cosmos-sdk/docs/basics/app-anatomy.md

Anatomy of an SDK Application

Pre-requisite reading

• High-level overview of the architecture of an SDK application
• Cosmos SDK design overview

Synopsis

This document describes the core parts of a Cosmos SDK application. The placeholder name for this application will be app.

• Node Client
• Core Application File
• Modules
• Interfaces
• Dependencies and Makefile

The core parts listed above will generally translate to the following directory tree:

./app
├── cmd/
│   ├── appd
│   └── appcli
├── app.go
├── x/
│   ├── auth
│   ├── ...
│   └── bank
├── go.mod
└── Makefile

Node Client

The Daemon, or Full-Node Client, is the core process of an SDK-based blockchain. Participants in the network run this process to initialize their state-machine, connect with other full-nodes and update their state-machine as new blocks come in.

                ^  +-------------------------------+  ^
                |  |                               |  |
                |  |  State-machine = Application  |  |
                |  |                               |  |   Built with Cosmos SDK
                |  |            ^      +           |  |
                |  +----------- | ABCI | ----------+  v
                |  |            +      v           |  ^
                |  |                               |  |
Blockchain Node |  |           Consensus           |  |
                |  |                               |  |
                |  +-------------------------------+  |   Tendermint Core
                |  |                               |  |
                |  |           Networking          |  |
                |  |                               |  |
                v  +-------------------------------+  v

The blockchain full-node presents itself as a binary, generally suffixed by -d for "daemon" (e.g. appd for app or gaiad for gaia). This binary is built by running a simple main.go function placed in cmd/appd/. This operation usually happens through the Makefile.

To learn more about the main.go function, click here.

Once the main binary is built, the node can be started by running the start command. The core logic behind the start command is implemented in the SDK itself in the /server/start.go file. The main start command function takes a context and appCreator as arguments. The appCreator is a constructor function for the SDK application, and is used in the starting process of the full-node.

The start command function primarily does three things:

1. Create an instance of the state-machine defined in app.go using the appCreator.
2. Initialize the state-machine with the latest known state, extracted from the db stored in the ~/.appd/data folder. At this point, the state-machine is at height appBlockHeight.
3. Create and start a new Tendermint instance. Among other things, the node will perform a handshake with its peers. It will get the latest blockHeight from them, and replay blocks to sync to this height if it is greater than the local appBlockHeight. If appBlockHeight is 0, the node is starting from genesis and Tendermint sends an InitChain message via the ABCI to the app, which triggers the InitChainer.

To learn more about the start command, click here.

Core Application File

In general, the core of the state-machine is defined in a file called app.go. It mainly contains the type definition of the application and functions to create and initialize it.

Type Definition of the Application

The first thing defined in app.go is the type of the application. It is generally comprised of the following parts:

• A reference to baseapp. The custom application defined in app.go is an extension of the baseapp type. baseapp implements most of the core logic for the application, including all the ABCI methods and the routing logic. When a transaction is relayed by Tendermint to the application, the latter uses baseapp's methods to route them to the appropriate module.
• A list of store keys. The store, which contains the entire state, is implemented as a multistore (i.e. a store of stores) in the Cosmos SDK. Each module uses one or multiple stores in the multistore to persist their part of the state. These stores can be accessed with specific keys that are declared in the app type. These keys, along with the keepers, are at the heart of the object-capabilities model of the Cosmos SDK.
• A list of module's keepers. Each module defines an abstraction called keeper, which handles reads and writes for this module's store(s). The keeper's methods of one module can be called from other modules (if authorized), which is why they are declared in the application's type and exported as interfaces to other modules so that they are only allowed to access the authorized functions.
• A reference to a codec. The Cosmos SDK gives developers the freedom to choose the encoding framework for their application. The application's codec is used to serialize and deserialize data structures in order to store them, as stores can only persist []bytes. The codec must be deterministic. The default codec is amino.
• A reference to a module manager. The module manager is an object that contains a list of the application's module. It facilitates operations related to these modules, like registering routes, query routes or setting the order of execution between modules for various functions like InitChainer, BeginBlocker and EndBlocker.

You can see an example of application type definition here.

