►
From YouTube: Deep Dive: gRPC - Yuxuan Li, Google
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
Deep Dive: gRPC - Yuxuan Li, Google
gRPC-Go is a native implementation of gRPC spec in Go language, with rich features and flexible APIs. To fully exploit the power of gRPC-Go and be able to debug issues, one needs knowledge about how gRPC-Go is designed, the various support it provides and the flexibility of customization it offers. This presentation will dive into the architecture of gRPC-Go, mainly the surface (gRPC) layer, transport layer, resolver component and balancer component. It will also include a brief introduction to key gRPC features such as binary logging, channelz (runtime debugging), service config and rpc retry along the way.
To learn more click here: https://sched.co/FuLX
A
Okay,
let's
get
started
welcome
to
my
session
about
your
physical
architecture
and
the
features.
This
is
the
deep
dive
session,
so
I
assume
the
audience
to
have
some
level
of
similarity
with
geophysical
in
general,
my
name
is
usually
a
messenger
at
Google
and
I've
been
working
on
the
chair
physical
project
for
about
two
years
before
we
dive
into
the
main
topic
of
today,
I
would
like
to
give
you
guys
a
short
and
brief
introduction
to
our
PC,
because
you
guys
have
been
hearing
about
kubinashi
and
many
other
stuff
for
the
past
few
days.
A
So
I
need
I
want
to
do
a
contacts
with
you,
contacts
which,
with
you
guys.
So
what
is
your
PC
gerrae
PC?
Is
their
modern,
open-source,
hyper
homeless,
RPC
framework,
so
a
sports
force
multiple
languages
and
is
a
cross-platform
and
we
design
our
API
carefully
to
make
our
components
pluggable,
so
user
can
choose
the
implementation
or
even
implement
them
on
their
own
stuff?
That
best
suits
your
needs,
and
one
more
thing
to
highlight
about
your
PC's
is
rich
features.
A
We
have
core
features
like
bi-directional
streaming:
flow
control
and
many
other
add-on
features
like
binary
logging
channel
D
tracing,
so
it
can
provide
it
help
user
better,
manage
configure
or
debug
the
applications
based
on
top
of
your
PC.
This
is
a
picture
that
illustrates
our
typical
use
cases,
so
a
new
user
on
mobile
web
or
desktop
can
use
your
PC
to
communicate
with
micro
services
or
access
cloud
api's
and
between
micro
services.
They
can
also
communicate
with
your
PC,
moreover,
and
similarly,
this
micro
service
can
still
use
Java
C
to
access
those
claw
api's.
A
A
As
I
said
before,
ERP
system
for
multiple
languages
and
among
those
languages
we
have
three
implementation,
that's
natively
implemented,
there
are
PCC
Jerry,
Siegel
and
Joe
PC
Java
and
all
the
other
languages
are
a
wrapper
around
the
jar
PCC
core
implementation.
So
today,
I'm
going
to
talk
about
the
Derby
Seco
implementation,
which
is
one
of
the
three
native
implementation
and
a
little
bit
of
history
about
your
physical.
So
we
open
sourced
in
February
2015
and
we
get
a
better
version
out
in
September
2015.
A
A
A
So
later,
when
you
explore
our
code
base,
you
feel
more
confident
and
more
comfortable
I
hope
that
at
the
end
of
the
story,
you
guys
can
have
sort
of
a
jarppi
seen
map
in
your
mind.
So
you,
when
you
later
diving
on
what
dive
into
our
code
base,
you
will
never
get
lost
and
soldier.
Your
PC
is
many
about
two
things.
First
is
managing
the
life
of
network
connections
and
the
second
is
managing
the
average
cost
over
those
network
connections.
So
this
talk
will
be
focus
on
these
two
things.
A
So,
first,
let's
look
at
the
code
stack,
so
we
have
the
transport
of
transport
layer
on
the
bottom
and
we
have
the
HTTP
to
transport
by
default.
On
top
of
that
is
the
GPC
framework
layer.
It
contains
the
main
load
row
of
G
RPC
and
the
next
is
the
JP
capr
layer,
and
we
also
have
another
optional
layer
or
generated
code
of
API.
That's
ruptured
around
the
join
PC
API
user,
in
most
cases,
user
and
interact
with
post
generated
code
API
and
your
PC
API.
A
So
what
are
the
operations
involving
at
your
PC?
There
are
basically
four
of
them,
so
both
client
and
server
can
sell
had
her
to
set
up
a
physical
and
next
it
has
a
message
which
is
what
an
RPC
all
about
and
to
end
the
RPC
call
client
can
send
close
to
server
so
to
lessor
and
know
that
he
has
done
standing
all
the
message
you
want
to
send.
