Smart Optimistic Rollups

A rollup is a processing unit that receives, retrieves and interprets input messages to update its local state and to produce output messages targetting the Tezos blockchain. In this documentation, we will generally refer to the rollup under consideration as the Layer 2 on top of the Tezos blockchain, considered as the Layer 1.

Rollups are a permissionless scaling solution for the Tezos blockchain. Indeed, anyone can originate and operate one or more rollups, allowing to increase the throughput of the Tezos blockchain, (almost) arbitrarily.

The integration of these rollups in the Tezos protocol is optimistic: this means that when a participant publishes a claim about the state of the rollup, this claim is a priori trusted. However, a refutation mechanism allows anyone to economically punish someone who has published an invalid claim. Therefore, thanks to the refutation mechanism, a single honest participant is enough to guarantee that the input messages are correctly interpreted.

In the Tezos protocol, the subsystem of smart rollups is generic with respect to the syntax and the semantics of the input messages. More precisely, the originator of a smart rollup provides a program named a kernel (in one of the languages supported by Tezos) responsible for interpreting input messages. During the refutation mechanism, the execution of this kernel is handled by a Proof-generating Virtual Machine (PVM) for this language, provided by the Tezos protocol, which allows to prove that the result of applying an input message to the rollup context is correct. The rest of the time, any VM implementation of the chosen language can be used to run the smart rollup kernel, provided that it is compliant with the PVM.

The smart rollup infrastructure currently supports the WebAssembly language. A WASM rollup runs a kernel expressed in WASM. The role of the kernel is to process input messages, to update a state, and to output messages targeting the Layer 1 following a user-defined logic. Anyone can develop a kernel or reuse existing kernels. A typical use case of WASM rollups is to deploy a kernel that implements the Ethereum Virtual Machine (EVM) and to get as a result an EVM-compatible Layer 2 running on top of the Tezos blockchain. WASM rollups are not limited to this use case though: they are fully programmable, hence their names, smart optimistic rollups, as they are very close to smart contracts in terms of expressiveness.

The purpose of this documentation is to give:

  1. an overview of the terminology and basic principles of smart rollups;

  2. a complete tour of smart rollups related workflows;

  3. a reference documentation for the development of a WASM kernel.


Just like smart contracts, smart rollups are decentralized software components. However, contrary to smart contracts that are processed by the network validators automatically, a smart rollup requires a dedicated rollup node to function.

Any user can originate, operate, and interact with a rollup. For the sake of clarity, we will distinguish three kinds of users in this documentation: operators, kernel developers, and end-users. An operator deploys the rollup node to make the rollup progress. A kernel developer writes a kernel to be executed within a rollup. An end-user interacts with the rollup through Layer 1 operations or Layer 2 input messages.


When a smart rollup is originated on the Layer 1, a unique address is generated to uniquely identify it. A smart rollup address starts with the prefix sr1 (see also the kinds of address prefixes in Tezos).


There are two channels of communication to interact with smart rollups:

  1. a global rollups inbox allows the Layer 1 to transmit information to all the rollups.

  2. a reveal data channel allows each rollup to retrieve data coming from data sources external to the Layer 1. Rollups request data through that channel to the runner of that rollup kernel (i.e. the smart rollup node).

Rollups inbox

A single global inbox serves all rollups and contains two kinds of messages: external messages are pushed through a Layer 1 manager operation while internal messages are pushed by Layer 1 smart contracts or by the protocol itself. All messages (external and internal) pushed to the inbox also contain the Layer 1 level of their insertion and a counter. The counter is the index of the message and it is reset at each Layer 1 level.

External messages

Anyone can push a message to the rollups inbox. This message is a mere sequence of bytes following no particular underlying format. The interpretation of this sequence of bytes is the responsibility of each kernel.

There are two ways for end-users to push an external message to the rollups inbox: first, they can inject the dedicated Layer 1 operation using the Octez client (see command send smart rollup message <messages> from <src>); second, they can use the batcher of a smart rollup node. More details can be found in Sending an external inbox message.

Internal messages

Contrary to external messages, which are submitted by the end users, internal messages are constructed by the Layer 1.

At the beginning of every Tezos block, the Layer 1 pushes two internal messages: “Start of level”, and “Info per level”. “Start of level” does not have any payload associated to it, while “Info per level” provides to the kernel the timestamp and block hash of the predecessor of the current Tezos block. If the Tezos block is the first block of a protocol, then the Layer 1 pushes another message “Protocol migration” just after the “Info per level” that provides the new protocol version (i.e. <proto-name>_<NNN>).

A rollup is identified by an address and has an associated Michelson type (defined at origination time). Any Layer 1 smart contract can perform a transfer to this address with a payload of this type. This transfer is realized as an internal message pushed to the rollups inbox.

Finally, after the application of the operations of the Tezos block, the Layer 1 pushes one final internal message “End of level”. Similarly to “Start of level“, these internal messages does not come with any payload.

Reveal data channel

The reveal data channel is a communication interface that allows the rollup to request data from sources that are external to the inbox and can be unknown to the Layer 1. The rollup node has the responsibility to answer the rollup requests.

A rollup can do the following requests through the reveal data channel:

  1. preimage requests: The rollup can request arbitrary data of at most 4kBytes, provided that it knows its (blake2b) hash. The request is fulfilled by the rollup node Populating the reveal channel.

  2. metadata requests The rollup can request information from the protocol, namely the address and the origination level of the rollup node itself. The rollup node retrieves this information through RPCs to answer the rollup.

Information passing through the reveal data channel does not have to be considered by the Layer 1: for this reason, the volume of information is not limited by the bandwidth of the Layer 1. Thus, the reveal data channel can be used to upload large volumes of data to the rollup.


A smart rollup is characterized by: - the kind of Proof-generating Virtual Machine (PVM), - the kernel written in a language that the PVM can interpret, - the Michelson type of the entrypoint used by Layer 1 smart contracts to send internal messages to it, and - an optional list of addresses used as a white-list of allowed stakers (see Private rollup).

All these characteristics are provided when originating a new smart rollup.