Constructor Function

This function constructs a new application of the type defined above. It is called every time the full-node is started with the start command. Here are the main actions performed by this function:

• Instantiate a new application with a reference to a baseapp instance, a codec and all the appropriate store keys.
• Instantiate all the keepers defined in the application's type using the NewKeeper function of each of the application's modules. Note that keepers must be instantiated in the correct order, as the NewKeeper of one module might require a reference to another module's keeper.
• Instantiate the application's module manager with the AppModule object of each of the application's modules.
• With the module manager, initialize the application's routes and query routes. When a transaction is relayed to the application by Tendermint via the ABCI, it is routed to the appropriate module's handler using the routes defined here. Likewise, when a query is received by the application, it is routed to the appropriate module's querier using the query routes defined here.
• With the module manager, register the application's modules' invariants. Invariants are variables (e.g. total supply of a token) that are evaluated at the end of each block. The process of checking invariants is done via a special module called the InvariantsRegistry. The value of the invariant should be equal to a predicted value defined in the module. Should the value be different than the predicted one, special logic defined in the invariant registry will be triggered (usually the chain is halted). This is useful to make sure no critical bug goes unnoticed and produces long-lasting effects that would be hard to fix.
• With the module manager, set the order of execution between the InitGenesis, BegingBlocker and EndBlocker functions of each of the application's modules. Note that not all modules implement these functions.
• Set the remainer of application's parameters:
  • InitChainer: used to initialize the application when it is first started.
  • BeginBlocker, EndBlocker: called at the beginning and the end of every block).
  • anteHandler: used to handle fees and signature verification.
• Mount the stores.
• Return the application.

Note that this function only creates an instance of the app, while the actual state is either carried over from the ~/.appd/data folder if the node is restarted, or generated from the genesis file if the node is started for the first time.

You can see an example of application constructor here.

InitChainer

The InitChainer is a function that initializes the state of the application from a genesis file (i.e. token balances of genesis accounts). It is called when the application receives the InitChain message from the Tendermint engine, which happens when the node is started at appBlockHeight == 0 (i.e. on genesis). The application must set the InitChainer in its constructor via the SetInitChainer method.

In general, the InitChainer is mostly composed of the InitGenesis function of each of the application's modules. This is done by calling the InitGenesis function of the module manager, which in turn will call the InitGenesis function of each of the modules it contains. Note that the order in which the modules' InitGenesis functions must be called has to be set in the module manager using the SetOrderInitGenesis method. This is done in the application's constructor, and the SetOrderInitGenesis has to be called before the SetInitChainer.

You can see an example of an InitChainer here.

BeginBlocker and EndBlocker

The SDK offers developers the possibility to implement automatic execution of code as part of their application. This is implemented through two function called BeginBlocker and EndBlocker. They are called when the application receives respectively the BeginBlock and EndBlock messages from the Tendermint engine, which happens at the beginning and at the end of each block. The application must set the BeginBlocker and EndBlocker in its constructor via the SetBeginBlocker and SetEndBlocker methods.

In general, the BeginBlocker and EndBlocker functions are mostly composed of the BeginBlock and EndBlock functions of each of the application's modules. This is done by calling the BeginBlock and EndBlock functions of the module manager, which in turn will call the BeginBLock and EndBlock functions of each of the modules it contains. Note that the order in which the modules' BegingBlock and EndBlock functions must be called has to be set in the module manager using the SetOrderBeginBlock and SetOrderEndBlock methods respectively. This is done in the application's constructor, and the SetOrderBeginBlock and SetOrderEndBlock methods have to be called before the SetBeginBlocker and SetEndBlocker functions.

As a sidenote, it is important to remember that application-specific blockchains are deterministic. Developers must be careful not to introduce non-determinism in BeginBlocker or EndBlocker, and must also be careful not to make them too computationally expensive, as gas does not constrain the cost of BeginBlocker and EndBlocker execution.

You can see an example of BeginBlocker and EndBlocker functions here.

Register Codec

The MakeCodec function is the last important function of the app.go file. The goal of this function is to instantiate a codec cdc (e.g. amino) initiliaze the codec of the SDK and each of the application's modules using the RegisterCodec function.

To register the application's modules, the MakeCodec function calls RegisterCodec on ModuleBasics. ModuleBasics is a basic manager which lists all of the application's modules. It is instanciated in the init() function, and only serves to easily register non-dependant elements of application's modules (such as codec). To learn more about the basic module manager, click here.

You can see an example of a MakeCodec here

Modules

Modules are the heart and soul of an SDK application. They can be considered as state-machines within the state-machine. When a transaction is relayed from the underlying Tendermint engine via the ABCI to the application, it is routed by baseapp to the appropriate module in order to be processed. This paradigm enables developers to easily build complex state-machines, as most of the modules they need often already exist. For developers, most of the work involved in building an SDK application revolves around building custom modules required by their application that do not exist, and integrating them with modules that do already exist into one coherent application. In the application directory, the standard practice is to store modules in the x/ folder (not to be confused with the SDK's x/ folder, which contains already-built modules).

