►
From YouTube: Layer5 Community Meeting (Dec 4th, 2020)
Description
Nisarg Joshi introduces @envoyproxy Nighthawk's adaptive load controller.
B
A
A
A
B
A
A
A
A
A
We
were
having
a
little
bit
of
a
trouble
with
the
virus
here
in
here
where
I
am
people
enjoyed
thanksgiving
and
then
now
the
now,
I
think
my
my
kids
will
probably
stay
do
school
virtually
for
the
rest
of
the
year,
just
because
it's
it's
going
round,
so
I
hope
that
you're
all
well.
I
hope
that
none
of
you
have
know
people
that
are
affected.
A
Okay
on
to
some
better
notes,
I'm
good
to
see
all
of
you
here.
I
recognize
each
of
the
names
of
all
of
you,
including
mr
alvarez,
neil,
but
neil.
This
is
your
first
time
on
the
community.
Call,
I
think,
neil,
do
you
mind
introducing
yourself
saying
hi
to
everybody
just
talking
about
where
you're
from
what
you're
interested
in
letting
people
get
to
know
you.
A
A
A
A
A
But
speaking
of
workshops,
we're
gonna
we'll
have
a
couple
more
workshops
on
istio
coming
up
this
week.
These
are
both
being
held
through
o'reilly's
platform.
So
I
think
it's
been
a
long
very
well
understood
and
kind
of
a
long-standing
long-standing
understanding
by
each
of
you
that,
as
you
want
to
get
more
into
service
mesh
technology
that
there's
there's
workshop
material
for
you
to
take
and
self-study.
A
A
Guess
what
it's
no
surprise
that
you're
internet
famous
or
your
work
is
being
shown.
So
you
know,
I
honestly,
I
didn't
read
this
article
but
of
a
couple
of
the
talks
that
we
gave.
People
are
paying
attention
they're.
Looking
at
the
landscape,
noting
the
differences
between
service
meshes
and
trying
to
understand
these
things.
A
They
are
also
talking
about
measuring
and
service
mesh
performance.
So
so
they're
talking
about
our
abstractions
smp
I've
been
talking
about
smp.
With
a
few
of
you,
a
few
of
you
who
are
working
on
mesh
map.
Meshmap
is
going
to
represent
visually
service
meshes
and
as
it
does,
it
needs
to
track
all
kinds
of
information
metadata
about
the
service
meshes
as
we
catalog
that
metadata
it
turns
out,
like
that's
in
large
part.
What
smp
is
so
briefly
for
those
that
aren't
as
familiar
with
this
this
project
s
p.
A
So
when
you,
when
you
go
out
and
look
at
the
project,
here's
the
project
site
the
code-
that's
in
it
not
very
much
most
of
the
code
is
just
the
project
site
itself,
but
the
actual
project.
The
actual
specification
is
here
under
protos
and
right
now,
there's
three
of
them.
I
think
the
largest
one
being
the
one
called,
no
I'm
sorry,
the
largest
one
being
called
service
mesh
performance,
proto
and
there's
a
bunch
of
things
that
we
want
to
track
about
a
service
mesh.
A
Let
me
let
me
go
back
to
this
more
simple
one
here.
This
very
short
one
just
to
convey
the
notion
to
you
that
very
simply,
everyone
here
knows
that
there's
more
than
one
service
mesh
out
there
to
use
and
this
file
this
protobuf.
This
proto
is
a
a
way
of
categorically
enumerating
the
name
of
or
identifying
a
service
mesh
saying,
hey
as
measury,
and
other
tools
want
to
identify
that
one
cert
that
a
service
mesh
is
console
it
it.
A
A
That's
what
this
news
is
about.
So
this
is
actually
the
first
news
article
that
I've
seen
that
where
someone
has
written
up
about
smp-
and
I
know
if
pratya
or
jabir
were
on
the
call
right
now-
they
would
be
just
really
perturbed
about
this.
They
would
be
perturbed
that
they
killed
off
the
s
in
s
p.
I
think
in
a
good
way.
A
But
I
just
wanted
to
take
a
moment
to
congratulate
you
all,
because
I
didn't.
I
didn't
tell
this
news
source
to
come.
Write
this
up
or
to
come.
Do
anything
they're,
finding
it
very
interesting
and
they're
writing
it
up.
So
so
kudos
to
all
of
you
is
pretty
cool.
