| title |
|---|
Object and Package Versioning |
You reference every object stored on chain by an ID and version. When a transaction modifies an object, it writes the new contents to an on-chain reference with the same ID but a later version. This means that a single object (with ID I) might appear in multiple entries in the distributed store:
(I, v0) => ...
(I, v1) => ... # v0 < v1
(I, v2) => ... # v1 < v2
Despite appearing multiple times in the store, only one version of the object is available to transactions -- the latest version (v2 in the previous example) -- and only one transaction can modify the object at that version to create a new version, guaranteeing a linear history (v1 was created in a state where I was at v0, and v2 was created in a state where I was at v1).
Versions are strictly increasing and (ID, version) pairs are never re-used. This structure allows node operators to prune their stores of old object versions that are now inaccessible, if they choose. This is not a requirement, though, as node operators might keep prior object versions around to serve requests for an object's history, either from other nodes that are catching up, or from RPC requests.
Sui uses Lamport timestamps in its versioning algorithm for objects. The use of Lamport timestamps guarantees that versions never get re-used as the new version for objects touched by a transaction is one greater than the max version among all input objects to the transaction. For example, a transaction transferring an object O at version 5 using a gas object G at version 3 updates both O and G versions to 1 + max(5, 3) = 6.
The relevance of Lamport versions for maintaining the "no (ID, version) re-use" invariant or for accessing an object as a transaction input changes depending on that object's ownership, as detailed in the following sections.
You must reference address-owned transaction inputs at a specific ID and version. When a validator signs a transaction with an owned object input at a specific version, that version of the object is locked to that transaction. Validators reject requests to sign other transactions that require the same input (same ID and version).
If F + 1 validators sign one transaction that takes an object as input, and a different F + 1 validators sign a different transaction that takes the same object as input, that object (and all the other inputs to both transactions) are equivocated meaning they cannot be used for any further transactions in that epoch. This is because neither transaction can form a quorum without relying on a signature from a validator that has already committed the object to a different transaction, which it cannot get. All locks are reset at the end of the epoch, which frees the objects again.
Only an object's owner can equivocate it, but this is not a desirable thing to do. You can avoid equivocation by carefully managing the versions of address-owned input objects.
Like address-owned objects, you reference immutable objects at an ID and version, but they do not need to be locked as their contents and versions do not change. Their version is relevant because they could have started life as an address-owned object before being frozen. The given version identifies the point at which they became immutable.
Specifying a shared transaction input is slightly more complex. You reference it by its ID, the version it was shared at, and a flag indicating whether it is accessed mutably. You don't specify the precise version the transaction accesses because it is decided by consensus during transaction scheduling. When scheduling multiple transactions that touch the same shared object, validators agree the order of those transactions, and pick each transaction's input versions for the shared object accordingly (one transaction's output version becomes the next transaction's input version, and so on).
Shared transaction inputs that you reference immutably participate in scheduling, but don't modify the object or increment its version.
The make_wrapped function in the following example creates an Inner object, wrapped in an Outer object, which is sent back to the transaction sender.
module example::wrapped {
use sui::object::{Self, UID};
use sui::transfer;
use sui::tx_context::{Self, TxContext};
struct Inner has key, store {
id: UID,
x: u64,
}
struct Outer has key {
id: UID,
inner: Inner,
}
entry fun make_wrapped(ctx: &mut TxContext) {
let inner = Inner {
id: object::new(ctx),
x: 42,
};
let outer = Outer {
id: object::new(ctx),
inner,
};
transfer::transfer(outer, tx_context::sender(ctx));
}
}
You can't access wrapped objects by their ID in the object store, you must access them by the object that wraps them. In the prior example, the owner of Outer must specify it as the transaction input and then access its inner field to read the instance of Inner. Validators refuse to sign transactions that specify wrapped objects (like the inner of an Outer) as inputs. As a result, you don't need to specify a wrapped object's version in a transaction that reads that object.