To learn more about modules, click here

Application Module Interface

Modules implement two interfaces defined in the Cosmos SDK, AppModuleBasic and AppModule. The former implements basic non-dependant elements of the module, such as the codec, while the latter handles the bulk of the module methods (including methods that require references to other modules' keepers). Both the AppModule and AppModuleBasic types are defined in a file called ./module.go.

AppModule exposes a collection of useful methods on the module that facilitates the composition of modules into a coherent application. Important methods include:

• Route() and QueryRoute(): These methods the name of the route and querier route for the module, for messages to be routed to the module's handler and queries to be routes to the module's querier.
• NewHandler() and NewQuerierHandler(): These methods return a handler and querierHandler respectively, in order to process a message or a query once they are routed.
• BeginBlock(), EndBlock() and InitGenesis(): These methods are executed respectively at the beginning of each block, at the end of each block and at the start of the chain. They implement special logic the module requires to be triggered during those events. For example, the EndBlock function is frequently used by modules where voting occurs to tally the result of the votes.
• RegisterInvariants(): This method registers the invariants for the module. Invariants are checked at the end of every block to make sure no unpredicted behaviour is occuring.

AppModule's methods are called from the module manager(./modules.md#module-manager), which manages the application's collection of modules.

To learn more about the application module interface, click here.

Message Types

A message is a custom type defined by each module that implements the message interface. Each transaction contains one or multiple messages. When a valid block of transactions is received by the full-node, Tendermint relays each one to the application via DeliverTx. Then, the application handles the transaction:

1. Upon receiving the transaction, the application first unmarshalls it from []bytes.
2. Then, it verifies a few things about the transaction like fee payment and signatures before extracting the message(s) contained in the transaction.
3. With the Type() method, baseapp is able to know which modules defines the message. It is then able to route it to the appropriate module's handler in order for the message to be processed.
4. If the message is successfully processed, the state is updated.

For a more detailed look at a transaction lifecycle, click here.

Module developers create custom message types when they build their own module. The general practice is to prefix the type declaration of the message with Msg. For example, the message type MsgSend allows users to transfer tokens. It is processed by the handler of the bank module, which ultimately calls the keeper of the auth module in order to update the state.

To learn more about messages, click here.

Handler

The handler refers to the part of the module responsible for processing the message after it is routed by baseapp. handler functions of modules (except those of the auth module) are only executed if the transaction is relayed from Tendermint by the DeliverTx ABCI message. If the transaction is relayed by CheckTx, only stateless checks and fee-related (i.e. auth module-related) stateful checks are performed. To better understand the difference between DeliverTxand CheckTx, as well as the difference between stateful and stateless checks, click here.

The handler of a module is generally defined in a file called handler.go and consists of:

• A switch function NewHandler to route the message to the appropriate handler function. This function returns a handler function, and is registered in the AppModule to be used in the application's module manager to initialize the application's router. See an example of such a switch here.
• One handler function for each message type defined by the module. Developers write the message processing logic in these functions. This generally involves doing stateful checks to ensure the message is valid and calling keeper's methods to update the state.

Handler functions return a result of type sdk.Result, which informs the application on whether the message was successfully processed.

To learn more about handlers, click here.

Keeper

Keepers are the gatekeepers of their module's store(s). To read or write in a module's store, it is mandatory to go through one of its keeper's methods. This is ensured by the object-capabilities model of the Cosmos SDK. Only objects that hold the key to a store can access it, and only the module's keeper should hold the key(s) to the module's store(s).

Keepers are generally defined in a file called keeper.go. It contains the keeper's type definition and methods.

The keeper type definition generally consists of:

• Key(s) to the module's store(s) in the multistore.
• Reference to other module's keepers. Only needed if the keeper needs to access other module's store(s) (either to read or write from them).
• A reference to the application's codec. The keeper needs it to marshal structs before storing them, or to unmarshal them when it retrieves them, because stores only accept []bytes as value.

Along with the type definition, the next important component of the keeper.go file is the keeper's constructor function, NewKeeper. This function instantiates a new keeper of the type defined above, with a codec, store keys and potentially references to other modules' keepers as parameters. The NewKeeper function is called from the application's constructor.