And
finally,
sir,
was
Sangha
trailer
to
to
to
client
which
contains
the
final
status
of
the
RPC
and
mark
the
RPC
as
complete.
A
A
Next,
it
will
clue
you
will
send
a
hop
close
to
server
to
say
that
I'm
done
with
sending
all
the
messages
and
the
server
will
send
first
and
the
header
to
claim
to
set
up
the
RPC
at
the
client
side
then
send
the
response
message
to
client
and
finally
says
a
trailer
containing
the
final
status
of
the
RPC.
To
mark
the
RPC
as
complete
engine
PC,
there
are
we
categorize
RPC
into
four
different
types
by
the
messaging
pattern.
So
the
first
one
is
the
unary
four
unary
type.
A
Basically
client
will
send
a
single
message
to
silver
and
the
service
and
a
single
message
back
to
client,
and
this
is
the
most
commonly
used
type
of
RPC
nowadays,
and
the
second
type
is
the
client
streaming
type
for
this
type.
Client
was
from
Boston,
send
a
stream
of
send
a
stream
or
message
to
a
server,
and
the
server
will
respond
with
a
single
message
and
similarly
for
service
streaming
as
kind
of
a
send
a
single
message
to
server
and
the
server
will
stream
multiple
message
back
to
client.
A
Finally,
the
Biden
streaming
is
the
shorthand
for
bi-directional
streaming.
This
is
the
so
for
this
type
of
RPC.
There's
no
restriction
on
the
message
sent
by
the
client
or
the
server
and
actually
GRP
see
implemented
production
of
streaming
natively
and
the
all
the
other
three
types
are
just
special
cases
of
the
Medeiros.
No
streaming.
The
picture
on
the
right
hand,
side
illustrate
our
at
the
concept
of
directional
streaming.
A
A
It
is
a
bi-directional
streaming
opposite
hype,
and
basically
is
the
client
just
exchanged
the
location
info
with
the
resistor
broadcast
service
continuously,
and
we
also,
let's
also
assume
that
for
the
host
name
ABC
that
XYZ,
we
have
three
backends
to
serve
a
surrounded,
our
hostname,
which
is
a
very
common
casing
nowadays
to
do
load
balancing
so
so.
The
first
step
to
making
our
physical
will
be
to
create
a
channel
to
transmit
the
RPC
code.
If
there's
no
easiest
in
one
that
you
can
use.
A
So
let's
see
how
to
create
a
channel
in
new
RPC
before
that.
Last
differential
will
very
important
concept
in
your
PC,
which
are
channel,
and
the
sub
Channel,
so
channel
is
an
abstraction
over
connection
that
an
RPC
can
be
issued
to
you
can
simply
think
of
it
as
a
a
logical
connection
between
client
and
the
host
name
target.
Next
is
the
sub
Channel.
It
is
an
abstraction
over
a
connection
and
use
for
a
lot
of
balancing.
You
can
simply
regard
it
as
a
connection
between
the
client
and
an
individual
back
end.
A
Now,
let's
look
at
the
step
to
create
a
channel,
so
the
first,
the
user
will
diode
will
call
diode
who
will
call
the
opposite
diode
to
connect
to
the
host
name
and
it
prefix
the
host
name
by
the
scheme
that
the
host
name
is
fully,
and
here
we
are
using
the
DNS
scheme
and
so
Champy's
you've
got
the
dial
function
code.
It
will
parse
the
target
stream
and
starts
the
resolver
according
to
the
stream.
Here,
since
we
are
using
a
DNS
dream,
work
will
will
only
initiate
start
a
DNS
resolver.
A
Resolver
was,
and
then
the
resolver
work,
or
to
go
to
to
go
to
resolve
the
hostname
and
return
the
addresses
from
return.
The
resolved
addresses
to
the
ER
PC
and
there
is
an
optional
service
on
Theoden.
Cows
will
be
reserved
at
the
same
time
and
the
ones
that
your
PC
receives
the
address
list.
It
will
start
a
balancer
based
on
the
user
config
or
the
service
and
fake.
Both
of
them
can
configure
to
configure
the
parents
are
to
be
used
for
this
channel,
and
so
your
PC
will
start.
A
Let's
assume
that
we,
your
PC,
decides
to
use
the
wrong
wrong
robin
balancer,
so
it
was
starts
the
wrong
robin
balancer
and
the
past
around
return,
a
list
of
address
to
the
balancer,
so
the
bouncer
will,
depending
on
its
balancing
policy,
it
will
create
connections
to
the
well.
It
will
instruct
your
PC
to
connect
to
create
connections
to
the
back
ends
a
little
bit
about
balancer
here.