Each time a Tezos block is finalized, a rollup reacts to three kinds of events: the beginning of the block, the input messages possibly contained in that block, and the end of the block. A rollup node implements this reactive process: it downloads the Tezos block and interprets it according to the semantics of the PVM. This interpretation can require updating a state, downloading data from other sources, or performing some cryptographic verifications. The state of the rollup contains an outbox, which is a sequence of latent calls to Layer 1 contracts.

The behavior of the rollup node is deterministic and fully specified by a reference implementation of the PVM embedded in the protocol. Notice that the PVM implementation is meant for verification, not performance: for this reason, a rollup node does not normally run a PVM to process inputs but a fast execution engine (e.g., based on the Wasmer runtime for the WASM PVM in the case of the rollup node distributed with Octez). This fast execution engine implements the exact same semantics as the PVM. The PVM is only ever used by the rollup node when it needs to produce a proof during the last step of the refutation mechanism.


Starting from the rollup origination level, levels are partitioned into commitment periods of 60 consecutive blocks.

A commitment claims that the interpretation of all inbox messages published during a given commitment period, and applied on the state of a parent commitment, led to a given new state by performing a given number of execution steps of the PVM. Execution steps are called ticks in Smart Rollups terminology. A commitment must be published on the Layer 1 after each commitment period to have the rollup progress. A commitment is always based on a parent commitment (except for the genesis commitment that is automatically published at origination time).

Since the PVM is deterministic and the inputs are completely determined by the Layer 1 rollups inbox and the reveal channel, there is only one honest commitment. In other words, if two distinct commitments are published for the same commitment period, one of them must be wrong.

Notice that, to publish a commitment, an operator must provide a deposit of 10,000 tez. For this reason, the operator is said to be a staker. Several users can stake on the same commitment. When a staker S publishes a new commitment based on a commitment that S is staking on, S does not have to provide a new deposit: the deposit also applies to this new commitment.

There is no need to synchronize between operators: if two honest operators publish the same commitment for a given commitment period, the commitment will be published with two stakes on it.

A commitment is optimistically trusted but it can be refuted until it is said to be cemented (i.e., final, unchangeable). Indeed, right after a commitment is published, a two-weeks refutation period starts. During the refutation period, anyone noticing that a commitment for a given commitment period is invalid can post a concurrent commitment for the same commitment period to force the removal of the invalid commitment. If no one posts such a concurrent commitment during the refutation period, the commitment can be cemented with a dedicated operation injected in Layer 1, and the outbox messages can be executed by the Layer 1 by an explicit Layer 1 operation (see Triggering the execution of an outbox message), typically to transfer assets from the rollup to the Layer 1.

The outbox messages can follow three different formats. Firstly, the Layer 1 operations contained in the outbox messages can be left untyped, meaning only the Micheline expression is provided by the kernel. Before executing the transaction, the Layer 1 typechecks said expression against the expected type of the targeted entrypoint. Since Nairobi, it is also possible for the kernel to provide its expected type of the targeted entrypoint. This additional safety mechanism is to avoid type confusion: namely, a kernel transferring a tuple that the Layer 1 interprets as a ticket. Lastly, the outbox message can contain a white-list update. This message can only be executed for a rollup that is private since its origination (see Private rollup).


Because of concurrent commitments, a rollup is generally related to a commitment tree where branches correspond to different claims about the rollup state.

By construction, only one view of the rollup state is valid (as the PVM is deterministic). When two concurrent branches exist in the commitment tree, the cementation process is stopped at the first fork in the tree. To unfreeze the cementation process, a refutation game must be started between two concurrent stakers of these branches. Refutation games are automatically played by rollup nodes to defend their stakes: honest participants are guaranteed to win these games. Therefore, an honest participant should not have to worry about refutation games. Finally, a running refutation game does not prevent new commitments to be published on top of the disputed commitments.

A refutation game is decomposed into two main steps: a dissection mechanism and a final conflict resolution phase. During the first phase, the two stakers exchange hashes about intermediate states of the rollups in a way that allows them to converge to the very first tick on which they disagree. The exact number of hashes exchanged at a given step is PVM-dependent. During the final phase, the stakers must provide a proof that they correctly interpreted this conflicting tick.

The Layer 1 PVM then determines whether these proofs are valid. There are only two possible outcomes: either one of the stakers, that we dub S in the sequel, has provided a valid proof, then S wins the game, and is rewarded with half of the opponent’s deposit (the other half being burnt); or, both stakers have provided an invalid proof and they both lose their deposit. In the end, at most one stake will be kept in the commitment tree. When a commitment has no more stake on it (because all stakers have lost the related refutation games), it is removed from the tree. An honest player H must therefore play as many refutation games as there are stakes on the commitments in conflict with H’s own commitment.

Finally, notice that each player is subject to a timer similar to a chess clock, allowing each player to play only up to one week: after this time is elapsed, a player can be dismissed by any Layer 1 user playing a timeout operation. Thus, the refutation game played by the two players can last at most 2 weeks.

There is no timeout for starting a refutation game after having published a concurrent commitment. However, assuming the existence of an honest participant H, then H will start the refutation game with all concurrent stakers to avoid the rollup getting stuck.

Private rollup

A private smart rollup guarantees that private data can’t be leaked by any means, whereas in a public rollup, one can force a rollup to leak part of the data by starting a refutation game. This is achieved by restricting the set of allowed stakers with a white-list. With that restriction only addresses on the white-list can publish commitments and therefore participate in a refutation game.

The white-list is optionally defined at origination. The rollup is considered public if no white-list is defined, private otherwise. The white-list can be updated with a specific outbox message. This message contains an optional list, the new list completely replaces the stored white-list in layer 1. If the message contains no list, then the rollup becomes public. In turn, it is forbidden to make a public rollup private by sending an outbox message with a non-empty white-list.

It is the responsibility of the kernel to maintain the white-list by submitting outbox messages. Kernels must therefore implement their own access control list logic to add and remove addresses.

Also, it is important to remember that because of the refutation logic, an outbox message can only be executed when the associated commitment has been cemented (see Triggering the execution of an outbox message).



Smart rollups come with two new executable programs: the Octez rollup node and the Octez rollup client.

The Octez rollup node is used by a rollup operator to deploy a rollup. The rollup node is responsible for making the rollup progress by publishing commitments and by playing refutation games.

Just like the Octez node, the Octez rollup node provides an RPC interface. The services of this interface can be called directly with HTTP requests or indirectly using the Octez rollup client.


