This repository is a tutorial for building a complete application on top of the Tendermint ABCI, implementing a key-value store with a compare-and-swap API.
- The goal
- Tendermint v Cosmos
- Tendermint concepts
- ABCI methods
- Our demo application
- The abci-cli
- System architecture
- Operations
- Building and running
We're going to build a distributed system in Go. Each node will have an HTTP API implementing compare-and-swap semantics for a key-value store. The keys and values will be reliably and consistently replicated between nodes by Tendermint.
$ curl -Ss -XPOST 'http://localhost:10001/x?new=foo'
{
"key": "x",
"value: "foo"
}
$ curl -Ss -XPOST 'http://localhost:10002/x?old=foo&new=bar'
{
"key": "x",
"value: "bar"
}
$ curl -Ss -XGET 'http://localhost:10003/key'
{
"key": "x",
"value: "bar"
}
Hopefully, this exercise will expose you to enough of the Tendermint programming model that you can confidently build your own Tendermint ABCI applications.
Tendermint is a distributed, byzantine fault-tolerant consensus system designed to replicate arbitrary state machines. You can build on top of Tendermint by plugging into it at different levels of abstraction, depending on what kind of application you're building.
At the lowest level, Tendermint defines an API, called the Application BlockChain Interface, or ABCI. Applications that implement the ABCI can be replicated with the Tendermint protocol.
One level above Tendermint is Cosmos, a federated network of blockchains; or, more accurately, the Cosmos SDK, which allows users to build Cosmos-compatible applications. The Basecoin demo application is built on the Cosmos SDK.
+------------+
| Basecoin |
+------------+ +------------+
| Cosmos SDK | | This repo |
+------------+-+------------+
| ABCI |
+---------------------------+
| Tendermint |
+---------------------------+
A picture is worth a thousand words; this diagram provides a great conceptual overview of all of the interacting components in a Tendermint network. In the full node, we're going to implement the blue box titled ABCI App. To do that, we need to implement the ABCI interface.
type Application interface {
Info(RequestInfo) ResponseInfo
SetOption(RequestSetOption) ResponseSetOption
Query(RequestQuery) ResponseQuery
CheckTx(tx []byte) ResponseCheckTx
InitChain(RequestInitChain) ResponseInitChain
BeginBlock(RequestBeginBlock) ResponseBeginBlock
DeliverTx(tx []byte) ResponseDeliverTx
EndBlock(RequestEndBlock) ResponseEndBlock
Commit() ResponseCommit
}Understanding these methods requires understanding the Tendermint state machine. Let's understand things at a high level first, and then describe each method in detail.
It's expected that your application wraps some state, which Tendermint transactions (Tx) manipulate. Furthermore, it's expected that your application state has a notion of a commit, which should
- Count the number of commits made
- Persist the state to long-term storage
- Hash the complete state at the time of commit
Tendermint takes care of delivering transactions to your application via DeliverTx. Those transactions are guaranteed to come in the same order to all instances of your application, on all nodes in the network. Each transaction must have the same deterministic effect on application state on all nodes.
Transactions are bundled into blocks, demarcated by BeginBlock and EndBlock calls. One block will contain zero or more transactions, delivered via DeliverTx. After EndBlock, Tendermint will always call Commit, which should trigger the commit steps enumerated above. Tendermint is guaranteed to call (BeginBlock, DeliverTx, DeliverTx, ..., EndBlock, Commit) in exactly the same order, with exactly the same data, on all nodes.
All changes to state must occur via DeliverTx exclusively.
The Tendermint ABCI method Query is used to read application state. Your application has a lot of leeway to decide how to create, interpret, and service Query requests. The only thing that Tendermint stipulates is that some query paths (a string field in the query request) are reserved. Otherwise, the only contract is that queries must not mutate state.
We've talked about how state is written to and read from. But how is the state machine itself initialized?
The abstract, global state machine replicated by Tendermint is created in an event known as genesis. Genesis involves creating a chain ID, which uniquely identifies the state machine, as well as other parameters, like the initial set of participating nodes. This configuration information is collected into a genesis file, which must be securely distributed to each initial node in the Tendermint network. All nodes must share exactly the same genesis file.
The concrete, specific state machine instance managed by a given Tendermint node is created at process start. If the node is starting for the first time, it will have an empty application state; if the node has e.g. rebooted, it should load its application state from persistent storage into memory.
Three ABCI methods manage state machine initialization.
InitChain is called once, when a node starts for the first time. It tells the application about some aspects of the state machine, or chain, from Tendermint's perspective, including the chain ID, consensus parameters, and any initial application state that's been provided by the network operator. The application can use this information to make itself ready to receive transactions.
