Zanzibar is an extensible framework to build configuration driven web applications. The goal of Zanzibar is to simplify application development into two steps:
- write configurations for the application and its components;
- write code to implement and test business logic.
Based on the configurations, Zanzibar generates boilerplates and glue code, wires them up with your business domain code and the runtime components Zanzibar provides to create a deployable binary.
The builtin components of Zanzibar makes it easy to develop microservices and gateway services that proxy or orchestrate microservices. It is also simple to extend Zanzibar with custom plugins to ease the development of applications that suit your specific needs.
Zanzibar is built on three pillars: module, config, code generation.
Modules are the components that a Zanzibar application is made of. A module belongs to a ModuleClass
, has a type
and can have dependencies on other modules.
ModuleClass
abstracts the functionality of a specific class of components. Zanzibar predefines a few module classes, i.e., client
, endpoint
, middleware
and service
. Each represents a corresponding abstraction:
ModuleClass | Abstraction |
---|---|
client | clients to communicate with downstreams, e.g., database clients and RPC clients |
endpoint | application interfaces exposed to upstreams |
middleware | common functionality that has less to do with business logic, e.g., rate limiting middleware |
service | a collection of endpoints that represents high level application abstraction, e.g., a demo service that prints "Hello World!" |
The module type
differentiates module instances of the same ModuleClass
with further classification. Types are somewhat arbitrary as they are not necessarily abstractions but indications about how Zanzibar should treat the modules.
A client module could be of type http
, tchannel
or custom
, where http
or tchannel
means Zanzibar will generate a client with given configuration that speaks that protocol while custom
means the client is fully provided and Zanzibar will use it as is without code generation. In other words, http
and tchannel
clients are configuration driven (no user code) whereas custom
clients are user-defined and can be "smart clients".
An endpoint
module could also be of type http
or tchannel
, which determines the protocol that the endpoint will be made available to invoke externally via the Zanzibar router. While endpoint
modules do not have custom
type, each method of an endpoint
module has a workflowType
that indicates the type of workflow the endpoint method fulfills. The builtin workflow type is httpClient
, tchannelClient
and custom
, where httpClient
and tchannelClient
means the endpoint method workflow is to proxy to a client, and custom
means the workflow is fulfilled by user code, see more in Custom Workflow.
Note that workflow type is likely to be deprecated in the future so that proxy to a client will be no longer a builtin option.
The builtin type of middleware module is default
.
The builtin service type is gateway
(it is likely to change in the future, because default
is probably a better name).
Note Zanzibar has support for user-defined module classes and module types in case the builtin types are not sufficient. The preferred way of extending Zanzibar is through user defined module classes and module types.
Module dependencies describe the relationships among various modules. The dependency relationship is critical to correctly assemble the modules to a full application.
A module is expected to define its immediate or direct dependencies. Zanzibar generates a module constructor with dependent modules as parameters, and passes the dependencies to the constructor during initilizaiton.
Zanzibar also constructs a full global dependency graph for a given set of modules. This graph is used to initialize modules in the correct order, e.g. leaf modules are initialized first and then passed to the constructors of parent modules for initialization.
To establish and enforce abstraction boundaries, dependency rules at ModuleClass
level are necessary. Zanzibar predefines the following dependency rules for built module classes:
ModuleClass | DependsOn | DependedBy |
---|---|---|
client | N/A | middleware, endpoint |
middleware | client | endpoint |
endpoint | client, middleware | service |
service | endpoint | N/A |
This table exhausts the possible immediate or direct dependency relationships among builtin module classes. Take endpoint module class for example, an endpoint module can depend on client or middleware modules but not endpoint or service modules. The reasoning for such rules aligns with the abstractions the module classes represent.
The ModuleClass
struct has DependsOn
and DependedBy
public fields, which makes it simple to extend the dependency rules with custom module class, e.g., we can define a custom module class task
that abstracts common business workflow by setting its DependsOn
field to client and DependedBy
field to endpoint.
Configurations are the interface that developers interact with when using the Zanzibar framework, they make up most of Zazibar's API. Various configurarions contain essential meta information of a Zanzibar application and its components. They are source of truth of the application.
