12 KiB

Raw Blame History

Glossary

This defines many of the terms that are used in the other documents. If there is every a concept that seems unclear, check here. This is mainly to provide a background and general understanding of the different words and concepts that are used. Other documents will explain in more detail how to combine these concepts to build a particular application.

Transaction

A transaction is a packet of binary data that contains all information to validate and perform an action on the blockchain. The only other data that it interacts with is the current state of the chain (kv store), and it must have a deterministic action. The transaction is the main piece of one request.

We currently make heavy use of go-wire and go-data to provide automatic binary and json encodings (and decodings) for objects, even when they embed many interfaces inside. There is one public TxMapper in the basecoin root package, and all modules can register their own transaction types there. This allows us to deserialize the entire tx in one location (even with types defined in other repos), to easily embed an arbitrary Tx inside another without specifying the specific type, and provide an automatic json representation to provide to users (or apps) to inspect the chain.

Note how we can wrap any other transaction, add a fee level, and not worry about the encoding in our module any more?

type Fee struct {
  Fee   coin.Coin      `json:"fee"`
  Payer basecoin.Actor `json:"payer"` // the address who pays the fee
  Tx    basecoin.Tx    `json:"tx"`
}

Context

As the request passes through the system, it can pick up information, that must be carried along with it. Like the authorized it has received from another middleware, or the block height it runs at. This is all deterministic information from the context in which the request runs (based on the tx and the block it was included in) and can be used to validate the tx.

Data Store

To be able to provide proofs to tendermint, we keep all data in one key-value store, indexed with a merkle tree. This allows us to easily provide a root hash and proofs for queries without requiring complex logic inside each module. Standarizing this also allows powerful light-client tooling as it knows how to verify all data in the store.

The downside is there is one quite simple interface that the application has to Get and Set data. There is not even a range query. Although there are some data structures like queues and range queries that are also in the state package to provide higher-level functionality in a standard format.

Isolation

One of the main arguments for blockchain is security. So while we encourage the use of third-party modules, we must be vigilant against security holes. If you use the stack package, it will provide two different types of sandboxing for you.

The first step, is that when DeliverTx is called on a module, it is never given the entire data store, but rather only its own prefixed section. This is achieved by prefixing all keys transparently with <module name> + 0x0, using the null byte as a separator. Since module name must be a string, no clever naming scheme can lead to a collision. Inside the module, we can write anywhere we want, without worry that we have to touch some data that is not ours.

The second step involves the permissions in the context. The context can say that this tx was signed by eg. Rigel. But if any module can add that permission, it would be too easy to forge accounts. Thus, each permission is associated with the module that granted it (in this case auth), and if a module tries to add a permission for another module, it will panic. There is also protection if a module creates a brand new fake context to trick the downstream modules.

This means that modules can confidently write to their local section of the database and trust the permissions associated with the context, without concern of interferance from other modules. (Okay, if you see a bunch of C-code in the module traversing through all the memory space of the application, then get worried....)

Handler

The ABCI interface is handled by app, which translates these data structures into an internal format that is more convenient, but unable to travel over the wire. The basic interface for any code that modifies state is the Handler interface, which provides four methods:

  Name() string
  CheckTx(ctx Context, store state.KVStore, tx Tx) (Result, error)
  DeliverTx(ctx Context, store state.KVStore, tx Tx) (Result, error)
  SetOption(l log.Logger, store state.KVStore, module, key, value string) (string, error)

Note the Context, Store, and Tx as principal carriers of information. And that Result is always success, and we have a second error return for errors (which is much more standard go that res.IsErr())

The Handler interface is designed to be the basis for all modules that execute transaction, and this can provide a large degree of code interoperability, much like http.Handler does in golang web development.

Middleware

Middleware is a series of processing steps that any request must travel through before (and after) executing the registered Handler. Some examples are a logger (that records the time before executing the tx, then outputs info - including duration - after the execution), of a signature checker (which unwraps the tx by one layer, verifies signatutes, and adds the permissions to the Context before passing the request along).

In keeping with the standardazation of http.Handler and inspired by the super minimal negroni package, we just provide one more Middleware interface, which has an extra next parameter, and a Stack that can wire all the levels together (which also gives us a place to perform isolation of each step).