To experiment with the commands described in this section, we use the Dailynet. In this section, we assume that ${OPERATOR_ADDR} is a valid implicit account on Dailynet owned by the reader.

Notice that you need a specific development version of Octez to participate to Dailynet. This version is either available from docker images or can be compiled from sources. Please refer to the Dailynet website for installation details.

An Octez rollup node needs an Octez node to run. We assume that an Octez node has been launched locally, typically by issuing:

octez-node config init --data-dir "${ONODE_DIR}" --network "${NETWORK}"
octez-node run --data-dir "${ONODE_DIR}" --network "${NETWORK}" --rpc-addr

in a terminal where ${NETWORK} is of the form and ${ONODE_DIR} is a path for the Octez node store, by default ~/.tezos-node.

The commands will only work when proto_alpha is activated. This can be checked by:

octez-client rpc get /chains/main/blocks/head/protocols

that must return:

{ "protocol": "ProtoALphaALphaALphaALphaALphaALphaALphaALphaDdp3zK",
  "next_protocol": "ProtoALphaALphaALphaALphaALphaALphaALphaALphaDdp3zK" }

In case you do not already have an implicit account, you can generate one with:

octez-client gen keys "${ACCOUNT_NAME}"
octez-client show address "${ACCOUNT_NAME}"

Then, the ${OPERATOR_ADDR} can be set to the hash value (tz1...) returned.

Finally, you need to check that your balance is greater than 10,000 tez to make sure that staking is possible. In case your balance is not sufficient, you can get test tokens for the tz1 address from a faucet, after your node gets synchronized with Dailynet.

octez-client get balance for "${OPERATOR_ADDR}"


Anyone can originate a smart rollup with the following invocation of the Octez client:

octez-client originate smart rollup "${SOR_ALIAS}" \
  from "${OPERATOR_ADDR}" \
  of kind wasm_2_0_0 \
  of type bytes \
  with kernel "${KERNEL}" \
  --burn-cap 999

where ${SOR_ALIAS} is an alias to memorize the smart rollup address in the client. This alias can be used in any command where a smart rollup address is expected. ${KERNEL} is a hex representation of a WebAssembly bytecode serving as an initial program to boot on. From a WASM bytecode file named kernel.wasm, such representation can be obtained through

xxd -ps -c 0 <kernel.wasm> | tr -d '\n'

To experiment, we propose that you use the value ${KERNEL} defined in the given file.

source # defines shell variable KERNEL

If everything went well, the origination command results in:

This sequence of operations was run:
  Manager signed operations:
    From: tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX
    Fee to the baker: ꜩ0.000357
    Expected counter: 36
    Gas limit: 1000
    Storage limit: 0 bytes
    Balance updates:
      tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX ... -ꜩ0.000357
      payload fees(the block proposer) ....... +ꜩ0.000357
    Revelation of manager public key:
      Contract: tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX
      Key: edpkukxtw4fHmffj4wtZohVKwNwUZvYm6HMog5QMe9EyYK3QwRwBjp
      This revelation was successfully applied
      Consumed gas: 1000
  Manager signed operations:
    From: tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX
    Fee to the baker: ꜩ0.000956
    Expected counter: 37
    Gas limit: 2849
    Storage limit: 6572 bytes
    Balance updates:
      tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX ... -ꜩ0.000956
      payload fees(the block proposer) ....... +ꜩ0.000956
    Smart rollup origination:
      Kind: wasm_2_0_0
      Parameter type: bytes
      Kernel Blake2B hash: '24df9e3c520dd9a9c49b447766e8a604d31138c1aacb4a67532499c6a8b348cc'
      This smart rollup origination was successfully applied
      Consumed gas: 2748.269
      Storage size: 6552 bytes
      Address: sr1RYurGZtN8KNSpkMcCt9CgWeUaNkzsAfXf
      Genesis commitment hash: src13wCGc2nMVfN7rD1rgeG3g1q7oXYX2m5MJY5ZRooVhLt7JwKXwX
      Balance updates:
        tz1fp5ncDmqYwYC568fREYz9iwQTgGQuKZqX ... -ꜩ1.638
        storage fees ........................... +ꜩ1.638

The address sr1RYurGZtN8KNSpkMcCt9CgWeUaNkzsAfXf is the smart rollup address. Let’s write it ${SOR_ADDR} from now on.

Deploying a rollup node

Now that the rollup is originated, anyone can make it progress by deploying a rollup node.

First, we need to decide on a directory where the rollup node stores its data. Let us assign ${ROLLUP_NODE_DIR} with this path, by default ~/.tezos-smart-rollup-node.

The rollup node can then be run with:

octez-smart-rollup-node --base-dir "${OCLIENT_DIR}" \
                 run operator for "${SOR_ALIAS_OR_ADDR}" \
                 with operators "${OPERATOR_ADDR}" \
                 --data-dir "${ROLLUP_NODE_DIR}"

where ${OCLIENT_DIR} is the data directory of the Octez client, by default ~/.tezos-client.

The log should show that the rollup node follows the Layer 1 chain and processes the inbox of each level.

Notice that distinct Layer 1 addresses could be used for the Layer 1 operations issued by the rollup node simply by editing the configuration file to set different addresses for publish, add_messages, cement, and refute.

In addition, a rollup node can run under different modes:

  1. operator activates a full-fledged rollup node. This means that the rollup node will do everything needed to make the rollup progress. This includes following the Layer 1 chain, reconstructing inboxes, updating the states, publishing and cementing commitments regularly, and playing the refutation games. In this mode, the rollup node will accept transactions in its queue and batch them on the Layer 1.

  2. batcher means that the rollup node will accept transactions in its queue and batch them on the Layer 1. In this mode, the rollup node follows the Layer 1 chain, but it does not update its state and does not reconstruct inboxes. Consequently, it does not publish commitments nor play refutation games.

  3. observer means that the rollup node follows the Layer 1 chain to reconstruct inboxes, to update its state. However, it will neither publish commitments, nor play a refutation game. It does not include the message batching service either.

  4. maintenance is the same as the operator mode except that it does not include the message batching service.

  5. accuser follows the layer1-chain and computes commitments but does not publish them. Only when a conflicting commitment (published by another staker) is detected will the “accuser node” publish a commitment and participate in the subsequent refutation game.