SetOption may be called to set arbitrary application configuration parameters. This is only done if by user request, and Tendermint doesn't interpret these commands, or route them through its consensus system. This means that any changes made via SetOption may be different on different nodes, and therefore must not have any effect on how transactions or queries are processed. This is sometimes referred to as being non-consensus-critical or non-deterministic.
Info is called at each process start, after the application has restored any application state from persistent storage, so that Tendermint can know the last block height (a.k.a. the commit count) and app state hash of the application. Tendermint will calculate the diff between what your application reports in Info, and what Tendermint knows the current state of the global state machine to be, and will replay blocks of transactions to your application, until the block height and app hash match.
Tendermint speaks to your application exclusively through the ABCI interface, but it does so through three independent connections. Calls are serialized on each connection, but may be concurrent across different connections. Each connection only calls a subset of the ABCI methods.
The query connection is responsible for read operations by calling Query. It also handles initialization, calling InitChain, SetOption, and Info.
The consensus connection is responsible for write operations, calling BeginBlock, DeliverTx, EndBlock, and Commit.
There is a third connection which introduces a new concept to your application state. When a transaction arrives at a Tendermint node, before it's given to the consensus machinery and replicated to the rest of the network, it's first sent to the local application, over the mempool connection, to the CheckTx method. This is ultimately an optimization step, giving the application the opportunity to validate the transaction (for example, checking that the transaction body is properly encoded) and stop invalid transactions before they're broadcast. If the application needs to implement replay protection (for example, to protect against double-spend attacks) it should also perform that accounting in CheckTx.
To support CheckTx and the mempool connection, it's recommended that applications actually keep two separate in-memory representations of their state: the consensus state, updated by DeliverTx; and the mempool state, updated by CheckTx. The mempool state should be updated by CheckTx transactions in the same way the consensus state is updated by DeliverTx transactions, with one important difference: when the consensus state is committed by Commit, it should be copied to and fully overwrite the mempool state. This is because Tendermint may deliver the same transaction via CheckTx more than once, though it will only do so if that transaction is checked but doesn't make it in to the consensus block.
To be clear, this step is optional. Applications may choose to skip managing a separate mempool state, and simply return an OK result for every CheckTx call. This should not affect correctness, only efficiency.
All user requests must be routed to the application through Tendermint via Tendermint's RPC mechanism. Requests must never hit the application or its state directly. To make RPC requests through Tendermint to our application, we use an RPC client. The relevant part of that interface is ABCIClient.
type ABCIClient interface {
ABCIInfo() (*ctypes.ResultABCIInfo, error)
ABCIQuery(path string, data cmn.HexBytes) (*ctypes.ResultABCIQuery, error)
ABCIQueryWithOptions(path string, data cmn.HexBytes, opts ABCIQueryOptions) (*ctypes.ResultABCIQuery, error)
BroadcastTxCommit(tx types.Tx) (*ctypes.ResultBroadcastTxCommit, error)
BroadcastTxAsync(tx types.Tx) (*ctypes.ResultBroadcastTx, error)
BroadcastTxSync(tx types.Tx) (*ctypes.ResultBroadcastTx, error)
}We never need to implement an ABCIClient, we just need to construct one. That client should be taken as a dependency to our user-facing API, and all user requests should be proxied through it. Reads should go through ABCIQuery, and writes should go through one of the BroadcastTx methods. The difference between those methods relates to how long they block before returning a result. BroadcastTxCommit blocks the longest, and waits until the transaction has been committed into a block by a quorum of nodes in the network. BroadcastTxSync waits until the transaction has been accepted by the consensus connection, but not necessarily committed. BroadcastTxAsync only waits until the Tendermint machinery has received the transaction, and returns before it's been received by any node.
Now that we have a high-level understanding of Tendermint, let's get into detail about each ABCI method.
RequestInfo
- Version: The version of Tendermint, e.g. "0.25.0".
ResponseInfo
- Data: An arbitrary string containing information about the application, not parsed by Tendermint. Optional.
- Version: An arbitrary string containing the version of the application, used in the Tendermint version handshake. Optional.
- LastBlockHeight: The height of the blockchain (number of commits) on this node. Taken from persisted consensus state. Required.
- LastBlockAppHash: The SHA256 hash of the last committed application state on this node. Taken from persisted consensus state. Required.
See Initialization.
RequestSetOption
- Key: An arbitrary string defining the option key.
- Value: An arbitrary string defining the option value.
ResponseSetOption
- Code: Response code; zero for OK, non-zero for error. Required.
- Log: Arbitrary string containing non-deterministic data intended for literal output via the application's logger. Optional.
- Info: Arbitrary string containing non-deterministic data in addition to log. Optional.
See Initialization.
RequestInitChain
- Time: The timestamp in the genesis file.
- ChainId: The chain ID string in the genesis file.
- ConsensusParams: Parameters that govern Tendermint's consensus behavior.
- Validators: The current set of validator nodes in the network.