Because configurations are the core of a Zanzibar application, we create a root directory to host configuration files when starting a Zanzibar application. There are a few typical directories and files under the root directory. Take example-gateway for example:
example-gateway # root directory
├── bin # directory for generated application binaries
│  └── example-gateway # generated example-gateway binary
├── build # directory for all generated code
│  ├── clients # generated mocks and module initializers for clients
│  ├── endpoints # generated mocks and module initializers for endpoints
│  ├── gen-code # generated structs and (de)serializers by Thrift compiler
│  ├── middlewares # generated module initializers for middlewares
│  │ └── default # generated module initializers for default middlewares
│  └── services # generated mocks and module intialziers for services
├── build.yaml # config file for Zanzibar code generation, see below for details
├── clients # config directory for modules of client module class
│  └── bar # config directory for a client named 'bar'
├── config # config directory for application runtime properties
│  ├── production.yaml # config file for production environment
│  └── test.yaml # config file for test environment
├── copyright_header.txt # optional copyright header for open source application
├── endpoints # config directory for modules of endpoint module class
│  └── bar # config directory for an endpoint named 'bar'
├── idl # idl directory for all thrift files
│  ├── clients # idl directory for client thrift files
│  └── endpoints # idl directory for endpoint thrift files
├── middlewares # config directory for modules of middleware module class
│  ├── transform-response # config directory for a middleware named 'transform-response'
│  ├── default # directory for all default middlewares
│  │ └── log-publisher # config directory for a default middleware named 'log-publisher'
│  └── default.yaml # config file describing default middlewares and their execution order
└── services # config directory for modules of service module class
└── example-gateway # config directory for a service named 'example-gateway'
Each module must have a config file so that it can be recognized by Zanzibar. This section explains how the module config files are organized and what goes into them.
Under the application root directory, there should be a corresponding top level config directory for each module class. For Zanzibar builtin module classes, the name of the directory is the plural of the module class name, e.g., a clients
directory for client
module class. The directory name is used when registering generator for a module class (example). While it is not required, the same directory naming convention should be followed when defining custom module classes.
Under a module class directory, there should be a corresponding config directory for each module, e.g., the clients
directory has a few subdirectories and each of them corresponds to a module.
Under a module directory, there should be a YAML file that contains the meta information of that module. It is required that the file is named of {$ModuleClass}-config.yaml
, e.g. the path to the YAML config file of bar
client module is clients/bar/client-config.yaml
, similarly the path to the YAML config file of bar
endpoint module is endpoints/bar/endpoint-config.yaml
.
Besides the YAML config file, the module directory also contains other necessary directories/files. For example, the quux client is a custom (non-generated) client, its module config directory has following layout:
quxx # client module config directory
├── client-config.yaml # client module config file
├── fixture # directory for fixtures used for testing
│  └── fixure.go # fixtures that can be used by a generated mock client for testing
└── quux.go # custom client implementation, package is imported by generated code
For client and endpoint modules of builtin type custom
, Zanzibar expects user code to be placed in the module directory. This is important because Zaznibar-generated code refers to user code by importing the package of the module directory path. Furthermore, user code of custom client and endpoint modules must also define and implement necessary public types and interfaces so that Zanzibar can wire up the modules.
For client module of custom type, user code must define a Client
interface and a NewClient
constructor that returns the Client
interface. Below is the example code snippet for the quux
custom client:
package quux
import "github.com/uber/zanzibar/examples/example-gateway/build/clients/quux/module"
type Client interface {
Echo(string) string
}
func NewClient(deps *module.Dependencies) Client {
return &quux{}
}
type quux struct{}
func (c *quux) Echo(s string) string { return s }
Note the type of deps
parameter passed to NewClient
constructor function is generated by Zanzibar, as indicated by the import path. Zanzibar takes care of initializing and passing in the acutal deps
argument, as mentioned in Dependency Injection.
