Concepts

Pact is designed to be used in distinct execution modes to address the performance requirements of rapid linear execution on a blockchain. These are:

In this mode, a large amount of code is sent into the blockchain to establish the smart contract, as comprised of modules (code), tables (data), and keysets (authorization). This can also include "transactional" (database-modifying) code, for instance to initialize data.

For a given smart contract, these should all be sent as a single message into the blockchain, so that any error will rollback the entire smart contract as a unit.

Keysets are customarily defined first, as they are used to specify admin authorization schemes for modules and tables. Definition creates the keysets in the runtime environment and stores their definition in the global keyset database.

Namespace declarations provide a unique prefix for modules and interfaces defined within the namespace scope. Namespaces are handled differently in public and private blockchain contexts: in private they are freely definable, and the root namespace (ie, not using a namespace at all) is available for user code. In public blockchains, users are not allowed to use the root namespace (which is reserved for built-in contracts like the coin contract) and must define code within a namespace, which may or may not be definable (ie, users might be restricted to "user" namespaces).

Namespaces are defined using define-namespace. Namespaces are "entered" by issuing the namespace command.

Modules contain the API and data definitions for smart contracts. They are comprised of:

When a module is declared, all references to native functions, interfaces, or definitions from other modules are resolved. Resolution failure results in transaction rollback.

Modules can be re-defined as controlled by their governance capabilities. Often, such a function is simply a reference to an administrative keyset. Module versioning is not supported, except by including a version sigil in the module name (e.g., "accounts-v1"). However, module hashes are a powerful feature for ensuring code safety. When a module is imported with use, the module hash can be specified, to tie code to a particular release.

As of Pact 2.2, use statements can be issued within a module declaration. This combined with module hashes provides a high level of assurance, as updated module code will fail to import if a dependent module has subsequently changed on the chain; this will also propagate changes to the loaded modules' hash, protecting downstream modules from inadvertent changes on update.

Module names must be unique within a namespace.

Interfaces contain an API specification and data definitions for smart contracts. They are comprised of:

Interfaces represent an abstract api that a module may implement by issuing an implements statement within the module declaration. Interfaces may import definitions from other modules by issuing a use declaration, which may be used to construct new constant definitions, or make use of types defined in the imported module. Unlike Modules, Interface versioning is not supported. However, modules may implement multiple interfaces.

Interface names must be unique within a namespace.

Tables are created at the same time as modules. While tables are defined in modules, they are created "after" modules, so that the module may be redefined later without having to necessarily re-create the table.

The relationship of modules to tables is important, as described in Table Guards.

There is no restriction on how many tables may be created. Table names are namespaced with the module name.

Tables can be typed with a schema.

"Transactions" refer to business events enacted on the blockchain, like a payment, a sale, or a workflow step of a complex contractual agreement. A transaction is generally a single call to a module function. However there is no limit on how many statements can be executed. Indeed, the difference between "transactions" and "smart contract definition" is simply the kind of code executed, not any actual difference in the code evaluation.

Querying data is generally not a business event, and can involve data payloads that could impact performance, so querying is carried out as a local execution on the node receiving the message. Historical queries use a transaction ID as a point of reference, to avoid any race conditions and allow asynchronous query execution.

Transactional vs local execution is accomplished by targeting different API endpoints; pact code has no ability to distinguish between transactional and local execution.

Pact presents a database metaphor reflecting the unique requirements of blockchain execution, which can be adapted to run on different back-ends.

A single message sent into the blockchain to be evaluated by Pact is atomic: the transaction succeeds as a unit, or does not succeed at all, known as "transactions" in database literature. There is no explicit support for rollback handling, except in multi-step transactions.

Blockchain execution can be likened to OLTP (online transaction processing) database workloads, which favor denormalized data written to a single table. Pact's data-access API reflects this by presenting a key-row model, where a row of column values is accessed by a single key.

As a result, Pact does not support joining tables, which is more suited for an OLAP (online analytical processing) database, populated from exports from the Pact database. This does not mean Pact cannot record transactions using relational techniques -- for example, a Customer table whose keys are used in a Sales table would involve the code looking up the Customer record before writing to the Sales table.

As of Pact 2.3, Pact offers a powerful query mechanism for selecting multiple rows from a table. While visually similar to SQL, the select and where operations offer a streaming interface to a table, where the user provides filter functions, and then operates on the rowset as a list data structure using sort and other functions.