As
a
matter
of
fact,
I
was
just
on
linkedin
a
moment
ago
and
heb,
which
is
a
national
grocer.
A
A
grocer
grocery
store
chain
in
the
u.s
and
the
biggest
one
in
texas,
which
is
where
I
am
and
which
is
texas,
is
a
massive.
It's
a
large
state
anyway.
Heb
is
a
big
old,
a
big
old
and
they
are
pointing
out
the
news
article
and
really
when
you
look
at
the
article
it
starts
with
us
and
it
ends
with
us,
which
means
it
probably
says
something
about
how
important
they
think
that
the
work
is
so
pretty
cool.
A
E
C
C
That
would
be
seen
on
this
individual
member
pages
so
like
it's
basically
two
to
three
things:
location,
skills
and
something
like
80
to
100
words
of
a
small
description
stating
from
where
you
are
or
what
like
exa.
What
work
have
you
done
with
layer
file
like
you
have
worked
with
mastery
or
you
have
worked
with
smp
or
you
worked
with
websites
or
something
and
yeah.
That's
like
that
is
very
quiet
thing
that
we
are
looking
for.
C
C
So
in
the
new
website,
we
wish
to
have
additional
info
for
all
the
members
and,
if
you,
if
you
think
there
can
be
some
other
addition
or
some
field
that
can
be
incorporated
for
you
but
hasn't
been
marked
till
now,
you
can
just
ping
us
and
we'll
try
to
create
something
like
that.
I
think
like
in
the
mail
itself
will
provide
the
link
to
the
vinayaks
the
page,
so
you
can
just
look
into
it
and
or
I
will
just
create
an
example
template
as
well,
if
required,.
C
A
A
This
of
an
ix
profile
is
a
good
example
because,
because
he
actually
has
a
short
bio
and
his
location
and
skills
actually
josh,
it
strikes
me
that
we
need
to
communicate.
Vinayak
is
the
he
has
a
bio
because
he's
the
one
that's
been
enhancing
the
page
to
look
a
little
bit
better
and
have
have
this
info.
I
don't.
I
think
that
there's
one
more
iteration
of
what
we
were
doing
here.
I
think
that
vinayak
was
going
to
publish.
C
A
A
Much
less
is
that
actually
done
in
a
standard
way
of
the
standard
way
that
they
have
a
little
anyway.
It
wouldn't
work
or
work
it
wouldn't,
because
even
at
that,
then
some
people
might
say
yeah.
But
I
want
my
thing
to
say
something
different
here
than
it
does
over
there
and
like
what
would
be
better
is
to
try
to
grab
that
from
linkedin.
But
then,
even
at
that,
like
nope
people
can
just
put
in
a
little
bit
of
text
and
put
in
the
other
stuff
once
so.
A
F
C
A
Yeah
good
good,
maybe
maybe
a
good
topic
for
mondays
mondays
websites
call
or
for
but
but
kishore.
I
thought
it
would
be
nice
for
everyone
to
get
a
little
bit
of
a
look
at
the
the
new
site,
how
it's
shaping
up.
And
since
we
were
kind
of
talking
about
since
you're,
actively
working
on
that
section
and
we're
kind
of
talking
about
a
couple
of
upcoming
workshops
that
good
for
people
to
see
how
that's
changing.
E
E
D
E
A
C
A
A
I
have
nothing
else
to
add
other
than
that
that
makes
sense
makes
makes
a
lot
of
sense.
There's
you
know
a
few
people
have
been
asking
for
like
embroidered
shirts,
and
we
actually
have
a
stack
of
shirts
here
to
send
out
to
folks
I
gotta
say
like
it's
incredibly
expensive
to
ship
things
internationally,
and,
and
so
we
have
only
sent
so
many
but
participating
with
a
store
that
will
print
them
might
be
the
right
thing
to
do
so.
A
D
Yeah,
so
I'm
not
quite
sure
that
how
like
should
I
just
talk
about
it
or
because
I
don't
have
a
website
that
I
can
demonstrate
you
don't
have
a
cool
stuff
store
for
the
istio
adapter,
what's
going
on
yeah,
so
should
I
go
through
the
pr,
because
I
think
that
might
just
bore
a
lot
of
people.
Should
I
just
talk
about
that?
What
the
pr
ants.
