►
From YouTube: Webinar - What’s New in Prometheus 2.0
Description
Monitored environments have become increasingly more dynamic. With projects like Kubernetes and Mesos rapidly changing how software is being deployed and managed, Prometheus has grown alongside with its simple yet extremely powerful monitoring philosophy. The Prometheus 2.0 release includes many exciting performance enhancements and refinements underway.
In this webinar, members of the Prometheus 2.0 release team covers some of the major features in this release.
We will give a broad overview of all the features in the release and demo some of the highlights in the areas of storage, staleness handling, snapshot backups, and other improvements.
Join us for KubeCon + CloudNativeCon in Barcelona May 20 - 23, Shanghai June 24 - 26, and San Diego November 18 - 21! Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy and all of the other CNCF-hosted projects.
A
All
right,
the
number
of
participants
seems
to
be
stabilizing
Samina
get
started:
hello,
I'm,
Mark,
Coleman,
I'm,
the
client
computing
foundation,
marketing
chair
today.
We've
got
a
Prometheus
to
release
webinar
and
we're
joined
by
brian
browser
from
robust
perception
and
Ben
Kochi
from
gitlab
and
as
a
CNC
F
webinar.
First,
we
also
have
thingy
the
cat,
apparently
slightly
off
camera
there.
There
we
go
hello,
thingy,
so
Before
we
jump
into
it
a
couple
of
things
as
somebody
we
all
know
who
joined
us
earlier.
A
We
used
to
do
all
the
questions
at
the
end,
but
these
days
we
encourage
you
to
ask
questions
throughout
the
webinar.
You
can't
actually
talk
so
just
drop
them
into
either
the
chat
window
or
the
Q&A
box
and
we'll
find
opportune
moments
to
interrupt
obsesses
and
ask
them
those
now.
That's
all
that's
left,
for
me
to
say,
is
Brian
and
Ben
the
thingy.
Are
you
ready.
B
C
B
So,
thank
you
all
for
coming
so
as
stated
I'm,
Brian,
Bravo
and
we've
got
Ben
Kochi
there,
and
we
talked
a
bit
about
Prometheus
2-0,
which
was
released
just
over
three
weeks
ago.
So
when
this
webinar
was
Lukas
and
give
some
background
about,
prometheus
is
and
then
as
well
looking
at
where
it
is
to
zero.
What's
changed,
what's
improved
and
so
on
so
to
give
some
general
background
and
philetus
is
a
time
series
based
monitoring
system,
and
that
means
it
doesn't
deal
with
blogs.
It
doesn't
deal
with
traces,
it
doesn't
no
profiling.
B
Any
query
you
can
really
take
about
you
can
do
it
prompt
UL,
because
it's
pretty
powerful.
So
if
you
can
subtract
things
defined
things
whenever
you
want-
and
this
is
pretty
handy
because
as
well
with
problem
ql-
it
isn't
just
for
graphing,
for
which
we
would
normally
recommend
graph
ana
Roddick
menteng
inbuilt
in
3
dias
itself.
Anything
you
can
graph.
You
can
also
alert
on.
So
there
isn't
a
big
division
for
graphing.
You
know
nursing,
it's
all
the
same
thing:
impurities.
B
If
for
things
that
are
most
important
for
you,
both
technically
and
for
a
business
standpoint,
if
you
want
to
track
the
number
of
orders
that
customers
have
put
into
your
system
for
a
second,
that's
something
you
can
do
with
your
medias
and
that's
normally
only
two
or
three
lines
of
code.
So
it's
pretty
simple
for
more
infrastructure
stuff,
where
you
cannot
change
the
code.
B
There
are
over
two
hundred,
probably
over
300
by
now,
integrations
to
expose
metrics
from
basically
everything
there
is
so
we've
got
things
like
C
advisor
for
all
your
container
stuff
by
sequel,
post
res
Oracle,
for
all
your
database
stuff,
an
MPF
we've
got
never
devices
Linux
and
kernel
stats
for
all
the
UNIX
systems.
There's
also
WI
exported
for
Windows
Cassandra
and
all
your
databases,
and
some
people
have
also
hooked
into
things
like
minecraft
and
factorial,
because
they
really
care
about
how
many
FPS
targeting
because
they
use
way
too
many
mods.
B
But
while
shrink
can
go
beyond
just
the
normal
thing,
because
radius
is
very
flexible.
We
want
monitoring
because
Furius
is
as
light
as
an
open
platform,
so
furry
teases
tend
to
be
to
corridor
monitoring
system
that
you
can
then
hook
into
as
you
watch.
So
you
can
write
your
own
exporter
if
one
of
the
200
existing
ones
doesn't
do
what
you
want,
if
you're
not
using
a
system
like
kubernetes
or
AWS
or
GCE
or
a
resort,
and
you
have
some
other
way
of
scheduling
your
chains,
you
can
have
your
own
service
discovery
integrations
video
Jason.