A
So
in
Jeff
you
see
we
have
been
to
two
balancer
which
are
wrong:
robbie
and
a
peek
first,
so
Ronghua
being
baffled
by
rubbing
balancer
either
what
tries
to
connects
to
all
the
back
ends
that
returns
by
the
resolver
and
for
the
pic
first
balancer.
It
will
connects
onto
only
one
of
the
back
hands
return
the
by
the
resolver
so
and
the
peak
first
and
the
peak
first
balancer
is
or
whatever
default
balancer
into
our
PC.
A
So
so
the
journalist
advanced,
instructed
repositories,
established
connections
and
AM
PC
will
returns
a
client
count,
which
is
an
abstraction
of
channel
inter
physical
to
user.
So
user
wallet
will
later
then
use
the
kind
country
to
initialize
the
generated
code.
Stop
that
user
can
turn
invoke.
Rpc
method,
on
top
of
so
last
to
me,
and
look
at
how
balancer
in
structure
RPC
to
create
those
connections
to
the
backend.
A
So
balancer
was
the
last
created
sub
channel
to
IP,
address
1
2
3
4,
which
the
first
pack,
which
is
the
first
back
end
and
the
DRAM
PC,
will
call
the
function,
quit
transport
the
trend
and
it
will
start
a
procedure
to
create
a
transport
credit
report.
Without
that
backhand
and
the
transport
the
world
then
kind
of
tries
to
connect
to
the
server.
Essentially,
here
is
doing
a
network
dial
and
then
on
the
server
side.
A
We
need
received
a
signal
that
there
is
a
coming
connection
request,
so
it
will
pass
along
the
info
to
the
G
RPC
saying
that
there's
a
new
coming
connection,
whether
you
want
to
accept
it
or
not,
your
PC
will
decide.
Ok,
let's
accept
the
connection
and
the
transport
will
pass
alone
to
accept
a
signal
back
to
the
client
transport
and
then
the
transfer.
Now
by
now,
it's
like
fully
set
up,
so
the
transport
will
return
and
return
itself
as
a
ready
transport
to
G
RPC.
A
So
so,
each
sub
channel
creation
process
is
following
the
step
from
1
to
7,
and
the
passo
may
request
multiple
connection
to
be
created,
and
they
are
all
under
the
same
procedure
here
and
now
we
have
the
channel.
We
are
ready
to
create
an
RPC
over
the
channel,
so
so
the
RPC
started
by
the
user,
calling
the
root
chat,
coroner
each
a
method
on
the
prod
under
generated
code
provided,
stop
initial
initiated
by
the
clang,
come
returned
by
the
char
PC
and
then
the
judge,
the
generated
code
were
called.
A
A
new
new
stream
function
is
to
let
your
PC
node
to
create
a
new
stream
for
the
RPC
and
equipment,
and
now
quick
tip
on
here
is
that
in
your
PC,
no
matter
what
is
the
generated
code,
layer
stream
or
GOP
silly
or
stream
or
strand
for
layering?
They
all
correspond
to
RPC
code,
so
you
can
simply
equal
them.
Eco
stream,
with
RPC
code,
when
you're
looking
at
our
code
base
so
champion,
will
tries
to
create
as
new
stream
for
the
RPC.
What
it
does
is.
A
A
And
next,
when
the
transport
returns,
the
transports
train
to
to
our
PC
jam
PC
will
wraps
will
create
a
wrapper
around
the
transport
stream
to
provide
functionality
like
compression
encoding,
retry
and
many
other
features
and
return.
The
rap
and
the
return.
The
rap
stream
to
the
generated
code
and
finally
generate
occur
well
created
a
very
C
wrapper
around
the
ER
PC,
just
pc
layer
stream
to
provide
type
Sinclair
type
safety
to
the
user,
which
is
much
better,
which
is
a
much
much
better
programming,
the
faithful
user
to
use.
A
So
user
will
write
a
user
function
based
on
the
interface
exported
by
the
defined
by
the
generated
code,
and
it
will
registered
it's
the
server
well,
the
user
will
register
function,
we
start
generating
code
and
then
generated
code
well,
will
create
a
wrapper
around
the
user
function
to
provide
set
type
safety
and
register
it
in
register
draft
handler
in
the
gr
PC.
So
that's
the
some
setup
here
and
so
the
on
the
client
side.
A
A
It
will
read
by
read
the
bytes
of
the
stream
and
passed
along
to
the
G
RPC,
and
the
danger
piece
of
a
path
alone
will
do
the
decompression
and
decoding
and
pass
along
the
message
back
to
the
generated
code
and
the
generated
code
will
pass
on
pass
along
democrat
request
to
the
server
and
the
same
same
stand,
and
we
see
procedure
happens
like
for
the
server
and
the
client.
