One goal of a distributed ledger like Sawtooth, indeed the defining goal, is to distribute a ledger among participating nodes. The ability to ensure a consistent copy of data amongst nodes in Byzantine consensus is one of the core strengths of blockchain technology.
Sawtooth represents state for all transaction families in a single instance of a Merkle-Radix tree on each validator. The process of block validation on each validator ensures that the same transactions result in the same state transitions and that the resulting data is the same for all participants in the network.
The state is split into namespaces which allow flexibility for transaction family authors to define, share, and reuse global state data between transaction processors.
Merkle-Radix Tree Overview
Sawtooth uses an addressable Merkle-Radix tree to store data for transaction families. Let's break that down: The tree is a Merkle tree because it is a copy-on-write data structure which stores successive node hashes from leaf-to-root upon any changes to the tree. For a given set of state transitions associated with a block, we can generate a single root hash which points to that version of the tree. By placing this state root hash on the block header, we can gain consensus on the expected version of state in addition to the consensus on the chain of blocks. If a validator's state transitions for a block result in a different hash, the block is not considered valid. For more information about general concepts, see the Merkle page on Wikipedia.
The tree is an addressable Radix tree because addresses uniquely identify the paths to leaf nodes in the tree where information is stored. An address is a hex-encoded 70 character string representing 35 bytes. In the tree implementation, each byte is a Radix path segment which identifies the next node in the path to the leaf containing the data associated with the address. The address format contains a 3 byte (6 hex character) namespace prefix which provides 2^24^ (16,777,216) possible different namespaces in a given instance of Sawtooth. The remaining 32 bytes (64 hex characters) are encoded based on the specifications of the designer of the namespace, and may include schemes for subdividing further, distinguishing object types, and mapping domain-specific unique identifiers into portions of the address. For more information about general concepts, see the Radix page on Wikipedia.
In addition to questions regarding the encoding of addresses, namespace designers also need to define the mechanism of serialization and the rules for serializing/deserializing the data stored at addresses. The domain-specific Transaction Processor makes get(address) and set(address, data) calls against a version of state that the validator provides. get(address) returns the byte array found at that address and set(address, data) sets the byte array stored at that address. The byte array is opaque to the core system. It only has meaning when deserialized by a domain-specific component based on the rules of the namespace. It is critical to select a serialization scheme which is deterministic across executions of the transaction, across platforms, and across versions of the serialization framework. Data structures which don't enforce ordered serialization (e.g. sets, maps, dicts) should be avoided. The requirement is to consistently produce the same byte array across space and time. If the same byte array is not produced, the leaf node hash containing the data will differ, as will every parent node back to the root. This will result in transactions and the blocks that contain them being considered valid on some validators and invalid on others, depending on the non-deterministic behavior. This is considered bad form.