B
B
Another
key
value
of
fermius
to
talk
about
is
reliability,
because
this
is
often
something
that
comes
up
and
what
per
medias
is
designed,
which
reliability
is
one
of
the
most
important
things
to
have,
because
at
the
end
of
the
day
it
is
you're,
alerting
system.
That's
going
to
be
waking
you
up
when
there's
a
real
problem.
So
for
that
reason
we
avoid
clustering.
B
Instead,
every
single
server
is
standalone
and
it
only
depends
on
local
disk,
the
networks
of
in
action
mode
ratings,
and
this
is
all
designed
so
that
you
know
you
can
set
things
up
without
a
single
point
of
failure.
So
if
you
want
redundancy
in
nursing,
you
just
run
two
identical
media
servers
which
will
both
go
to
the
earth
manager
your
manager
uses
as
ap
clustering.
B
B
So
historically,
Prometheus
was
created
which
in
South
Kent,
because
there
was
scanning
challenges
with
something
existing
watering
systems,
is
using
things
like
stats,
T,
and
that
was
a
moderately
than
I
Mac
environment.
But
in
last
three,
four
years
college,
where
she
increase
in
docker
containers
and
things
like
kubernetes.
The
profile
of
how
users
are
using
moorings
and
managing
your
systems
is
change.
B
They've
gone
from
a
situation
where
you
might
have
a
hundred
or
thousand
ten
thousand
VMs,
but
just
static
and
live
for
a
long
time
for
a
long
time
is
weeks,
at
least
to
where
you
might
be
deploying
new
pots
on
kubernetes
several
times
an
hour,
and
this
means
you're,
creating
and
destroying
a
time
series
are
lost
because
every
single
one
of
those
has
a
pot
ID
Amish
and
that
causes
performance
issues
in
the
previous
Regis
design.
Because
in
previous
one
we
were
using
a
four
time
series
and
that
causes
problems
when
you
have
lost
insured.
B
So
the
big
change
in
previous
two
is
a
new
storage
engine
which
is
indeed
our
turd
storage
engine,
and
it
is
a
brand-new
design.
It's
based
on
time,
based
blocks
and
normally
about
for
about
two
hours
and
there's
been
a
lot
of
optimization
done,
so
it
all
will
see
typical
numbers
from
users
that
RAM
is
cut
by
3
to
5
X
CPU
usage.
B
Similarly
and
I
ops
are
down
by
like
a
factor
of
a
hundred
commonly
so
well,
basically,
nothing
which
is
great
and
as
well
because
it's
using
in
virtually
index
and
these
posting
lists
within
each
two-hour
time
block
it
means
it
can
handle
this
chart.
That's
happening,
we're
seeing
in
all
these
cloud-based
environments
with
a
lot
of
changing
a
time
series.
Yes,
so
that's
great
for
skinny
guys
over
time
as
well.
B
It's
using
page
cast
restoring
previous
one
that
was
doing
his
own
memory
management
and
talk
of
what
the
kernel
was
doing
in
prometheus
to
were
just
relying
on
the
kernel
because
well
Linux
has
been
doing
this
for
well
over
a
decade
now,
if
knows
what
it's
doing,
just
no
need
for
us.
A
second
and
another
change
to
consider
is
that
some
users
like
to
take
backups
of
breaches
in
prometheus
ward.
This
wasn't
really
practical
because
it
involved
shutting
down
Prometheus,
which
might
take
a
few
minutes
and
now
over
ets.
B
B
It
read
here
is
one
if
you
had
a
time
series
and
it
went
away
such
as
the
pod,
the
kubernetes
pod
went
away.
Previous
would
continue
to
return
that
for
five
minutes
afterwards,
which
causes
a
bit
of
problems
if
you're
tried,
for
example,
inertia
on
that
going
away
because
it's
going
to
add
a
5
minutes
later,
and
it's
also
caused
double-counting.
B
If,
for
example,
another
pod
comes
in
to
replace
it
and
their
boat
in
there
inside
that
five
minute
window
so
in
prometheus,
this
is
no
longer
an
issue,
because
Prometheus
will
now
notice
that
the
targets
by
the
way
Rockets
sail
along
a
return.
So
this
is
something
pretty
nice
that
most
users
will
just
get
for
free.
B
Another
change
is
rules.
We
now
have
row
groups
previously
in
previous
one
all
the
rules
you
have
so
recording
rules
allow
you
to
execute
a
query
and
save
the
result
inside
previous,
so
one
called
standing
queries
or
whatever,
and
this
is
commonly
used
for
aggregation
and
historically
how
it
worked.
Is
that
all
the
rules
you
have
just
one
big
block
and
they're
all
run
simultaneously?
So
there
are
race
conditions.
B
There
is
where
you're
getting
the
information
from
a
previous
run,
with
the
current
run,
or
maybe
a
mix
of
those
two
so
of
radius
tuning
down
the
notion
of
row
groups
at
again
of
multiple
groups.
Each
of
them
has
a
list
of
rules
which
are
just
run
from
top
to
bottom.