A
D
D
What
it
allows
you
to
do
is
you
can
use
it
to
basically
expose
any
kind
of
services
and
deployments
in
any
of
the
namespaces
right
now
we
are
using
in
linkery
adapter
to
expose
the
prometheus
instance
that
lingerie,
like
linkedin
service
missions,
deploys
on
its
own,
but
that's
something
that
we
are
using
it
for
now,
but
as
in
general,
this
particular
function
can
be
used
for
any
to
basic.
This
is
exactly
energy
to
what
cube
ct
expose
does,
and
it's
now
present
in
the
linkedin
adapter.
A
A
A
How
do
you
access
those
services?
You?
Well
you
oops!
Excuse
me:
you
need
those
exposed
outside
of
the
mesh,
so
you
have
like
a
url
and
a
port
number
and
there's
something
going
on
with
my
system
and
and
you
and
you
want
to
you-
want
to
get
at
them.
Well,
that
exposure
happens
in
exposing
the
service
happens
in
different
ways,
depending
upon
what
service
mesh
you're
running
and
so
as
you're
doing
this
for
linker
d.
A
Let
me
ask
you
some
questions
around
that,
because
I'm
I'm
really
excited
about
this,
that
this
is.
This
is
great,
so
everyone,
like
you
know
I
I
just
started
a
fresh
system.
I
deployed
linker
d,
part
of
what
came
inside
of
linker
d
in
the
control
plane
was
a
small
copy
of
grafana,
a
small
copy
of
prometheus
good.
If
I
and
meshrey
will
connect
to
those
two
things
and
retrieve
metrics.
A
A
This
is
a
bad
example,
because
the
way
that
linker
d
will
deploy
grafana
is
a
bit
funky
or
a
bit
specific
to
linker
d
in
the
way
that
it
configures
grafana
it
like
it,
doesn't
necessarily
allow
external
systems
other
than
linker
d
to
get
to
grafana,
but
the
way
that
it
deploys
prometheus
isn't
the
same.
An
external
system
like
mesher
is
able
to
get
to
it.
So.
A
D
A
Refresh
my
memory
cube
ctl
expose
is
it
is
that
is
not
creating.
That
is
not
doing
port
forwarding,
that
is
configuring
or
updating
a
service
that
represents
that
end
point.
It
does
it
require
having
an
existing
kubernetes
service
and
it's
updating
the
config
or
what
what
is
cube
ctl
exposed
doing
under
the
coverage.
D
They
basically
create
another
service
just
to
expose
it.
You
can
actually,
as
far
as
remember,
you
can't
specify
the
name
similarly
like
in,
like
the
function
that
I've
added.
What
we
do
is
we
create,
and
the
service
which
is
so
responsibility
is
to
allow
us
to
communicate
it,
allow
us
to
communicate
with
the
prometheus
or
whatever
the
deployment
or
service
we
want
to
communicate
with.
D
Similarly,
in
the
period
after
the
function
that
I've
added,
that
also
allows
us
to
specify
that,
if
you
want
to
expose
it,
we
want
to
get
an
end
point
rest
id
based
or
not
code
based
or
load
balancer
base,
but
by
default
and
the
function
that
linkery
adapter
that
that
will
be
getting
is
to
add
a
notepod
based
endpoint,
so
that
every
node
can
have
access
to
it.
But
obviously
like
load
balancer
cost
money,
so
we
don't
go
for
it
by
default.
D
A
This
functionality
that
you're
putting
into
the
linker
d
adapter
is,
is
actually
desirable
in
mesherie's
server
as
well,
because
because
that's
just
it's
a
generic
capability
that
we
would
want
to
be
able
to,
you
know,
manipulate
kubernetes
configuration
to
expose
any
service.
So
this
is.
This
is
really
good.
I'm
really!
It's
really
neat
to
see
that
that's
being
incorporated.
A
A
A
A
So
clearly,
there's
some
sort
of
there's
a
couple:
there's
a
couple
of
challenges
around
this.
If
you
have
multiple
prometheuses
promethei
by
the
way,
there's
been
a
anyway
anyway.
If
you
have
multiple
prometheuses
in
the
environment,
which
one
should
you
connect
to,
or
rather
maybe
it's
time
that
mescheri
as
a
tool
has
the
ability
to
connect
to
multiple
at
a
time,
that's
not
hard
to
get
to
that.
That
code
isn't
difficult
to
write
the
the
point
of
discussion
of
doing
that,
though,
revolves
around
the
concept
of
a
context
or
of
an
environment.
A
We
really
need
to.