For increasing overall system resiliency, zanzibar uses a Circuit Breaker which avoids calling client when there is an increase in failure rate beyond a set threshold. After a sleepWindowInMilliseconds, client calls are attempted recovery by going in half-open and then close state.
circuitBreakerDisabled: Default false. To disable the circuit-breaker:
"clients.<clientID>.circuitBreakerDisabled" : true
maxConcurrentRequests: Default 50. To set how many requests can be run at the same time, beyond which requests are rejected:
"clients.<clientID>.maxConcurrentRequests": 50
errorPercentThreshold: Default 20. To set error percent threshold beyond which to trip the circuit open:
"clients.<clientID>.errorPercentThreshold": 20
requestVolumeThreshold: Default 20. To set minimum number of requests that will trip the circuit in a rolling window of 10 (For example, if the value is 20, then if only 19 requests are received in the rolling window of 10 seconds the circuit will not trip open even if all 19 failed):
"clients.<clientID>.requestVolumeThreshold" : true
sleepWindowInMilliseconds: Default 5000. To set the amount of time, after tripping the circuit, to reject requests before allowing attempts again to determine if the circuit should again be closed:
"clients.<clientID>.sleepWindowInMilliseconds" : true
For endpoint module of custom workflow type, user code must define a New{$endpoint}{$method}Workflow
constructor that returns the Zanzibar-generated {$endpoint}{$method}Workflow
interface which has a sole Handle
method. Below is the example code snippet for the contacts
custom endpoint:
package contacts
import (
"context"
"github.com/uber/zanzibar/examples/example-gateway/build/endpoints/contacts/module"
"github.com/uber/zanzibar/examples/example-gateway/build/endpoints/contacts/workflow"
contacts "github.com/uber/zanzibar/examples/example-gateway/build/gen-code/endpoints-idl/endpoints/contacts/contacts"
zanzibar "github.com/uber/zanzibar/runtime"
"go.uber.org/zap"
)
func NewContactsSaveContactsWorkflow(
c *module.ClientDependencies,
l *zap.Logger,
) workflow.ContactsSaveContactsWorkflow { return &saveContacts{ ... } }
type saveContacts struct { ... }
func (w *saveContacts) Handle(
ctx context.Context,
headers zanzibar.Header,
r *contacts.SaveContactsRequest,
) (*contacts.SaveContactsResponse, zanzibar.Header, error) { ... }
The idea of the workflow constructor is similar to the client constructor, with a couple of differences:
- the first parameter is specifically
ClientDependencies
and there is an additional logger parameter, this will be changed in the future so that the dependency parameter is generalized; - the return value is an interface generated by Zanzibar, the parameter and return value of the
Handle
method refers to structs generated by Thrift compiler based on the endpoint thrift file configured in the endpoint-config.yaml, see more in Config Schema.
Zanzibar allows nesting module config directories in the sub-directories of a module class config directory. This is useful to group related modules under a sub-directory. For example, the tchannel directory groups all TChannel endpoint modules:
endpoints
├── ...
└── tchannel # this directory does not correspond to a module, it represents a group
└── baz # module config directory under the 'tchannel' group
├── baz_call.go
├── baz_call_test.go
├── call.yaml
└── endpoint-config.yaml
Modules of different ModuleClass
and type
are likely to have different fields in their config files. Developers are expected to write module config files according to the schemas.
Note: fields are absent in config schemas but present in examples are experimental.
The endpoint module config is different from other module classes as it has multiple YAML files, where each endpoint method corresponds to a YAML file and the endpoint-config.yaml file refers to them.
endpoints/multi
├── endpoint-config.yaml # has a field 'endpoints' that is a list and contains helloA and helloB
├── helloA.yaml # config file for method helloA
└── helloB.yaml # config file for method helloB
The reason for such layout is to avoid a large endpoint-config.yaml file when an endpoint has many methods.
Besides the module configs, Zanzibar also expects a YAML file that configures necessary properties to boostrap the code generation process of a Zanzibar application. The schema for application config is defined here.
Unlike the module configs, there is no restriction on how this config file should be named. It can be named {$appName}.yaml
or build.yaml
as it is in example-gateway, as long as it is passed correctly as an argument to the code generation runner.
In this config file, you can specify the paths from which to discover modules. You can also specify default dependencies
.
Default Dependencies
allow module classes to include instances of other module classes as default dependencies. This means that no explicit configurations are required for certain module instances to be included as a dependency. e.g., we can include clients/logger
as a default dependency for endpoint
, and every endpoint will have clients/logger
as a dependency in its module/dependencies.go
file, even if the endpoint's endpoint-config.yaml
file does not list clients/logger
as a dependency.
Note that these paths support Glob
patterns.
Zanzibar provides HTTP and TChannel runtime components for both clients and servers. Once all the configs are properly defined, Zanzibar is able to parse the config files and generate code and wire it up with the runime components to produce a full application. All generated code is placed in the build
directory.
