1. Google confidential │ Do not
distribute
Google confidential │ Do not
distribute
Bringing learnings from
Googley microservices with
gRPC
Microservices Summit
Varun Talwar
2. Contents
1. Context: Why are we here?
2. Learnings from Stubby experience
a. HTTP/JSON doesnt cut it
b. Establish a Lingua Franca
c. Design for fault tolerance and control: Sync/Async, Deadlines, Cancellations, Flow control
d. Flying blind without stats
e. Diagnosing with tracing
f. Load Balancing is critical
3. gRPC
12. Key learnings
1. HTTP/JSON doesnt cut it !
2. Establish a lingua franca
3. Design for fault tolerance and provide control knobs
4. Dont fly blind: Service Analytics
5. Diagnosing problems: Tracing
6. Load Balancing is critical
13. HTTP1.x/JSON doesn’t cut it !
1. WWW, browser growth - bled into services
2. Stateless
3. Text on the wire
4. Loose contracts
5. TCP connection per request
6. Nouns based
7. Harder API evolution
1
14. Establish a lingua franca
1. Protocol Buffers - Since 2003.
2. Start with IDL
3. Have a language agnostic way of agreeing on data semantics
4. Code Gen in various languages
5. Forward and Backward compatibility
6. API Evolution
2
16. Google Cloud Platform
Service Definition (weather.proto)
syntax = "proto3";
service Weather {
rpc GetCurrent(WeatherRequest) returns (WeatherResponse);
}
message WeatherRequest {
Coordinates coordinates = 1;
message Coordinates {
fixed64 latitude = 1;
fixed64 longitude = 2;
}
}
message WeatherResponse {
Temperature temperature = 1;
float humidity = 2;
}
message Temperature {
float degrees = 1;
Units units = 2;
enum Units {
FAHRENHEIT = 0;
CELSIUS = 1;
KELVIN = 2;
}
}
17. Design for fault tolerance and control
Sync and Async APIs
Need fault tolerance: Deadlines, Cancellations
Control Knobs: Flow control, Service Config, Metadata
3
18. Google Cloud Platform 18
First-class feature in gRPC.
Deadline is an absolute point in time.
Deadline indicates to the server how
long the client is willing to wait for an
answer.
RPC will fail with DEADLINE_EXCEEDED
status code when deadline reached.
gRPC Deadlines
19. Google Cloud Platform
Deadline Propagation
Gateway
90 ms
Now =
1476600000000
Deadline =
1476600000200
40 ms
20 ms
20 ms 60 ms
withDeadlineAfter(200, MILLISECONDS)
Now =
1476600000040
Deadline =
1476600000200
Now =
1476600000150
Deadline =
1476600000200
Now =
1476600000230
Deadline =
1476600000200
DEADLINE_EXCEEDED DEADLINE_EXCEEDED DEADLINE_EXCEEDED DEADLINE_EXCEEDED
20. Google Cloud Platform 20
Deadlines are expected.
What about unpredictable cancellations?
•User cancelled request.
•Caller is not interested in the result any
more.
•etc
Cancellation?
21. Google Cloud Platform
Cancellation?
GW
Busy Busy Busy
Busy Busy Busy
Busy Busy Busy
Active RPC Active RPC
Active RPC
Active RPC Active RPCActive RPC
Active RPC Active RPC
Active RPC
23. Google Cloud Platform 23
Automatically propagated.
RPC fails with CANCELLED status code.
Cancellation status be accessed by the
receiver.
Server (receiver) always knows if RPC is
valid!
Cancellation
24. Google Cloud Platform
BiDi Streaming - Slow Client
Fast Server
Request
Responses
Slow Client
CANCELLED
UNAVAILABLE
RESOURCE_EXHAUSTED
25. Google Cloud Platform
BiDi Streaming - Slow Server
Slow Server
Request
Response
Fast Client
CANCELLED
UNAVAILABLE
RESOURCE_EXHAUSTED
Requests
26. Google Cloud Platform 26
Flow-control helps to balance
computing power and network
capacity between client and server.
gRPC supports both client- and
server-side flow control.