At
that
point,
we
really
need
to
start
to
say:
mesherie
is
capable
of
actively
having
connections
to
multiple
kubernetes
clusters,
multiple
prometheuses,
multiple
grafanas,
and
we
need
to
empower
the
user
with
sort
of
a
simplified,
simple
control
over
that
and
you
oftentimes
find
the
word
environment
or
context
or
cluster
or
some
something
so
that
people
can.
You
know
we're
gonna
have
to
we're.
Gonna
have
to
bring
some
further
controls
to
allow
to
enhance
this,
but,
but
I
do
feel
like
hey.
D
Yeah,
so
there
was
a
minority,
that's
in
ist
adapter.
Now
you
should
be
able
to
like
earlier.
We
had
some
scripts
and
those
who
should
like
they
were
used
to
work
only
in
unix-based
environments
like
mac
or
linux.
But
now
you
should
be
able
to
run
this.
This
video,
sorry
yeah,
it's
the
adapter
in
a
windows
based
environment,
because
now
it
supports
dot,
ext
based
files
and
sets
up.
Basically
it
installs
it
for
you.
A
D
A
A
C
Okay,
yeah
hello,
so
yes,
essentially,
this
component
related
to
adaptive
load
control
in
nighthawk,
okay,
so
I'll
just
give
a
brief
about
nighthawk
and
what
sort
of
we
are
planning
to
do
with
it.
So
nighthawk
is
essentially
a
performance,
characterization
tool
and
yeah.
Okay
thanks.
So
basically
we
are
planning
to
use
it
as
a
part
of
service
service,
mesh
performance,
and
this
tool
would
help
us
to
essentially
carry
out
distributed
load
tests
on
the
service
mesh
and
gather
some
matrix
and
analyze
them.
B
B
C
C
Of
how
scaling
works
in
nighthawk
and
in
this
diagram,
if
it's
visible,
so
what
they
have
done,
is
they
just
outlined
how
the
components
work?
It's
a
very
straightforward
sort
of
an
architecture
that
we
would
be
leveraging,
so
basically
they
have
outlined
a
forwarding
or
distributing
service.
First
of
all,
here
you
send
out
your
sort
of
the
test
configuration
that
you
want
to
carry
out
and
some
details
about
it
and
what
they
do
internally.
C
So
this
is
like
a
single
point
to
which
you
talk
to,
or
your
client
talks
to,
and
then
this
service
itself
generates
multiple
load
test
clients,
so
these
are
grpc
services
that
are
going
to
actually
execute
the
load
test
based
on
the
configurations
and
then
there's
a
sync.
So
this
is
one
of
the
updates,
I
think
from
the
previous
design,
where
now
they
have
a
sync,
where
you
can
actually
query
the
result
at
a
later
stage,
instead
of
waiting
for
the
results
or
something,
so
you
can
actually
get
an
id
back
from
this.
C
This
forwarding
service
that
tells
you
to
like,
using
that
you
can
later
on
query
the
results
associated
with
that
test
id
or
session
id,
so
to
say
so.
These
load
test
clients
are
going
to
carry
out
the
load
test,
and
so
yeah
here
is
sort
of
an
outline
here.
This
component,
they
have
showed,
is
like.
C
If
you
have
a
proxy
service,
then
that's
what
they're
fitted
in
there
and
after
they
have
carried
out
this
test,
they
are
going
to
aggregate
and
send
the
results
to
the
sink,
and
you,
the
sync
service
will
carry
out
some
result
aggregations
and
you
can
query
it.
So
it's
a
like
a
straightforward
architecture
that
is
outlined
in
this
document.
C
Apart
from
that,
I
think
they
have
discussed
like
some
code
segments-
and
there
are
some
comments
here
and
there
I
don't
think
we
want
to
go
into
some
details
about
the
other
thing
about
the
adaptive
load
control
component.
I
think
that
is
one
of
bigger
interest
for
us,
so
we
are
going
to
use
this
nighthawk
tool
and
there's
a
component
in
this
tool
called
adaptive
load
control.
C
Yeah,
okay,
so
about
this
component,
adaptive
load
control
again,
I
am
myself
in
some
initial
stages
of
understanding
this
completely,
but
I'll
give
an
overview
of
what
it
is
and
sort
of
how
it
works.
So
the
idea
is
in
a
load
test.
What
if
you
want
to
adaptively,
find
out
some
amount
of
load?