The following table summarizes the operation modes, focusing on the L1 operations which are injected by the rollup node in each mode.

Add messages


























Yes *




An accuser node will publish commitments only when it detects conflicts; for such cases it must make a deposit of 10,000 tez.

Configuration file

The rollup node can also be configured via one configuration file stored in its own data directory, with the following command that uses the same arguments as the run command:

octez-smart-rollup-node --base-dir "${OCLIENT_DIR}" \
                 init operator config for "${SOR_ALIAS_OR_ADDR}" \
                 with operators "${OPERATOR_ADDR}" \
                 --data-dir "${ROLLUP_NODE_DIR}"

where ${OCLIENT_DIR} must be the directory of the client, containing all the keys used by the rollup node, i.e. ${OPERATOR_ADDR}.

This creates a smart rollup node configuration file:

Smart rollup node configuration written in ${ROLLUP_NODE_DIR}/config.json

Here is the content of the file:

  "data-dir": "${ROLLUP_NODE_DIR}",
  "smart-rollup-address": "${SOR_ADDR}",
  "smart-rollup-node-operator": {
    "publish": "${OPERATOR_ADDR}",
    "add_messages": "${OPERATOR_ADDR}",
    "cement": "${OPERATOR_ADDR}",
    "refute": "${OPERATOR_ADDR}"
  "fee-parameters": {},
  "mode": "operator"

The rollup node can now be run with just:

octez-smart-rollup-node -d "${OCLIENT_DIR}" run --data-dir ${ROLLUP_NODE_DIR}

The configuration will be read from ${ROLLUP_NODE_DIR}/config.json.

Rollup node in a sandbox

The node can also be tested locally with a sandbox environment. (See sandbox documentation.)

Once you initialized the “sandboxed” client data with ./src/bin_client/, you can run a sandboxed rollup node with octez-smart-rollup-node run.

A temporary directory /tmp/tezos-smart-rollup-node.xxxxxxxx will be used. However, a specific data directory can be set with the environment variable SCORU_DATA_DIR.

Sending an external inbox message

The Octez client can be used to send an external message into the rollup inbox. Assuming that ${EMESSAGE} is the hexadecimal representation of the message payload, one can do:

octez-client -d "${OCLIENT_DIR}" -p ProtoALphaAL \
 send smart rollup message "hex:[ \"${EMESSAGE}\" ]" \
 from "${OPERATOR_ADDR}"

to inject such an external message. So let us focus now on producing a viable content for ${EMESSAGE}.

The kernel used previously in our running example is a simple “echo” kernel that copies its input as a new message to its outbox. Therefore, the input must be a valid binary encoding of an outbox message to make this work. Specifically, assuming that we have originated a Layer 1 smart contract as follows:

octez-client -d "${OCLIENT_DIR}" -p ProtoALphaAL \
  originate contract go transferring 1 from "${OPERATOR_ADDR}" \
  running 'parameter string; storage string; code {CAR; NIL operation; PAIR};' \
  --init '""' --burn-cap 0.4

and that this contract is identified by an address ${CONTRACT} (a KT1... address), then one can encode an outbox transaction using the Octez rollup client as follows:

MESSAGE='[ { \
  "destination" : "KT1...", \
  "parameters" : "\"Hello world\"", \
  "entrypoint" : "%default" } ]'

EMESSAGE=$(octez-smart-rollup-client-alpha encode outbox message "${MESSAGE}")

Triggering the execution of an outbox message

Once an outbox message has been pushed to the outbox by the kernel at some level ${L}, the user needs to wait for the commitment that includes this level to be cemented. On Dailynet, the cementation process of a non-disputed commitment is 40 blocks long while on Mainnet, it is 2 weeks long.

When the commitment is cemented, one can observe that the outbox is populated as follows:

octez-smart-rollup-client-alpha rpc get \

Here is the output for this command:

[ { "outbox_level": ${L}, "message_index": "0",
   { "transactions":
       [ { "parameters": { "string": "Hello world" },
           "destination": "${CONTRACT}",
           "entrypoint": "%default" } ] } } ]

At this point, the actual execution of a given outbox message can be triggered. This requires precomputing a proof that this outbox message is indeed in the outbox. In the case of our running example, this proof is retrieved as follows:

PROOF=$(octez-smart-rollup-client-alpha get proof for message 0 \
  of outbox at level "${L}")

Finally, the execution of the outbox message is done as follows:

"${TEZOS_PATH}/octez-client" -d "${OCLIENT_DIR}" -p ProtoALphaAL \
        execute outbox message of smart rollup "${SOR_ALIAS_OR_ADDR}" \
        from "${OPERATOR_ADDR}" for commitment hash "${LCC}" \
        and output proof "${PROOF}"

where ${LCC} is the hash of the latest cemented commitment. Notice that anyone can trigger the execution of an outbox message (not only an operator as in this example).

One can check in the receipt that the contract has indeed been called with the parameter "Hello world" through an internal operation. More complex parameters, typically containing assets represented as tickets, can be used as long as they match the type of the entrypoint of the destination smart contract.

Sending an internal inbox message

A smart contract can push an internal message in the rollup inbox using the Michelson TRANSFER_TOKENS instruction targeting a specific rollup address. The parameter of this transfer must be a value of the Michelson type declared at the origination of this rollup.

Remember that our running example rollup has been originated with:

octez-client originate smart rollup "${SOR_ALIAS}" \
  from "${OPERATOR_ADDR}" \
  of kind wasm_2_0_0 \
  of type bytes \
  booting with "${KERNEL}" \
  -burn-cap 999

The fragment of type bytes of this command declares that the rollup is expecting values of type bytes. (Notice any Michelson type could have been used instead. To transfer tickets to a rollup, this type must mention tickets.)

Here is an example of a Michelson script that sends an internal message to the rollup of our running example. The payload of the internal message is the value passed as parameter of type bytes to the rollup.

parameter bytes;
storage unit;
    PUSH address "${SOR_ADDR}";
    CONTRACT bytes;
    IF_NONE { PUSH string "Invalid address"; FAILWITH } {};
    PUSH mutez 0;
    DIG 2;
    NIL operation;

Populating the reveal channel

It is the responsibility of rollup node operators to get the data passed through the reveal data channel when the rollup requested it.