Zanzibar expects non-custom clients and endpoints to define their interfaces using Thrift (Zanzibar Thrift file semantics). For example, the bar
endpoint defines its interfaces using the bar.thrift as specified in hello.yaml. The data types in such thrift files must have their equivalents in Go.
- For tchannel clients/endpoints, network communication is Thrift over TChannel. Zanzibar uses thriftrw to generate Go structs and thrift (de)serializers;
- For http clients/endpoints, network communication is JSON over HTTP. Zanzibar uses thriftrw to generate Go structs and then uses easyjson to generate JSON (de)serializers.
The pre-steps.sh script takes care of this part of the code generation, and places the generated code under build/gen-code
directory.
Everything except gen-code
under build
directory is generated by Zanzibar. Zanzibar parses config files for each module to gathers meta information and then executing various templates by applying them to the meta data. Here is what is generated for each builtin module class:
- client: dependency type, client interface and constructor if non-custom, mock client constructor
- middleware: dependency type, middleware type and constructor (unstable)
- endpoint: dependency type, endpoint type and constructor, workflow interface, workflow if non-custom, mock workflow constructor if custom
- service: dependency type and initializer, main.go, mock service constructor, service constructor
Assuming you are using a vendor package management tool like Glide, then the minimal glide.yaml file would look like:
- package: go.uber.org/thriftrw
version: ^1.8.0
- package: github.com/mailru/easyjson
version: master
- package: github.com/uber/zanzibar
version: master
After installing the packages, create your module configs and application config in your application root directory. Then you are ready to run the following script to kick off code generation:
# put this script in application root directory
CONFIG_DIR="."
BUILD_DIR="$CONFIG_DIR/build"
THRIFTRW_SRCS=""
# find all thrift files specified in the config files
config_files=$(find "." -name "*-config.yaml" ! -path "*/build/*" ! -path "*/vendor/*" | sort)
for config_file in ${config_files}; do
dir=$(dirname "$config_file")
yaml_files=$(find "$dir" -name "*.yaml")
for yaml_file in ${yaml_files}; do
thrift_file=$(yq -r '.. | .idlFile? | select(strings | endswith(".thrift"))' "$yaml_file")
[[ -z ${thrift_file} ]] && continue
[[ ${THRIFTRW_SRCS} == *${thrift_file}* ]] && continue
THRIFTRW_SRCS+=" $CONFIG_DIR/idl/$thrift_file"
done
done
bash ./vendor/github.com/uber/zanzibar/codegen/runner/pre-steps.sh "$BUILD_DIR" "$CONFIG_DIR" "zanzibar" "$THRIFTRW_SRCS"
go run ./vendor/github.com/uber/zanzibar/codegen/runner/runner.go --config="$CONFIG_DIR/build.yaml"
Note the above script will be abstracted for easier usage in the future.
Zanzibar comes with builtin integration testing frameworks to help test business logic with ease. Setting genMock to true will trigger Zanzibar to generate mock client, workflow and service constructors. The mock clients, being the leaf nodes in the dependency graph, are wired with the rest modules to create a testing application, which you can test against by setting expectations of the mock clients. The generated test helpers make writing tests straightforward and concise.
Currently Zanzibar provides two entry points to write integration tests: service and endpoint.
Service level integration testing treats your application as a black box. Zanzibar starts a local server for your application and you write tests by sending requests to the server and verify the response is expected.
func TestSaveContacts(t *testing.T) {
ms := ms.MustCreateTestService(t)
ms.Start()
defer ms.Stop()
ms.MockClients().Contacts.ExpectSaveContacts().Success()
endpointReqeust := &endpointContacts.SaveContactsRequest{
Contacts: []*endpointContacts.Contact{},
}
rawBody, _ := endpointReqeust.MarshalJSON()
res, err := ms.MakeHTTPRequest(
"POST", "/contacts/foo/contacts", nil, bytes.NewReader(rawBody),
)
if !assert.NoError(t, err, "got http error") {
return
}
assert.Equal(t, "202 Accepted", res.Status)
}
Endpoint level integration testing allows focusing on testing the business logic without a full server setup. It is lightweighted and feels more like unit tests.
func TestSaveContactsCallWorkflow(t *testing.T) {
mh, mc := mockcontactsworkflow.NewContactsSaveContactsWorkflowMock(t)
mc.Contacts.ExpectSaveContacts().Success()
endpointReqeust := &endpointContacts.SaveContactsRequest{
UserUUID: "foo",
Contacts: []*endpointContacts.Contact{},
}
res, resHeaders, err := mh.Handle(context.Background(), nil, endpointReqeust)
if !assert.NoError(t, err, "got error") {
return
}
assert.Nil(t, resHeaders)
assert.Equal(t, &endpointContacts.SaveContactsResponse{}, res)
}
The above snippets can be found in save_contacts_test.go.
