions
about
upcoming
work
or
even
just
how
to
operate.
A
What
you're
working
with
now,
if
you're
having
issues
running
tempo
or
just
want
to
talk
about
anything,
that's
going
on
in
your
tracing
world,
please
feel
free
to
unmute
and
ask,
or
you
can
just
chuck
it
in
the
old
chat
or
you
can
even
type
it
in
the
agenda.
I
guess
maybe
people
have
permission
to
use
the
type
in
the
agenda
doc,
I'm
not
even
sure
it's
possible.
I
have
permission
to
type
in
the
agenda,
doc.
C
A
All
right,
maybe
I'll,
ask
marty
a
question
marty.
What
are
you
gonna
do
this
fall.
A
A
We
submitted
a
paper
for
tempo
and
it
got
what
was
it
marty?
What
was
our
it
was
not.
It
was
wait
listed
right,
yeah
yeah.
We
submitted
one
about
the
new
columner,
back-end
storage
and
some
of
the
other
storage
types,
we'd
attempted
flat
buffers
and
proto,
and
all
this
and
it
was
wait
listed
and
we
haven't
heard
anything
back.
So
that's
the
way
it
goes.
I
guess.
A
Oh
heads
will
be
there
cool
I'll,
see
you
heads
all
right.
Well,
if
there
are
no
questions
or
thoughts
from
anyone
else,
we
can
go
ahead
and
wrap
this
up.
I
want
to
thank
everyone
for
showing
up.
This
is
a
pretty
big
call.
A
Just
keep
coming
to
these
calls
we'll
keep
keeping
you
updated
with
parquet
and
traceql
and
keep
you
in
line
with
what
we're
working
on
and
how
far
along
we
are,
and
I
expect
every
month
there
to
be
new
news
and
we'll
hopefully
deliver
this
soon
question
mark
soon,
tm
soon
in
the
next
few
months,
we'll
see
we're
working
hard
over
here.
I
appreciate
the
support
from
the
community
and
everything
you
all
are
doing.
Keep
following
issues.
A
Keep
following
prs,
keep
doing
what
you're
doing.
I
think
we're
building
something
awesome
here.
Everyone
take
care
and
I'll
see
you
next
month.