Flow-Control
Photo taken by Andrey Borisenko.
27. Google Cloud Platform 27
Policies where server tells client what
they should do
Can specify deadlines, lb policy,
payload size per method of a service
Loved by SREs, they have more control
Discovery via DNS
Service Config
28. Metadata Exchange - Common cross-cutting concerns
like authentication or tracing rely on the exchange of
data that is not part of the declared interface of a
service. Deployments rely on their ability to evolve these
features at a different rate to the individual APIs
exposed by services.
Metadata helps in exchange of useful information
29. Don’t fly blind: Stats4
What is the mean latency time per RPC?
How many RPCs per hour for a service?
Errors in last minute/hour?
How many bytes sent? How many connections to my server?
30. Data collection by arbitrary metadata is useful
Any service’s resource usage and performance stats in real time by (almost)
any arbitrary metadata
1. Service X can monitor CPU usage in their jobs broken down by the name of the invoked RPC
and the mdb user who sent it.
2. Ads can monitor the RPC latency of shared bigtable jobs when responding to their requests,
broken down by whether the request originated from a user on web/Android/iOS.
3. Gmail can collect usage on servers, broken down by according POP/IMAP/web/Android/iOS.
Layer propagates Gmail's metadata down to every service, even if the request was made by an
intermediary job that Gmail doesn't own
31. Diagnosing problems: Tracing5
1/10K requests takes very long. Its an ad query :-) I need to find out.
Take a sample and store in database; help identify request in sample which
took similar amount of time
I didnt get a response from the service. What happened? Which link in the
service dependency graph got stuck? Stitch a trace and figure out.
Where is it taking time for a trace? Hotspot analysis
What all are the dependencies for a service?
32. Load Balancing is important !5
Iteration 1: Stubby Balancer
Iteration 2: Client side load balancing
Iteration 3: Hybrid
Iteration 4: gRPC-lb
33. ● Round-robin-over-list - Lists not sets → ability to represent weights
● For anything more advanced, move the burden to an external "LB Controller", a
regular gRPC server and rely on a client-side implementation of the so-called
gRPC LB policy.
client LB Controller
backends
1) Control RPC
2) address-list
3) RR over addresses of
address-list
gRPC LB
Some new ideas !
Iteration 1: Stubby Balancer
Iteration 2: Client side load balancing
Iteration 3: Hybrid
Iteration 4: gRPC-lb
34. In summary, what did we learn
Contracts should be strict
Common language helps
Common understanding for deadlines, cancellations, flow control
Common stats/tracing framework is essential for monitoring, debugging
Common framework lets uniform policy application for control and lb
Single point of integration for logging, monitoring, tracing, service
discovery and load balancing makes lives much easier !
36. Open source on Github for C, C++, Java, Node.js,
Python, Ruby, Go, C#, PHP, Objective-C
gRPC core
gRPC Java
gRPC Go
37. 1.0 with stable APIs
Well documented with an active community
Reliable with continuous running tests on GCE
Deployable in your environment
Measured with an open performance dashboard
Deployable in your environment
Well adopted inside and outside Google
Where is the project today?
38. 1. Cross language & Cross platform matters !
2. Performance and Standards matter: HTTP/2
3. Pluggability matters: Interceptors, Name Resolvers,
Auth plugins
4. Usability matters !
More lessons
39. 1. Cross language & Cross platform matters !
2. Performance and Standards matter: HTTP/2
3. Pluggability matters: Interceptors, Name Resolvers,
Auth plugins
4. Usability matters !
More lessons
40. Google Cloud PlatformGoogle Cloud Platform
Coverage & Simplicity
The stack should be available on every popular
development platform and easy for someone to build
for their platform of choice. It should be viable on
CPU & memory limited devices.
gRPC Principles & Requirements
http://www.grpc.io/blog/principles
41. Google Cloud Platform
gRPC Speaks Your Language
● Java
● Go
● C/C++
● C#
● Node.js
● PHP
● Ruby
● Python
● Objective-C
● MacOS
● Linux
● Windows
● Android
● iOS
Service definitions and client libraries Platforms supported
43. 1. Cross language & Cross platform matters !