- AppStateBytes: Initial state, provided in the genesis file, that a node starting for the first time may need to make itself ready to receive transactions.
ResponseInitChain
- ConsensusParams: Any changes to the proposed consensus parameters that this node would like to propose. Optional.
- Validators: Any changes to the set of validators that this node would like to propose. Optional.
See Initialization. ConsensusParams and Validators are beyond the scope of this document, see the official documentation for details.
RequestQuery
- Data: The byte array from the user request.
- Path: The path string from the user request.
- Height: The desired height of the blockchain (in effect, the version of the state) against which the query should be run. A height of zero means the most recent state. To support this parameter, state needs to be implemented using a version-aware data structure, e.g. this IAVL tree.
- Prove: If true, include a Merkle proof of the query results in the response.
ResponseQuery
- Code: Response code; zero for OK, non-zero for error. Required.
- Log: Arbitrary string containing non-deterministic data intended for literal output via the application's logger. Optional.
- Info: Arbitrary string containing non-deterministic data in addition to log. Optional.
- Index: Related to the Merkle proof, if requested. Optional.
- Key: A byte array containing the key that's returned. Optional.
- Value: A byte array containing the data of the query response. Optional.
- Proof: A byte array containing the Merkle proof of the query, if requested. Optional.
- Height: The height of the blockchain (in effect, the version of the state) against which the query was run. Optional.
See Reading application state and Connections. Note that our demo application doesn't implement Height, Prove, or Proof. Merkle proofs are beyond the scope of this document, see the official documentation for details.
Query will probably want to read consensus state, for the most reliable and up-to-date view of the world. In some cases, it may want to read committed state, for example if a application-specific flag is defined in the query body, in order to return data that is guaranteed to be persistent in case of node failure. Or, it may want to read mempool state, to yield the most bleeding-edge version of events, with some risk of that state being rendered invalid in the future. These are all application decisions.
RequestBeginBlock
- Hash: The hash of the block.
- Header: The header details for the block.
- LastCommitInfo: Details about the most recent (previous) commit.
- ByzantineValidators: Evidence of malicious validators, if any, during the most recent (previous) commit.
ResponseBeginBlock
- Tags: A set of key-value pairs that can be used to denote properties about this block, which can later be searched. Optional.
See Writing application state. See also Application Development Guide: BeginBlock.
The only argument is an opaque byte slice, proxied without modification from the RPC connection's BroadcastTx methods to the application.
ResponseCheckTx
- Code: Response code; zero for OK, non-zero for error. Required.
- Data: Arbitrary byte array containing any result from the transaction. Optional.
- Log: Arbitrary string containing non-deterministic data intended for literal output via the application's logger. Optional.
- Info: Arbitrary string containing non-deterministic data in addition to log. Optional.
- GasWanted: Amount of gas request for the transaction. Optional.
- GasUsed: Amount of gas consumed by the transaction. Optional.
- Tags: A set of key-value pairs that can be used to denote properties about this transaction, which can later be searched. Optional.
See Connections. See also Mempool Connection. Observe that CheckTx has exactly the same signature as DeliverTx; the only difference is how to interpret the transaction body, i.e. which state (if any) to update.
The only argument is an opaque byte slice, proxied without modification from the RPC connection BroadcastTx methods to the application.
ResponseDeliverTx
- Code: Response code; zero for OK, non-zero for error. Required.
- Data: Arbitrary byte array containing any result from the transaction. Optional.
- Log: Arbitrary string containing non-deterministic data intended for literal output via the application's logger. Optional.
- Info: Arbitrary string containing non-deterministic data in addition to log. Optional.
- GasWanted: Amount of gas request for the transaction. Optional.
- GasUsed: Amount of gas consumed by the transaction. Optional.
- Tags: A set of key-value pairs that can be used to denote properties about this transaction, which can later be searched. Optional.
See Writing application state and Connections. See also DeliverTx. Observe that DeliverTx has exactly the same signature as CheckTx; the only difference is how to interpret the transaction body, i.e. which state (if any) to update.
RequestEndBlock
- Height: The height of the block.
ResponseEndBlock
- ValidatorUpdates: Updates to the set of validators, if any. Optional.
- ConsensusParamsUpdate: Updates to the consensus parameters, if any. Optional.
- Tags: A set of key-value pairs that can be used to denote properties about this block, which can later be searched. Optional.
See Writing application state and Connections. See also EndBlock.
Commit requests have no parameters.
ResponseCommit
- Data: A deterministic (Merkle) hash of the state root of the application. Required.
See Writing application state and Connections. See also Commit. It's expected that the application persist its state to disk during commit.
The code for the ABCI application, implementing our key-value store with compare-and-swap semantics, is available in internal/cas/application.go. The code for the state layer is available in internal/cas/state.go.