The rest of the file defines the keeper's methods, primarily getters and setters. You can check an example of a keeper implementation here.

To learn more about keepers, click here.

Querier

Queriers are very similar to handlers, except they serve user queries to the state as opposed to processing transactions. A query is initiated from an interface by an end-user who provides a queryRoute and some data. The query is then routed to the correct application's querier by baseapp's handleQueryCustom method using queryRoute.

The Querier of a module is defined in a file called querier.go, and consists of:

• A switch function NewQuerier to route the query to the appropriate querier function. This function returns a querier function, and is is registered in the AppModule to be used in the application's module manager to initialize the application's query router. See an example of such a switch here.
• • One querier function for each data type defined by the module that needs to be queryable. Developers write the query processing logic in these functions. This generally involves calling keeper's methods to query the state and marshalling it to JSON. See an example of querier functions here.

To learn more about queriers, click here.

Command-Line and REST Interfaces

Each module defines command-line commands and REST routes to be exposed to end-user via the application's interfaces. This enables end-users to create messages of the types defined in the module, or to query the subset of the state managed by the module.

CLI

Generally, the commands related to a module are defined in a folder called client/cli in the module's folder. The CLI divides commands in two category, transactions and queries, defined in client/cli/tx.go and client/cli/query.go respectively. Both commands are built on top of the Cobra Library:

• Transactions commands let users generate new transactions so that they can be included in a block and eventually update the state. One command should be created for each message type defined in the module. The command calls the constructor of the message with the parameters provided by the end-user, and wraps it into a transaction. The SDK handles signing and the addition of other transaction metadata. See examples of transactions commands here.
• Queries let users query the subset of the state defined by the module. Query commands forward queries to the application's query router, which routes them to the appropriate querier the queryRoute parameter supplied. See examples of query commands here.

To learn more about modules CLI, click here.

REST

The module's REST interface lets users generate transactions and query the state through REST calls to the application's light client daemon (LCD). REST routes are defined in a file client/rest/rest.go, which is composed of:

• A RegisterRoutes function, which registers each route defined in the file. This function is called from the main application's interface for each module used within the application. The router used in the SDK is Gorilla's mux.
• Custom request type definitions for each query or transaction creation function that needs to be exposed. These custom request types build on the base request type of the Cosmos SDK.
• One handler function for each request that can be routed to the given module. These functions implement the core logic necessary to serve the request.

See an example of a module's rest.go file here.

To learn more about modules REST interface, click here.

Application Interface

Interfaces let end-users interact with full-node clients. This means querying data from the full-node or creating and sending new transactions to be relayed by the full-node and eventually included in a block.

The main interface is the Command-Line Interface. The CLI of an SDK application is built by aggregating CLI commands defined in each of the modules used by the application. The CLI of an application generally has the -cli suffix (e.g. appcli), and defined in a file called cmd/appcli/main.go. The file contains:

• A main() function, which is executed to build the appcli interface client. This function prepares each command and adds them to the rootCmd before building them. At the root of appCli, the function adds generic commands like status, keys and config, query commands, tx commands and rest-server.
• Query commands are added by calling the queryCmd function, also defined in appcli/main.go. This function returns a Cobra command that contains the query commands defined in each of the application's modules (passed as an array of sdk.ModuleClients from the main() function), as well as some other lower level query commands such as block or validator queries. Query command are called by using the command appcli query [query] of the CLI.
• Transaction commands are added by calling the txCmd function. Similar to queryCmd, the function returns a Cobra command that contains the tx commands defined in each of the application's modules, as well as lower level tx commands like transaction signing or broadcasting. Tx commands are called by using the command appcli tx [tx] of the CLI.
• A registerRoutes function, which is called from the main() function when initializing the application's light-client daemon (LCD) (i.e. rest-server). registerRoutes calls the RegisterRoutes function of each of the application's module, thereby registering the routes of the module to the lcd's router. The LCD can be started by running the following command appcli rest-server.

See an example of an application's main command-line file here.

To learn more about interfaces, click here.

Dependencies and Makefile

This section is optional, as developers are free to choose their depencency manager and project building method. That said, the current most used framework for versioning control is go.mod. It ensures each of the libraries used throughout the application are imported with the correct version. An example can be found here.

For building the application, a Makefile is generally used. The Makefile primarily ensures that the go.mod is run before building the two entrypoints to the application, appd and appcli. An example of Makefile can be found here.

Next

Learn more about the Lifecycle of a transaction.