To answer a request for a page of hash H, the rollup node tries to read the content of a file H named ${ROLLUP_NODE_DIR}/wasm_2_0_0.

Notice that a page cannot exceed 4KB. Hence, larger pieces of data must be represented with multiple pages that reference each other through hashes. It is up to the kernel to decide how to implement this. For instance, one can classify pages into two categories: index pages that are hashes for other pages and leaf pages that contain actual payloads.

Configure WebAssembly fast execution

When the rollup node advances its internal rollup state under normal operation, it does so using the fast execution engine.

This engine uses Wasmer for running WebAssembly code. You may configure the compiler used for compiling WebAssembly code, via the OCTEZ_WASMER_COMPILER environment variable.

The choice of a compiler primarily affects the performance of the WebAssembly code execution vs the compilation time. Some compilers offer certain security guarantees in a blockchain context, such as compiling in linear time to avoid JIT bombs.

The available options are:

Wasmer compiler options






When to use Singlepass



When to use Cranelift

Note that while the rollup node is generally capable of using Wasmer’s LLVM-based compiler, Octez does not currently ship with it.

When the environment variable is undefined, Cranelift is used by default.

Developing WASM Kernels

A rollup is primarily characterized by the semantics it gives to the input messages it processes. This semantics is provided at origination time as a WASM program (in the case of the wasm_2_0_0 kind) called a kernel. More concretely, the kernel is a WASM module encoded in the binary format defined by the WASM standard.

Except for necessary restrictions to ensure determinism (a key requirement for any web3 technology), we support the full WASM language. More precisely, determinism is ensured by the following restrictions:

  1. Instructions and types related to floating-point arithmetic are not supported. This is because IEEE floats are not deterministic, as the standard includes undefined behavior operations.

  2. The length of the call stack of the WASM kernel is bounded.

Modulo the limitations above, a valid kernel is a WASM module that satisfies the following constraints:

  1. It exports a function kernel_run that takes no argument and returns nothing.

  2. It declares and exports exactly one memory.

  3. It only imports the host functions exported by the (virtual) module smart_rollup_core.

For instance, the mandatory example of a hello, world! kernel is the following WASM program in text format.

  (import "smart_rollup_core" "write_debug"
     (func $write_debug (param i32 i32) (result i32)))
  (memory 1)
  (export "mem" (memory 0))
  (data (i32.const 100) "hello, world!")
  (func (export "kernel_run")
    (local $hello_address i32)
    (local $hello_length i32)
    (local.set $hello_address (i32.const 100))
    (local.set $hello_length (i32.const 13))
    (drop (call $write_debug (local.get $hello_address)
                             (local.get $hello_length)))))

This program can be compiled to the WASM binary format with general-purpose tool like WABT.

wat2wasm hello.wat -o hello.wasm

The contents of the resulting hello.wasm file is a valid WASM kernel, though its relevance as a decentralized application is debatable.

One of the benefits of choosing WASM as the programming language for smart rollups is that WASM has gradually become a ubiquitous compilation target over the years. Its popularity has grown to the point where mainstream, industrial languages like Go or Rust now natively compile to WASM. Thus, cargo —the official Rust package manager— provides an official target to compile Rust to .wasm binary files, which are valid WASM kernels. This means that, for this particular example, one can build a WASM kernel while enjoying the strengths and convenience of the Rust language and the Rust ecosystem.

The rest of the section proceeds as follows.

  1. First, we explain the execution environment of a WASM kernel: when it is parsed, executed, etc.

  2. Then, we explain in more details the API at the disposal of WASM kernel developers.

  3. Finally, we demonstrate how Rust in particular can be used to implement a WASM kernel.

Though Rust has become the primary language whose WASM backend has been tested in the context of smart rollups, the WASM VM has not been modified in any way to favor this language. We fully expect that other mainstream languages such as Go are also good candidates for implementing WASM kernels.

Execution Environment

In a nutshell, the life cycle of a smart rollup is a never-ending loop of fetching inputs from the Layer 1, and executing the kernel_run function exposed by the WASM kernel.


The smart rollup carries two states:

  1. A transient state, that is reset after each call to the kernel_run function and is akin to RAM.

  2. A persistent state, that is preserved across kernel_run calls. The persistent state consists in an inbox that is regularly populated with the inputs coming from the Layer 1, the outbox which the kernel can populate with contract calls targeting smart contracts in the Layer 1, and a durable storage which is akin to a file system.

The durable storage is a persistent tree, whose contents are addressed by path-like keys. A path in the storage may contain: a value (also called file) consisting of a sequence of raw bytes, and/or any number of subtrees (also called directories), that is, the paths in the storage prefixed by the current path. Thus, unlike most file systems, a path in the durable storage may be at the same time a file and a directory (a set of sub-paths).

The WASM kernel can write and read the raw bytes stored under a given path (the file), but can also interact (delete, copy, move, etc.) with subtrees (directories).

The values and subtrees under the key /readonly are not writable by a kernel, but can be used by the PVM to give information to the kernel.

WASM PVM Versioning

One of Tezos distinguishing features is its native support for upgrades. At its core, Tezos is a Layer 1 designed to evolve via a self-updating mechanism, subject to an on-line governance process. The self-updating mechanism is also implemented by the smart rollup infrastructure.

The WASM PVM is versioned. Kernels can read the version of the underlying WASM PVM (which is currently interpreting them) by reading the contents of the file stored under the key /readonly/wasm_version in their durable storage.

New WASM PVM versions are introduced by new Layer 1’s protocol upgrades. The WASM PVM will upgrade itself when it reads the Protocol_migration internal message.









The changes in each WASM PVM version can be found by searching for string “PVM” in the corresponding protocol’s changelog, section Smart Rollups (e.g. this section for protocol Alpha).

Control Flow

When a new block is published on Tezos, the inbox exposed to the smart rollup is populated with all the inputs published on Tezos in this block. It is important to keep in mind that all the smart rollups which are originated on Tezos share the same inbox. As a consequence, a WASM kernel has to filter the inputs that are relevant for its purpose from the ones it does not need to process.

Once the inbox has been populated with the inputs of the Tezos block, the kernel_run function is called, from a clean “transient” state. More precisely, the WASM kernel is re-initialized, then kernel_run is called.