2. Performance and Standards matter: HTTP/2
3. Pluggability matters: Interceptors, Name Resolvers,
Auth plugins
4. Usability matters !
More lessons
44. Google Cloud Platform
• Single TCP connection.
• No Head-of-line
blocking.
• Binary framing layer.
• Request –> Stream.
• Header Compression.
HTTP/2 in One Slide
Transport(TCP)
Application (HTTP/2)
Network (IP)
Session (TLS) [optional]
Binary Framing
HEADERS Frame
DATA Frame
HTTP/2
POST: /upload
HTTP/1.1
Host: www.javaday.org.ua
Content-Type: application/json
Content-Length: 27
HTTP/1.x
{“msg”: “Welcome to 2016!”}
45. Google Cloud Platform
HTTP/2 breaks down the
HTTP protocol
communication into an
exchange of binary-
encoded frames, which
are then mapped to
messages that belong to a
stream, and all of which
are multiplexed within a
single TCP connection.
Binary Framing
Stream 1 HEADERS
Stream 2
:method: GET
:path: /kyiv
:version: HTTP/2
:scheme: https
HEADERS
:status: 200
:version: HTTP/2
:server: nginx/1.10.1
...
DATA
<payload>
Stream N
Request
Response
TCP
47. Google Cloud Platform
gRPC Service Definitions
Unary RPCs where the
client sends a single
request to the server
and gets a single
response back, just like
a normal function call.
The client sends a
request to the server
and gets a stream to
read a sequence of
messages back.
The client reads from
the returned stream
until there are no more
messages.
The client send a
sequence of messages
to the server using a
provided stream.
Once the client has
finished writing the
messages, it waits for
the server to read them
and return its response.
Client streaming
Both sides send a
sequence of messages
using a read-write
stream. The two
streams operate
independently. The
order of messages in
each stream is
preserved.
BiDi streamingUnary Server streaming
48. Google Cloud Platform 48
Messaging applications.
Games / multiplayer tournaments.
Moving objects.
Sport results.
Stock market quotes.
Smart home devices.
You name it!
BiDi Streaming Use-Cases
49. Open Performance Benchmark and Dashboard
Benchmarks run in GCE VMs per Pull Request for regression testing.
gRPC Users can run these in their environments.
Good Performance across languages:
Java Throughput: 500 K RPCs/Sec and 1.3 M Streaming messages/Sec on 32 core VMs
Java Latency: ~320 us for unary ping-pong (netperf 120us)
C++ Throughput: ~1.3 M RPCs/Sec and 3 M Streaming Messages/Sec on 32 core VMs.
Performance
50.
51.
52. 1. Cross language & Cross platform matters !
2. Performance and Standards matter: HTTP/2
3. Pluggability matters: Interceptors, Auth
4. Usability matters !
More lessons
53. Google Cloud PlatformGoogle Cloud Platform
Pluggable
Large distributed systems need security, health-
checking, load-balancing and failover, monitoring,
tracing, logging, and so on. Implementations should
provide extensions points to allow for plugging in
these features and, where useful, default
implementations.
gRPC Principles & Requirements
http://www.grpc.io/blog/principles
58. 1. Server reflection
2. Health Checking
3. Automatic retries
4. Streaming compression
5. Mechanism to do caching
6. Binary Logging
a. Debugging, auditing though costly
7. Unit Testing support
a. Automated mock testing
b. Dont need to bring up all dependent services just to test
8. Web support
Coming soon !
59. Microservices: in data centres
Streaming telemetry from network devices
Client Server communication/Internal APIs
Mobile Apps
Some early adopters