That
should
be
like
that
your
service
can
handle
for
some
amount
of
response,
time
or
service
level
performance
that
you
want
for
your
services.
So
that
is
what
this
component
can
help
you
do.
C
It
can
help
you
converge
to
some
value
of
load
that
will
meet
a
certain
criteria
or
constraint
that
you
want
from
your
service
in
terms
of
how
it
performs
basically
and
this
component.
So
this
file
is
like
some
header
file,
and
here
there's
a
nice
outline
actually
in
comments.
We
can
see
so
basically
a
load.
Adaptive
load
controller
has
a
set
of
helper
objects,
so
what
it
does
is
first
is
collect
some
matrix
from
the
tests
that
are
going
on.
After
that
it
uses
some
scoring
functions
to
score.
C
These
results
evaluate
these
matrix
and
then
you
have
to
use
a
step
controller
component
that
all
of
these
components-
you
can
implement
yourself
and
it
has
some
business
logic
that
you
care
about.
So
the
step
controller
actually
tries
to
converge
to
a
value
of
load
that
you
want
to
meet
according
to
the
spec.
Basically,
that
is
the
idea,
so
here
it's
written
out
that
through
helpers
it
performs
nighthawk
service
benchmarks
obtains
matrix
from
matrix
plugins.
So
you
can
write
your
custom
plugins
as
well.
They
have
written
some
plugins
already.
C
You
can
score
the
results
again.
These
scoring
functions
and
evaluation
functions
can
also
be
custom,
but
they
provide
some
baseline
functions
and
you
can
consult
a
step
controller
plug-in
again,
it's
a
plug-in,
so
you
can
write
your
own
step
controller
to
determine
next
load
and
detect
convergence,
so
all
of
these
plugins
are
specified
through
a
spec
that
you
provide
as
as
a
part
of
the
session
and
yeah.
So
this
is
like
a
header
file,
an
implementation
that
you
we
can
check.
I
just
want
to
show
a
couple
more
things.
B
C
That
the
matrix
plug-in
can
return
to
you
so
achieved
rps
attempted
rps
latency
like
different
matrix
on
latencies,
the
send
rate,
success
rate,
etc.
So
you
you
can
get
these
values
from
the
load
test
and
you
can
feed
it
to
the
next
component,
essentially
the
matrix
evaluation
component,
which
can
then
carry
out
some
kind
of
evaluations
and
feed
it
to
the
step
controller.
C
C
That
they
have
implemented
so
they
have
implemented
some
controllers
already,
which
can
be
used
to
find
a
convergence
value
for
the
load.
So
we
can
implement
our
own
custom
values
and
as
of
now
I
mean
I'm
not
sure
exactly
what
we
would
be
doing,
but
we
can
utilize
these
components
in
order
to
carry
out
performance.
A
A
A
A
A
So
the
the
weight,
the
weighting
like
on
line
33,
is
kind
of.
I
think
that
there's
a
couple
of
words
here
that
are
really
good,
like
the
use
of
weight,
the
the
the
threshold
and
scoring
those
are
things
that,
if
you've
heard
me
speak
about
mesh
mark
in
the
past
and
the
ability
to
mesh
mark
just
just
being
a
new,
a
new
yardstick
to
measure
the
performance
of
your
mesh
like
this
is
a
great
place
to
be
able
to
do
scoring
to
be
able
to
create
a
mesh
mark.
B
H
A
Because
it's
yeah,
it's
you
know
it's
exact.
It's
it's
there's
so
many
examples
of
like
questions
that
you
want
to
ask
which
is
hey.
I
want
you
know,
I'm
running
a
mesh,
I'm
running
a
set
of
services.
I
really
have
value
to
take
value
in
receiving
and
having
distributed,
tracing,
enabled
and
I'd
like
to
re
my
service
mesh
to
retrieve
as
many
I
like
to
sample
those
traces
as
much
as
possible.
So
I
can
have
a
high
fidelity
view
of
what's
happening
over
time
and
and
where
the
slowness
in
my
application
is.
A
Although
there's
a
cost
to
you
know
every
single
trace
that
you
collect,
there's
a
small
cost
and
so
where's.
What's
the
optimal
amount
of
visibility
versus
not,
and
some
of
that
actually
changes
over
time
based
on
parameters
that
the
load
generator
has
no
idea
about,
like
those
are
parameters
about,
is
it
black
friday
or
is
a
certain
user
signed
in
or
like.