So
basically,
this
is
the
the
read
pass
and
repass
of
her
RPC
and
I
would
like
to
just
stop
here.
A
So
you
guys
can
look
at
how
the
RPC
ends,
which
is
very
easy.
So
I
would
like
to
spend
a
couple
minutes
here
to
discuss
to
introduce
some
of
our
nice
features
of
Gerar
PC
select
I
said
before
we
have
a
pluggable
components
and
some
example:
here's
our
resolver
balancer,
ideal,
compressor,
codec
and
transport,
for
example,
for
ideal,
so
it
I
mean
beam.
A
Pluggable
means
that
you
don't
need
to
use
protobuf
to
be
able
to
use
a
RPC,
which
is
which
gives
you
so
much
more
complexity
here
and
also
I
would
like
to
briefly
introduce
some
of
our
recent
features.
Some
of
our
recent
features
like
panel
re
channel
tea
and
serve
something
for
your
feature.
I
won't
I,
don't
have
time
to
introduce
here.
You
can
go
to
our
repo
and
see
the
detailed
design
talk.
A
A
The
service
owner
will
need
to
push
the
service
config
file
to
a
service
account
in
hosting
service,
for
example,
a
DNS
server,
and
when
the
chirpies
client
tries
to
resolve
the
hostname,
it
will
try
to
see
whether
there's
a
service
config
under
their
hosts
name
and
if
there
is
it
will
pass
along
to
the
Jer,
PC
client
and
the
Jeff
is
declining,
would
then
apply
the
service
config
constraint
to
old
RPC.
For
that
for
the
connection
under
that
hostname
and,
for
example,
so,
what's
included
in
the
service
config,
for
example?
A
A
Oh
you'll
use
the
round-robin
server
and
he
also
says
about
the
RPC
timeouts,
our
the
request
response
message,
size
limit
and
actually
our
excuse
me
and
actually
our
new
features
like
ritual
RPC,
retrial,
health
tracking.
They
are
all
integrated
with
the
service
configure
a
central.
You
can
pass
the
configurations
through
the
service
campaign
and
next
the
banner
logging.
Finally
login
loss.
Rpc
RPC
is
in
binary
format.
A
Basically,
it
just
blocks
all
the
RPC
operations
in
a
in
a
binary
format
and
we,
along
with
the
time
stamp
and
the
coati
and
the
user,
needs
to
provide
a
logger
interface
a
lot
longer
to
be
to
log
listen
in
any
place.
You'd
like,
and
so
a
typical
use
case
of
for
binary
long
is
to
do
troubleshooting
service
finding
exception
during
load
testing
or
replaying
production
of
our
pieces.
A
Basically,
you
need
to
export
a
channel
D
service
in
your
binary
and
it
will
serve
the
job
and
you
need
to
serve
the
TRP,
see
our
universe
serve
the
chair,
PC
service,
so
so
the
end
user
can
can
get
info
from
those
service
and
what
are
the
info
provided
by
the
channel
T?
So
you
can
inquiry
like
how
many
scores
has
been
started,
succeeded
or
failed
at
a
channel
or
sub
Channel,
and
what
is
the
security
info?
Bada
transport
like
what
is
the
cycle
sweets?
A
Was
the
certificate
used
to
setup
this
transport
and
flow
control
info
like
how
many
cool
control
window
you
allow
for
the
current
life
for
this,
this
transport
and
the
last
and
freedom
who
also
input
like
what
is
the
last
are
busy
on
the
channel
and
how
many
streams
are
alive
currently
on
the
transport
and
there's
a
lot
a
lot
more.
There
seasick
for
them
to
yourself
and
I
have
wrote
a
very
short
introduction
to
channel.
So
you
can
find
the
link
on
the
bottom
and
the
so
channels.
A
A
Oh,
you
mean
the
entering
plan
for
a
piece
ago.
Implementation
I
think
like
right
now
so
I'd
like
said
here,
we
have
the
transport
to
be
pluggable
to
come.
That
feature
is
coming
soon
so
and
we
are
trying
to
make
our
transport
pluggable
and
after
we
made
that
change
them,
will
we
may
work
on
the
quick
transport
support.
A
B
B
A
Oh,
you
mean
where
you
can
configure
the
balancer,
so
we
provided
a
pig
first
balancer
and
they
roam
roam
a
palace
or
by
default.
So
if
you
want
to
any
other
behavior,
you
need
to
implement
a
palace
or
yourself.
So
you
can
decide
all
the
behavior
you
want.
We
have
about
a
nice
balance
or
API.
You
just
need
to
implement
that
to
interact
with
your
PC.