In a transactional setting, Pact database interactions are optimized for single-row reads and writes, meaning such queries can be slow and prohibitively expensive computationally. However, using the local execution capability, Pact can utilize the user filter functions on the streaming results, offering excellent performance.

The best practice is therefore to use select operations via local, non-transactional operations, and avoid using select on large tables in the transactional setting.

Pact has no concept of a NULL value in its database metaphor. The main function for computing on database results, with-read, will error if any column value is not found. Authors must ensure that values are present for any transactional read. This is a safety feature to ensure totality and avoid needless, unsafe control-flow surrounding null values.

The key-row model is augmented by every change to column values being versioned by transaction ID. For example, a table with three columns "name", "age", and "role" might update "name" in transaction #1, and "age" and "role" in transaction #2. Retrieving historical data will return just the change to "name" under transaction 1, and the change to "age" and "role" in transaction #2.

Pact guarantees identical, correct execution at the smart-contract layer within the blockchain. As a result, the backing store need not be identical on different consensus nodes. Pact's implementation allows for integration of industrial RDBMSs, to assist large migrations onto a blockchain-based system, by facilitating bulk replication of data to downstream systems.

With Pact 2.0, Pact gains explicit type specification, albeit optional. Pact 1.0 code without types still functions as before, and writing code without types is attractive for rapid prototyping.

Schemas provide the main impetus for types. A schema is defined with a list of columns that can have types (although this is also not required). Tables are then defined with a particular schema (again, optional).

Note that schemas also can be used on/specified for object types.

Any types declared in code are enforced at runtime. For table schemas, this means any write to a table will be typechecked against the schema. Otherwise, if a type specification is encountered, the runtime enforces the type when the expression is evaluated.

With the typecheck repl command, the Pact interpreter will analyze a module and attempt to infer types on every variable, function application or const definition. Using this in project repl scripts is helpful to aid the developer in adding "just enough types" to make the typecheck succeed. Successful typechecking is usually a matter of providing schemas for all tables, and argument types for ancillary functions that call ambiguous or overloaded native functions.

Pact's typechecker is designed to output a fully typechecked and inlined AST for generating formal proofs in the SMT-LIB2 language. If the typecheck does not succeed, the module is not considered "provable".

We see, then, that Pact code can move its way up a "safety" gradient, starting with no types, then with "enough" types, and lastly, with formal proofs.

Note that as of Pact 2.0 the formal verification function is still under development.

Pact is inspired by Bitcoin scripts to incorporate public-key authorization directly into smart contract execution and administration. Pact seeks to take this further by making single- and multi-sig interactions ubiquitous and effortless with the concept of keysets, meaning that single-signature mode is never assumed: anywhere public-key signatures are used, single-sig and multi-sig can interoperate effortlessly. Finally, all crypto is handled by the Pact runtime to ensure programmers can't make mistakes "writing their own crypto".

Also see Guards and Capabilities below for how Pact moves beyond just keyset-based authorization.

Keysets are defined by reading definitions from the message payload. Keysets consist of a list of public keys and a keyset predicate.

Examples of valid keyset JSON productions:

A keyset predicate references a function by its (optionally qualified) name, and will compare the public keys in the keyset to the key or keys used to sign the blockchain message. The function accepts two arguments, "count" and "matched", where "count" is the number of keys in the keyset and "matched" is how many keys on the message signature matched a keyset key.

Support for multiple signatures is the responsibility of the blockchain layer, and is a powerful feature for Bitcoin-style "multisig" contracts (i.e. requiring at least two signatures to release funds).

Pact comes with built-in keyset predicates: keys-all, keys-any, keys-2. Module authors are free to define additional predicates.

If a keyset predicate is not specified, keys-all is used by default.

Keysets can be rotated, but only by messages authorized against the current keyset definition and predicate. Once authorized, the keyset can be easily redefined.

When creating a table, a module name must also be specified. By this mechanism, tables are "guarded" or "encapsulated" by the module, such that direct access to the table via data-access functions is authorized only by the module's governance. However, within module functions, table access is unconstrained. This gives contract authors great flexibility in designing data access, and is intended to enshrine the module as the main "user data access API".

See also module guards for how this concept can be leveraged to protect more than just tables.

Note that as of Pact 3.5, the option has been added to selectively allow unguarded reads and transaction history access in local mode only, at the discretion of the node operator.