By default, the WASM kernel yields when kernel_run returns. In this case, the WASM kernel execution is put on hold while the inputs of the next inbox are being loaded. The inputs that were not consumed by kernel_run are dropped. kernel_run can prevent the WASM kernel from yielding by writing arbitrary data under the path /kernel/env/reboot in its durable storage. In such a case (known as reboot), kernel_run is called again, without dropping unread inputs. The value at /kernel/env/reboot is removed between each call of kernel_run, and the kernel_run function can postpone yielding at most 1,000 reboots for each Tezos level.

A call to kernel_run cannot take an arbitrary amount of time to complete, because diverging computations are not compatible with the optimistic rollup infrastructure of Tezos. To dodge the halting problem, the reference interpreter of WASM (used during the refutation game) enforces a bound on the number of ticks used in a call to kernel_run. Once the maximum number of ticks is reached, the execution of kernel_run is trapped (i.e., interrupted with an error). In turn, the fast execution engine does not enforce this time limit. Hence, it is the responsibility of the kernel developer to implement a kernel_run which does not exceed its tick budget.

The current bound is set to 11,000,000,000 ticks. octez-smart-rollup-wasm-debugger is probably the best tool available to verify the kernel_run function does not take more ticks than authorized.

The direct consequence of this setup is that it might be necessary for a WASM kernel to span a long computation across several calls to kernel_run, and therefore to serialize any data it needs in the durable storage to avoid losing them.

Finally, the kernel can verify if the previous kernel_run invocation was trapped by verifying if some data are stored under the path /kernel/env/stuck.

Host Functions

At its core, the WASM machine defined in the WASM standard is just a very evolved arithmetic machine. It needs to be enriched with so-called host functions in order to be used for greater purposes. The host functions provide an API to the WASM program to interact with an “outer world”.

As for smart rollups, the host functions exposed to a WASM kernel allow it to interact with the components of persistent state:


Loads the oldest input still present in the inbox of the smart rollup in the transient memory of the WASM kernel. This means that the input is lost at the next invocation of kernel_run if it is not written in the durable storage. Since version 2.0.0 of the WASM PVM.


Writes an in-memory buffer to the outbox of the smart rollup. If the content of the buffer follows the expected encoding, it can be interpreted in the Layer 1 as a smart contract call, once a commitment acknowledging the call to this host function is cemented. Since version 2.0.0 of the WASM PVM.


Can be used by the WASM kernel to log events which can potentially be interpreted by an instrumented rollup node. Since version 2.0.0 of the WASM PVM.


Returns the kind of data (if any) stored in the durable storage under a given path: a directory, a file, neither or both. Since version 2.0.0 of the WASM PVM.


Cuts both the value (if any) and any subdirectory under a given path out of the durable storage. Since version 2.0.0 of the WASM PVM.


Cuts the value under a given path out of the durable storage, but leaves the rest of the subtree untouched. Since version 2.0.0-r1 of the WASM PVM.


Copies the subtree under a given path to another key. Since the 2.0.0 version of the WASM PVM.


Behaves as store_copy, but also cuts the original subtree out of the tree. Since version 2.0.0 of the WASM PVM.


Loads at most 4,096 bytes from a file of the durable storage to a buffer in the memory of the WASM kernel. Since version 2.0.0 of the WASM PVM.*


Writes at most 2048 bytes from a buffer in the memory of the WASM kernel to a file of the durable storage, increasing its size if necessary. Note that files in the durable storage cannot exceed \(2^{31} - 1\) bytes (i.e. 2GB - 1). Since the 2.0.0 version of the WASM PVM.


Allocates a new file in the durable storage under a given key. Similarly to store_write, store_create cannot create files larger than the durable storage limits, that is 2GB - 1. Since the 2.0.0-r1 of the WASM PVM.


Returns the size (in bytes) of a file under a given key in the durable storage. Since version 2.0.0 of the WASM PVM.


Returns the number of child objects (either directories or files) under a given key. Since version 2.0.0 of the WASM PVM.


Loads in memory the preimage of a hash. The size of the hash in bytes must be specified as an input to the function. Since the 2.0.0 version of the WASM PVM.


Loads in memory the address of the smart rollup (20 bytes), and the Tezos level of its origination (4 bytes). Since the 2.0.0 version of the WASM PVM.

These host functions use a “C-like” API. In particular, most of them return a signed 32bit integer, where negative values are reserved for conveying errors, as shown in the next table.




Input is too large to be a valid key of the durable storage


Input cannot be parsed as a valid key of the durable storage


There is no file under the requested key


The host functions tried to read or write an invalid section (determined by an offset and a length) of the value stored under a given key


Cannot write a value beyond the 2GB size limit


Invalid memory access (segmentation fault)


Tried to read from the inbox or write to the outbox more than 4,096 bytes


Unknown error due to an invalid access


Attempt to modify a readonly value


Key has no tree in the storage


Outbox is full, no new message can be appended


Key has already a value in the storage

Implementing a WASM Kernel in Rust

Though WASM is a good fit for efficiently executing computation-intensive, arbitrary programs, it is a low-level, stack-based, memory unsafe language. Fortunately, it was designed to be a compilation target, not a language in which developers would directly write their programs.

Rust has several advantages that make it a good candidate for writing the kernel of a smart rollup. Not only does the Rust compiler treat WASM as a first class citizen when it comes to compilation targets, but its approach to memory safety eliminates large classes of bugs and vulnerabilities that arbitrary WASM programs may suffer from.

Setting-up Rust

rustup is the standard way to get Rust. Once rustup is installed, enabling WASM as a compilation target is as simple as running the following command.

rustup target add wasm32-unknown-unknown

Rust also proposes the wasm64-unknown-unknown compilation target. This target is not compatible with Tezos smart rollups, which only provides a 32bit address space.


This document is not a tutorial about Rust, and familiarity with the language and its ecosystem (e.g., how Rust crates are structured in particular) is assumed.