Zanzibar uses gomock to generate client mocks. To avoid manually setting the same fixture expectations again and again, Zanzibar augments gomock-generated mocks with fixture support. For example, the client-config.yaml file of contacts
client has a fixture
field:
fixture:
importPath: github.com/uber/zanzibar/examples/example-gateway/clients/contacts/fixture
scenarios:
SaveContacts:
- success
This basically says the saveContacts
method has a success
scenario which is defined in the fixture package indicated by the importPath
. The fixture package is provided by users and here is what it looks like:
package fixture
import (
mc "github.com/uber/zanzibar/examples/example-gateway/build/clients/contacts/mock-client"
gen "github.com/uber/zanzibar/examples/example-gateway/build/gen-code/clients-idl/clients/contacts/contacts"
)
var saveContactsFixtures = &mc.SaveContactsScenarios{
Success: &mc.SaveContactsFixture{
Arg0Any: true,
Arg1Any: true,
Arg2: &gen.SaveContactsRequest{
UserUUID: "foo",
},
Ret0: &gen.SaveContactsResponse{},
},
}
// Fixture ...
var Fixture = &mc.ClientFixture{
SaveContacts: saveContactsFixtures,
}
With that, in your tests you will be able to write
mc.Contacts.ExpectSaveContacts().Success()
rather than
s.mockClient.EXPECT().SaveContacts(arg0, arg1, arg2).Return(ret0, ret1, ret2)
Check out fixture abstraction to see how it works.
Once the concepts of module, config and code generation are clear, extending Zanzibar becomes straightforward. There are two ways to extend Zanzibar.
To extend Zanzibar with new module class or type is simply to extend each of its three pillars. For example, we want to add a new task
module class to abstract common business workflow, here is what we need to do for each pillar:
- module: understand what meta information is needed for each task module;
- config: add a
tasks
directory under the application root directory, define proper schema for task module class; - code generation: add templates for task if necessary, create a code generator that implements the BuildGenerator interface and register it onto the module system for the task module class.
The same idea applies for adding new types of an existing module class.
Zanzibar provides post-generation hooks which has access to the meta information of all modules. You can do whatever (mutating the input is probably not a good idea) suits your needs within a post-generation hook. Zanzibar invokes post-generation hooks as the very last step of code generation. In fact, mocks are all generated via post-generation hooks.
mkdir -p $GOPATH/src/github.com/uber
git clone [email protected]:uber/zanzibar $GOPATH/src/github.com/uber/zanzibar
cd $GOPATH/src/github.com/uber/zanzibar
GO111MODULE=off make install
make generate
make test
for i in `seq 5`; do make bench; done
First fetch wrk
git clone https://github.com/wg/wrk ~/wrk
cd ~/wrk
make
sudo ln -s $HOME/wrk/wrk /usr/local/bin/wrk
Then you can run the benchmark comparison script
# Assume you are on feature branch ABC
./benchmarks/compare_to.sh master
First create log dir...
sudo mkdir -p /var/log/my-gateway
sudo chown $USER /var/log/my-gateway
chmod 755 /var/log/my-gateway
sudo mkdir -p /var/log/example-gateway
sudo chown $USER /var/log/example-gateway
chmod 755 /var/log/example-gateway
make run
# Logs are in /var/log/example-gateway/example-gateway.log
We use glide @ 0.12.3 to add dependencies.
Download glide @ 0.12.3 and make sure it's available in your path
If we want to add a dependency:
- Add a new section to the glide.yaml with your package and version
- run
glide up --quick
- check in the
glide.yaml
andglide.lock
If you want to update a dependency:
- Change the
version
field in theglide.yaml
- run
glide up --quick
- check in the
glide.yaml
andglide.lock