Wrapped objects can eventually become "unwrapped", meaning that they are once again accessible at their ID:
module example::wrapped {
// ...
entry fun unwrap(outer: Outer, ctx: &TxContext) {
let Outer { id, inner } = outer;
object::delete(id);
transfer::transfer(inner, tx_context::sender(ctx));
}
}
The unwrap function in the previous code takes an instance of Outer, destroys it, and sends the Inner back to the sender. After calling this function, the previous owner of Outer can access Inner directly by its ID because it is now unwrapped. Wrapping and unwrapping of an object can happen multiple times across its lifespan, and the object retains its ID across all those events.
The Lamport timestamp-based versioning scheme ensures that the version that an object is unwrapped at is always greater than the version it was wrapped at, to prevent version re-use.
- After a transaction,
W, where objectIis wrapped by objectO,O's version is greater than or equal toI's. This means one of the following conditions is true:Iis an input so has a strictly lower version.Iis new and has an equal version.
- After a later transaction unwrapping
Iout ofO, the following must be true:O's input version is greater than or equal to its version afterWbecause it is a later transaction, so the version can only have increased.I's version in the output must be strictly greater thanO's input version.
This leads to the following chain of inequalities for I's version before wrapping:
- less than or equal to
O's version after wrapping - less than or equal to
O's version before unwrapping - less than
I's version after unwrapping
So I's version before wrapping is less than I's version after unwrapping.
From a versioning perspective, values held in dynamic fields behave like wrapped objects:
- They are only accessible via the field's parent object, not as direct transaction inputs.
- Based on the previous point, you do not need to supply their IDs or versions with the transaction inputs.
- Lamport timestamp-based versioning makes sure that when a field contains an object and a transaction removes that field, its value becomes accessible by its ID and the value's version has been incremented to a previously unused version.
One distinction to wrapped objects is that if a transaction modifies a dynamic object field, its version is incremented in that transaction, where a wrapped object's version would not be.
Adding a new dynamic field to a parent object also creates a Field object, responsible for associating the field name and value with that parent. Unlike other newly created objects, the ID for the resulting instance of Field is not created using sui::object::new. Instead, it is computed as a hash of the parent object ID and the type and value of the field name, so that it can be used to look-up the Field via its parent and name.
When a field is removed its associated Field is deleted, and if a new field with the same name is added, a new instance will be created, with the same ID. Versioning using Lamport timestamps, coupled with the fact that dynamic fields can only be accessed through their parent object ensures that (ID, version) pairs are not reused in the process:
- The transaction that deletes the original field will increment the parent's version to be greater than the deleted field's version.
- The transaction that creates the new version of the same field will create it with a version that is greater than the parent's version.
So the version of the new Field instance will be greater than the version of the deleted Field.
Move packages are also versioned and stored on chain, but follow a different versioning scheme to objects because they are immutable from their inception. This means that you refer to package transaction inputs (e.g. the package that a function is from for a move call transaction) by just their ID, and are always loaded at their latest version.
Every time you publish or upgrade a package a new ID is generated, a newly published package will have its version set to 1, whereas an upgraded package's version will be one greater than the package it is upgrading. Unlike objects, older versions of a package remain accessible even after being upgraded. For example, imagine a package P that is published and upgraded twice. It might be represented in the store as:
(0x17fb7f87e48622257725f584949beac81539a3f4ff864317ad90357c37d82605, 1) => P v1
(0x260f6eeb866c61ab5659f4a89bc0704dd4c51a573c4f4627e40c5bb93d4d500e, 2) => P v2
(0xd24cc3ec3e2877f085bc756337bf73ae6976c38c3d93a0dbaf8004505de980ef, 3) => P v3
In the example above, all three versions of the same package are at different IDs. The packages have increasing versions but it is possible to call into v1, even though v2 and v3 exist on-chain.
Framework packages (such as the Move standard library at 0x1 and the Sui Framework at 0x2) are a special-case because their IDs must remain stable across upgrades. The network can upgrade framework packages while preserving their IDs via a system transaction, but can only perform this operation on epoch boundaries because they are considered immutable like other packages. New versions of framework packages retain the same ID as their predecessor, but increment their version by one:
(0x1, 1) => MoveStdlib v1
(0x1, 2) => MoveStdlib v2
(0x1, 3) => MoveStdlib v3
The prior example shows the on-chain representation of the first three versions of the Move standard library.