The simplest kernel one can implement in Rust (the one that returns directly after being called, without doing anything particular) is the following Rust file (by convention named in Rust).

pub extern "C" fn kernel_run() {

This code can be easily computed with cargo with the following Cargo.toml.

name = 'noop'
version = '0.1.0'
edition = '2021'

crate-type = ["cdylib"]

The key line to spot is the crate-type definition to cdylib. As a side note, when writing a library that will eventually be consumed by a Kernel WASM crate, this line must be modified to

crate-type = ["cdylib", "rlib"]

Compiling our “noop” kernel is done by calling cargo with the correct argument.

cargo build --target wasm32-unknown-unknown

It is also possible to use the --release CLI flag to tell cargo to optimize the kernel.

To make the use of the target optional, it is possible to create a .cargo/config.toml file, containing the following line.

target = "wasm32-unknown-unknown"

lld = true%

The resulting project looks as follows.

├── .cargo
│   └── config.toml
├── Cargo.toml
└── src

and the kernel can be found in the target/ directory, e.g., ./target/wasm32-unknown-unknown/release/noop.wasm.

By default, Rust binaries (including WASM binaries) contain a lot of debugging information and possibly unused code that we do not want to deploy in our rollup. For instance, our “noop” kernel weighs 1.7MBytes. We can use wasm-strip to reduce the size of the kernel (down to 115 bytes in our case).

Host Functions in Rust

The host functions exported by the WASM runtime to Rust programs are exposed by the following API. The link pragma is used to specify the module that exports them (in our case, smart_rollup_core).

pub struct ReadInputMessageInfo {
    pub level: i32,
    pub id: i32,

#[link(wasm_import_module = "smart_rollup_core")]
extern "C" {
    /// Returns the number of bytes written to `dst`, or an error code.
    pub fn read_input(
        message_info: *mut ReadInputMessageInfo,
        dst: *mut u8,
        max_bytes: usize,
    ) -> i32;

    /// Returns 0 in case of success, or an error code.
    pub fn write_output(src: *const u8, num_bytes: usize) -> i32;

    /// Does nothing. Does not check the correctness of its argument.
    pub fn write_debug(src: *const u8, num_bytes: usize);

    /// Returns
    /// - 0 the key is missing
    /// - 1 only a file is stored under the path
    /// - 2 only directories under the path
    /// - 3 both a file and directories
    pub fn store_has(path: *const u8, path_len: usize) -> i32;

    /// Returns 0 in case of success, or an error code
    pub fn store_delete(path: *const u8, path_len: usize) -> i32;

    /// Returns the number of children (file and directories) under a
    /// given key.
    pub fn store_list_size(path: *const u8, path_len: usize) -> i64;

    /// Returns 0 in case of success, or an error code.
    pub fn store_copy(
        src_path: *const u8,
        scr_path_len: usize,
        dst_path: *const u8,
        dst_path_len: usize,
    ) -> i32;

    /// Returns 0 in case of success, or an error code.
    pub fn store_move(
        src_path: *const u8,
        scr_path_len: usize,
        dst_path: *const u8,
        dst_path_len: usize,
    ) -> i32;

    /// Returns the number of bytes written to the durable storage
    /// (should be equal to `num_bytes`, or an error code.
    pub fn store_read(
        path: *const u8,
        path_len: usize,
        offset: usize,
        dst: *mut u8,
        num_bytes: usize,
    ) -> i32;

    /// Returns 0 in case of success, or an error code.
    pub fn store_write(
        path: *const u8,
        path_len: usize,
        offset: usize,
        src: *const u8,
        num_bytes: usize,
    ) -> i32;

    /// Returns the number of bytes written at `dst`, or an error
    /// code.
    pub fn reveal_metadata(
        dst: *mut u8,
        max_bytes: usize,
    ) -> i32;

    /// Returns the number of bytes written at `dst`, or an error
    /// code.
    pub fn reveal_preimage(
        hash_addr: *const u8,
        hash_size: u8,
        dst: *mut u8,
        max_bytes: usize,
    ) -> i32;

These functions are marked as unsafe for Rust. It is possible to provide a safe API on top of them. For instance, the read_input host function can be used to declare a safe function which allocates a fresh Rust Vector to receive the input.

// Assuming the host functions are defined in a module `host`.

pub const MAX_MESSAGE_SIZE: u32 = 4096u32;

pub struct Input {
    pub level: u32,
    pub id: u32,
    pub payload: Vec<u8>,

pub fn next_input() -> Option<Input> {
    let mut payload = Vec::with_capacity(MAX_MESSAGE_SIZE as usize);

    // Placeholder values
    let mut message_info = ReadInputMessageInfo { level: 0, id: 0 };

    let size = unsafe {
            &mut message_info,

    if 0 < payload.len() {
        unsafe { payload.set_len(size as usize) };
        Some(Input {
            level: message_info.level as u32,
            id: as u32,
    } else {

Coupling Vec::with_capacity along with the set_len unsafe function is a good approach to avoid initializing the 4,096 bytes of memory every time you want to load data of arbitrary size into the WASM memory.

Testing your Kernel


octez-smart-rollup-wasm-debugger is available in the Octez distribution starting with Version 16.1.

Testing a kernel without having to start a rollup node on a test network is very convenient. We provide a debugger as a means to evaluate the WASM PVM without relying on any node and network: octez-smart-rollup-wasm-debugger.

octez-smart-rollup-wasm-debugger --kernel "${WASM_FILE}" --inputs "${JSON_INPUTS}" --rollup "${SOR_ADDR}"

octez-smart-rollup-wasm-debugger takes the target WASM kernel to be debugged as argument, either as a .wasm file (the binary representation of WebAssembly modules) or as a .wast file (its textual representation), and actually parses and typechecks the kernel before giving it to the PVM.