  Name() string
  CheckTx(ctx Context, store state.KVStore, tx Tx, next Checker) (Result, error)
  DeliverTx(ctx Context, store state.KVStore, tx Tx, next Deliver) (Result, error)
  SetOption(l log.Logger, store state.KVStore, module, key, value string, next Optioner) (string, error)

Modules

A module is a set of functionality that is more or less self-sufficient. It usually contains the following pieces:

transaction types (either end transactions, or transaction wrappers)
custom error codes
data models (to persist in the kv store)
handler (to handle any end transactions)
middleware (to handler any wrapper transactions)

To enable a module, you must add the appropriate middleware (if any) to the stack in main.go, as well as adding the handler (if any) to the dispatcher. One the stack is compiled into a Handler, then all tx are handled by the proper module.

Dispatcher

We usually will want to have multiple modules working together, and need to make sure the correct transactions get to the correct module. So we have have coin sending money, roles creating multi-sig accounts, and ibc following other chains all working together without interference.

After the chain of middleware, we can register a Dispatcher, which also implements the Handler interface. We then register a list of modules with the dispatcher. Every module has a unique Name(), which is used for isolating its state space. We use this same name for routing tx. Each tx implementation must be registed with go-wire via TxMapper, so we just look at the registered name of this tx, which should be of the form <module name>/xxx. The dispatcher grabs the appropriate module name from the tx name and routes it if the module is present.

This all seems a bit of magic, but really just making use of the other magic (go-wire) that we are already using, rather than add another layer. The only thing you need to remember is to use the following pattern, then all the tx will be properly routed:

const (
  NameCoin = "coin"
  TypeSend = NameCoin + "/send"
)

IPC (Inter-Plugin Communication)

But wait, there's more... since we have isolated all the modules from each other, we need to allow some way for them to interact in a controlled fashion. Some examples are the fee middleware, which wants to deduct coins from the calling account (in the coin module), or a vote that requires a payment.

If we want to make a call from the middleware, this is relatively simple. The middleware already has a handle to the next Handler, which will execute the rest of the stack. It can simple create a new SendTx and pass it down the stack. If it returns success, then do the rest of the processing (and send the original tx down the stack), otherwise abort.

However, if one Handler inside the Dispatcher wants to do this, it becomes more complex. The solution is that the Dispatcher accepts not a Handler, but a Dispatchable, which looks like a middleware, except that the next argument is a callback to the dispatcher to execute a sub-transaction. If a module doesn't want to use this functionality, it can just implement Handler and call stack.WrapHandler(h) to convert it to a Dispatchable that never uses the callback.

One example of this is the counter app, which can optionally accept a payment. If the tx contains a payment, it must create a SendTx and pass this to the dispatcher to deduct the amount from the proper account. Take a look at counter plugin for a better idea.

Permissions

This system requires a more complex permissioning system to allow the modules to have limited access to each other. Also to allow more types of permissions than simple public key signatures. So, rather than just use an address to identify who is performing an action, we can use a more complex structure:

type Actor struct {
  ChainID string     `json:"chain"` // this is empty unless it comes from a different chain
  App     string     `json:"app"`   // the app that the actor belongs to
  Address data.Bytes `json:"addr"`  // arbitrary app-specific unique id
}

ChainID is to be used for IBC, which is discussed below, but right now focus on App and Address. For a signature, the App is auth, and any modules can check to see if a specific public key address signed like this ctx.HasPermission(auth.SigPerm(addr)). However, we can also authorize a tx with roles, which handles multi-sig accounts, it checks if there were enough signatures by checking as above, then it can add the role permission like ctx = ctx.WithPermissions(NewPerm(assume.Role))

In addition to permissioning, the Actors are addresses just like public key addresses. So one can create a mulit-sig role, then send coin there, which can only be moved upon meeting the authorization requirements from that module. coin doesn't even know the existence of roles and one could build any other sort of module to provide permissions (like bind the outcome of an election to move coins or to modify the accounts on a role).

One idea (not implemented) is to provide scopes on the permissions. Right now, if I sign a tx to one module, it can pass it on to any other module over IPC with the same permissions. It could move coins, vote in an election, or anything else. Ideally, when signing, one could also specify the scope(s) that this signature authorizes. The oauth protocol also has to deal with a similar problem, and maybe could provide some inspiration.

Replay Protection

Is implemented as middleware. Rigel can add more info here. Or look at the github issue

IBC (Inter-Blockchain Communication)

Wow, this is a big topic. Also a WIP. Add more here...

12 KiB Raw Blame History