In
order,
so
you
can
have
no
dependencies
between
your
rules
in
the
same
fashion,
and
if
you
have
some
rules
there
doing
an
aggregation
hierarchy,
that's
great
because
around
can
depend
on
the
property
and
save
yourselves
in
CPU.
B
C
So
it'll
be
one
more
thing
on
rules.
That's
really
hammy
is
the
rules
can
now
have
individual
for
the
rural
groups
can
have
individual
execution
cycles
from
intervals
so
that
different
rule
groups
can
be
done
every
five
seconds
or
every
minute
or
every
five
minutes,
and
this
allows
for
some
some
interesting
techniques
to
do
down
to
do
internal
down
sampling.
Of
course,
the
Prometheus
server
won't
store
that
these
for
a
long
period
of
time,
but
you
can
use
Federation
or
remote
streaming
to
to
send
these
down
sample
metrics
to
an
external
server.
B
Yep
yep,
so
I,
then,
just
to
summarize
so
the
maintains
appreciates
2-0
is
new
storage,
which
is
much
more
efficient
and
scalable
the
improved
stainless
semantics
and
the
improved
rules.
There
are
smaller
additional
small
changes,
like
some
modifications
to
prompt
QL,
but
the
big
changes
are
there,
but
the
important
message
to
get
from
this
is
for
Regius
is
more
ready
than
ever
to
be
able
to
deal
with
kubernetes
and
others
out
major
environments
as
they
continue
to
be
rolled
out.
A
C
Yeah,
so
there's
there's
a
few
different
options
for
this.
The
biggest
one
we
recommend
is
is
starting
up
because,
because
of
the
polling
method
of
Prometheus,
you
can
have
multiple
Prometheus
servers
running
simultaneously
and
the
are
recommended
our
recommended
upgrade
technique
is
to
set
up
a
second
Prometheus
2.0
server
in
parallel
with
your
existing
server.
C
This
makes
it
really
really
easy
to
test
and
verify
that
everything
you
expect
to
be
working
with
prometheus
2.0
with
your
existing
one.
One
point
X
server
is:
is
working
as
expected:
you're
getting
alerts
sent
you're
getting
the
graphs
generated.
All
this
data
is
being
collected
and
you
have
no
performance
problems
and
then
once
you've
got
some
testing
on
that,
then
you
can
simply
just
throw
away
the
old
community
is
1.0
server
and
and
continue
on
with
2.0
there's.
C
Also
a
couple
of
different
people
in
the
parenthese
community
who
have
developed
or
are
in
the
process
of
developing
database
conversion
tools,
they're
kind
of
a
work
in
progress.
Of
course,
it
takes
quite
a
long
time
to
to
convert
a
database,
so
it
may
not.
It
may
take
longer
to
convert
the
database
than
to
simply
set
up
a
new
one
and
wait,
but
if
you
have
a
very
small
database-
and
you
want
to
just
do
it
to
do
it
an
offline
upgrade,
there
are
a
couple
of
tools
being
developed
for
that.
A
Thanks
Ben
I
think
relevant
question
would
be
since
we've
got
cloud
native
comb
coming
up
next
week.
There's
probably
gonna
be
a
hope.
I
mean
I
haven't
checked,
but
it's
probably
gonna
be
a
bunch
of
Prometheus
content
there
that
people
should
look
out
for.
Are
you
aware
of
any
any
tips
the
people
who
are
flying
over?
It's
awesome
so.
B
A
C
C
C
A
B
So
this
question
they're
asking
legacy
application
vulturing
on
five
six
hundred
machines
or
in
a
challenge
changed
codebase
and
it
sounds
like
it's
a
java
application.
So
in
that
sort
of
situation
there
is
the
JMX
exporter
which
I've
maintained
and
which
can
get
all
your
standard,
Java
stats.
So
all
your
memory
pools
your
tread
pools
how
many
classes
you
have
loaded
the
Java
version
initially
to
the
process.
B
B
So
the
question
is
about
histograms
and
how
to
graph
the
McGriff
error.
So
I
looked
familiar
with
these
particular
ones
here,
but
I
can
talk
about
free
cheese,
histograms
generally,
and
so
the
general
answer
is
that
you
get
the
history,
you
aggregated
upkeeping
d
le
label
and
then
you
call
histogram
quantile,
guess
you're
at
like
percentile,
your
75th
percentile
or
whatever
else
or,
and
then
you
graph
that
directly
in
your
graph,
and
it
just
lays,
would
everything
else
that
we
found
I
did
also
a
release
or
two
back
a
d8
a
heatmap.
B
A
All
right,
thanks
Brian,
it
seems
that
those
were
the
only
questions
that
we've
got
right
now.
I'm
just
gonna
check
the
chat,
no
there's
nothing
else
in
there.
So
there's
all
all.
That's
left
for
me
to
say
thank
you
very
much
to
Ben,
Brian
and,
of
course,
to
thingy
the
cat.
That
was
a
short
but
sweet
presentation
and
I'm
sure.
If
you
need
more
information,
you
can
just
go
to
the
URL
that
you
saw
with
the
blog
down
and
follow
it
up
to
the
top
to
find
everything
else.