Once you have a type implementing the ABCI application interface, you can do
basic tests by wrapping it with an abci/server.NewServer and
calling it with a tool called abci-cli.
The code to mount your application will look something like this.
func main() {
app := newMyApplication()
server, err := server.NewServer("127.0.0.1:8080", "socket", app)
if err != nil {
log.Fatal(err)
}
server.Start()
}See the abci-cli documentation for more details.
Now you have a working ABCI application. How can you connect it to the Tendermint machinery, so that it can communicate with other, identical nodes in a network?
Recall the original Tendermint architecture diagram. The Tendermint node, the green box, handles the heavy work of consensus. The validator signer, the purple box, can also be provided by Tendermint, and validates blocks, moving consensus forward. Our ABCI application, the blue box, is connected to the node exclusively. And our user API, in the diagram represented by Cosmos Voyager and the Light Client Daemon, is also connected only to the node.
These are the logical components, and they can be deployed in different physical arrangements depending on your needs. In the diagram, the Tendermint core node, the ABCI application, and the validator signer are all co-located on the same circle, representing a full node. The lines between them indicate communication, but that communication can occur in different ways. The components can be built into the same binary, executed as a single process, and the communication occur exclusively in-memory. Or, the components can be built into separate binaries, executed as different processes on the same machine, and the communication occur over e.g. UNIX domain sockets. Or, the components could be deployed to different physical machines, and the communication occur over e.g. TCP connections.
Similarly, in the diagram, the user is expected to interact with the network by using the Cosmos Voyager web application, deployed adjacent to a Light Client Daemon on the user's machine, speaking REST (HTTP) to each other. The Light Client Daemon connects to a Tendermint core node over HTTP, and performs the e.g. ABCIQuery and BroadcastTx requests that way.
The diagram has things arranged in this way, but you don't necessarily need to copy it. For example, a more secure deployment might move the validators onto their own single-purpose machines, heavily firewalled from the rest of the internet, with only a single connection to a different, larger set of full nodes, running only Tendermint core and your ABCI application.
For our demo, we'll model the user API as a separate HTTP API, but built into the same binary as all the other components, and run in the same process. The HTTP API is defined in cmd/tendermint-cas-demo/cas_api.go, and all of the components are wired together in cmd/tendermint-cas-demo/main.go.
The Tendermint node requires quite a lot of configuration to successfully start, including several files in well-defined locations on disk, such as the JSON genesis file, a TOML configuration file, and cryptographic keys for the node itself and its validators. The helper scripts bootstrap_1 and bootstrap_3 create these file structures for one- and three-node clusters on the local machine respectively. Studying them should give you a good start toward scripting your own deployment.
Building this repository requires a working Go installation. Most operating system package managers ship a reasonably up-to-date version of the Go development environment. If you're on a Mac and using Homebrew, I recommend
$ brew install go
Clone this repo into the correct location in your GOPATH.
$ mkdir -p $(go env GOPATH)/src/github.com/6thc
$ cd $(go env GOPATH)/src/github.com/6thc
$ git clone [email protected]:6thc/tendermint-cas-demo
$ cd tendermint-cas-demo
Then, build the binary.
$ make
$ ./tendermint-cas-demo -h
USAGE
tendermint-cas-demo [flags]
FLAGS
-api-addr 127.0.0.1:8081 HTTP API address
-app-file db.json application persistence file
-app-verbose false verbose logging of application information
-tendermint-dir tendermint Tendermint directory (config, data, etc.)
-tendermint-verbose false verbose logging of Tendermint information
You can also use the Makefile to bootstrap a 1- or 3-node-cluster on your local machine. Once everything is set up, it will print instructions on how to start the cluster.
$ make bootstrap_3
downloading tendermint_0.25.0_darwin_amd64.zip...
Archive: tendermint_0.25.0_darwin_amd64.zip
tendermint_0c9c3292c918617624f6f3fbcd95eceade18bcd5_darwin_amd64
extracting: tendermint
clearing any old state...
initializing three nodes...
capturing validators...
capturing peer addresses...
building a common genesis file...
writing common genesis file...
producing config files...
now you can run three nodes
./tendermint-cas-demo -api-addr 127.0.0.1:8081 -app-file a.json -tendermint-dir tendermint_a
./tendermint-cas-demo -api-addr 127.0.0.1:8082 -app-file b.json -tendermint-dir tendermint_b
./tendermint-cas-demo -api-addr 127.0.0.1:8083 -app-file c.json -tendermint-dir tendermint_c
other fun things to try
watch -n1 -- cat ?.json # watch state being updated
curl -Ss -XPOST 'localhost:8081/x?new=one' # set x=one
curl -Ss -XPOST 'localhost:8082/x?old=one&new=two' # set x=two
curl -Ss -XGET 'localhost:8083/x' # get x