Beside the kernel file, the debugger can optionally take an input file containing inboxes and a rollup address. The expected contents of the inboxes is a JSON value, with the following schema:

  [ { "payload" : <Michelson data>,
      "sender" : <Contract hash of the originated contract for the rollup, optional>,
      "source" : <Implicit account sending the message, optional>
      "destination" : <Smart rollup address> }
    // or
    { "external" : <hexadecimal payload> }

The contents of the input file is a JSON array of arrays of inputs, which encodes a sequence of inboxes, where an inbox is a set of messages. These inboxes are read in the same order as they appear in the JSON file. For example, here is a valid input file that defines two inboxes: the first array encodes an inbox containing only an external message, while the second array encodes an inbox containing two messages:

      "payload" : "0",
      "sender" : "KT1ThEdxfUcWUwqsdergy3QnbCWGHSUHeHJq",
      "source" : "tz1RjtZUVeLhADFHDL8UwDZA6vjWWhojpu5w",
      "destination" : "sr1RYurGZtN8KNSpkMcCt9CgWeUaNkzsAfXf"
    { "payload" : "Pair Unit False" }

Note that the sender, source and destination fields are optional and will be given default values by the debugger, respectively KT18amZmM5W7qDWVt2pH6uj7sCEd3kbzLrHT, tz1Ke2h7sDdakHJQh8WX4Z372du1KChsksyU and sr163Lv22CdE8QagCwf48PWDTquk6isQwv57. If no input file is given, the inbox will be assumed empty. If the option --rollup is given, it replaces the default value for the rollup address.

octez-smart-rollup-wasm-debugger is a debugger, as such it waits for user inputs to continue its execution. Its initial state is exactly the same as right after its origination. Its current state can be inspected with the command show status:

> show status
Status: Waiting for inputs
Internal state: Collect

When started, the kernel is in collection mode internally. This means that it is not executing any WASM code, and is waiting for inputs in order to proceed. The command load inputs will load the first inbox from the file given with the option --inputs, putting Start_of_level and Info_per_level before these inputs and End_of_level after the inputs.

> load inputs
Loaded 3 inputs at level 0

> show status
Status: Evaluating
Internal state: Snapshot

At this point, the internal input buffer can be inspected with the command show inbox.

> show inbox
Inbox has 3 messages:
{ raw_level: 0;
  counter: 0
  payload: Start_of_level }
{ raw_level: 0;
  counter: 1
  payload: 0000000023030b01d1a37c088a1221b636bb5fccb35e05181038ba7c000000000764656661756c74 }
{ raw_level: 0;
  counter: 2
  payload: End_of_level }

The first input of an inbox at the beginning of a level is Start_of_level, and is represented by the message \000\001 on the kernel side. We can now start a kernel_run evaluation:

> step kernel_run
Evaluation took 11000000000 ticks so far
Status: Waiting for inputs
Internal state: Collect

The memory of the interpreter is flushed between two kernel_run calls (at the Snapshot and Collect internal states), however the durable storage can be used as a persistent memory. Let’s assume this kernel wrote data at key /store/key:

> show key /store/key
`<hexadecimal value of the key>`

Since the representation of values is decided by the kernel, the debugger can only return its raw value. Please note that the command show keys <path> will return the keys under the given path. This can help navigate in the durable storage.

> show keys /store

It is also possible to inspect the memory by stopping the PVM before its snapshot internal state, with step result, and inspect the memory at pointer n and length l, and finally evaluate until the next kernel_run:

> step result
Evaluation took 2500 ticks so far
Status: Evaluating
Internal state: Evaluation succeeded

> show memory at p for l bytes
`<hexadecimal value>`

> step kernel_run
Evaluation took 7500 ticks so far
Status: Evaluating
Internal state: Snapshot

Once again, note that values from the memory are output as is, since the representation is internal to WASM.

Finally, it is possible to evaluate the whole inbox with step inbox. It will take care of the possible reboots asked by the kernel (through the usage of the /kernel/env/reboot_flag flag) and stop at the next collection phase.

> step inbox
Evaluation took 44000000000 ticks
Status: Waiting for inputs
Internal state: Collect

To obtain more information on the execution, the command bench will also run the kernel on a full inbox, consumed all inputs, run until more inputs are required, and output some information about the run.

> bench
Ran for 5 kernel_run call:
3173 ticks in 0.014739 seconds
4853 ticks in 0.004381 seconds
4914 ticks in 0.003762 seconds
23352 ticks in 0.008684 seconds
2369 ticks in 0.003198 seconds

Each cycle is a call of the kernel_run function. For each cycle, the number of _effective_ ticks used is shown (ticks corresponding to execution, and not used for padding), along with the duration in seconds.

It is also possible to show the outbox for any given level (show outbox at level 0)

> show outbox at level 0
Outbox has N messages:
{ unparsed_parameters: ..;
  destination: ..;
  entrypoint: ..; }

The reveal channel described previously is available in the debugger, either automatically or through specific commands. The debugger can fill automatically preimages from files in a specific directory on the disk, by default in the preimage subdirectory of the working directory. It can be configured with the option --preimage-dir <directory>. In case there is no corresponding file found for the requested preimage, the debugger will ask for the hexadecimal value of the preimage:

> step inbox
Preimage for hash 0000[..] not found.
> 48656c6c6f207468657265210a
Hello there!

Metadata are automatically filled with level 0 as origination level and the configured smart rollup address (or the default one).

Note that when stepping tick by tick (using the step tick command), it is possible to end up in a situation were the evaluation stops on Waiting for reveal. If the expected value is a metadata, the command reveal metadata will give the default metadata to the kernel. If the value expected is the preimage of a given hash, there are two possible solutions:

  • reveal preimage to read the value from the disk. In that case, the debugger will look for a file of the same name as the expected hash in the preimage subdirectory.

  • reveal preimage of <hex encoded value> can be used to feed a custom preimage hash.


  1. PVM: A Proof-generating Virtual Machine is a reference implementation for a device on top of which a smart rollup can be executed. This reference implementation is part of the Tezos protocol and is the unique source of truth regarding the semantics of rollups. The PVM is able to produce proofs enforcing this truth. This ability is used during the final step of refutation games.

  2. Inbox: A sequence of messages from the Layer 1 to smart rollups. The contents of the inbox are determined by the consensus of the Tezos protocol.

  3. Outbox: A sequence of messages from a smart rollup to the Layer 1. Messages are smart contract calls, potentially containing tickets. These calls can be triggered only when the related commitment is cemented (hence, at least two weeks after the actual execution of the operation).

  4. Commitment period: A period of 60 blocks during which all inbox messages must be processed by the rollup node state to compute a commitment. A commitment must be published for each commitment period.

  5. Refutation period: At the end of each commitment period, a period of two weeks starts to allow any commitment related to this commitment period to be challenged.

  6. Staker: An implicit account that has made a deposit on a commitment.

  7. Refutation game: A process by which the Tezos protocol solves a conflict between two stakers.