Keysets can be stored as a column value in a row, allowing for row-level authorization. The following code indicates how this might be achieved:

In the example, create-account reads a keyset definition from the message payload using read-keyset to store as "keyset" in the table. read-balance only allows that owner's keyset to read the balance, by first enforcing the keyset using enforce-keyset.

Namespaces are defined by specifying a namespace name and associating a keyset with the namespace. Namespace scope is entered by declaring the namespace environment. All definitions issued after the namespace scope is entered will be accessible by their fully qualified names. These names are of the form namespace.module.definition. This form can also be used to access code outside of the current namespace for the purpose of importing module code, or implementing modules:

Code may be appended to the namespace by simply re-entering the namespace and declaring new code definitions. All definitions must occur within a namespace, as the global namespace (the empty namespace) is reserved for Kadena code.

Examples of valid namespace definition and scoping:

Defining a namespace requires a keyset, and a namespace name of type string:

Members of a namespace may be accessed by their fully-qualified names:

Modules may be imported at a namespace, and interfaces my be implemented in a similar way. This allows the user to work with members of a namespace in a much less verbose and cumbersome way.

If one is simply appending code to an existing namespace, then the namespace prefix in the fully qualified name may be ommitted, as using a namespace works in a similar way to importing a module: all toplevel definitions within a namespace are brought into scope when (namespace 'my-namespace) is declared. Continuing from the previous example:

Pact 3.0 introduces powerful new concepts to allow programmers to express and implement authorization schemes correctly and easily: guards, which generalize keysets, and capabilities, which generalize authorizations or rights. In Pact 3.7, capabilities also function as events.

A guard is essentially a predicate function over some environment that enables a pass-fail operation, enforce-guard, to be able to test a rich diversity of conditions.

A keyset is the quintessential guard: it specifies a list of keys, and a predicate function to verify how many keys were used to sign the current transaction. Enforcement happens via enforce-keyset, causing the transaction to fail if the necessary keys are not found in the signing set.

However, there are other predicates that are equally useful:

Finally, we want guards to interoperate with each other, so that smart contract code doesn't have to worry about what kind of guard is used to mediate access to some resource or right. For instance, it is easy to think of entries in a ledger having diverse guards, where some tokens are guarded by keysets, while others are autonomously owned by modules, while others are locked in some kind of escrow transaction: what's important is that the guard always be enforced for the given account, not what type of guard it is.

Guards address all of these needs. Keysets are now just one type of guard, to which we add module guards, pact guards, and completely customizable "user guards". You can store any type of guard in the database using the guard type. The keyset type is still supported, but developers should switch to guard to enjoy the enhanced flexibility.

Capabilities are a new construct in Pact 3.0 that draws from capability theory to offer a system for managing runtime user rights in an explicit, literate, and principled fashion.

Simply put, a capability is a "ticket" that when acquired allows the user to perform some sensitive task. If the user is unable to acquire the ticket, portions of the transaction that demand the ticket will fail.

Code can demand that a capability be "already granted", that is, make no attempt to acquire the ticket, but fail if it was not acquired somewhere else. This is done with the construct require-capability.

Code can also directly attempt to acquire a capability, but only for a specific scope. This is done with the special form with-capability, which, like with-read, scopes a body of code. Here, the ticket is granted while this body of code is executing, and is revoked when the body leaves execution.

We've described capabilities like a "ticket", so let's continue by adding some attributes to this ticket:

Pact provides the defcap construct to do this.

ALLOW_ENTRY is the name or domain of the capability. user-id is a parameter. Together, they form the specification of a capability. Thus, (ALLOW_ENTRY 'dave) and (ALLOW_ENTRY 'carol) describe separate capailities. (Note that capability theory's notion of designation is indicated here, which we'll return to when we discuss capabilities and signatures).

The body implements the predicate function. It accesses whatever data it needs to perform necessary tests to protect against improper granting of the ticket. The body can do more than that -- it can import or compose additional capabilities, for instance -- and it can even modify database state. This might be used to ensure a capability cannot be granted ever again after the first time it is acquired, for example.

To acquire this capability, you would invoke with-capability:

To demand or require the capability, you would use require-capability:

Requiring capabilities allow for "private" or "restricted" functions than cannot be called directly. Here we see that do-entry can only be called "privately", by code inside the module somewhere. What's more, it can only be called in an outer operation for this user in particular, "restricting" it to that user.

A defcap can "import" other capabilities, for modular factoring of guard code, or to "compose" the outer capability from "smaller", "inner" capabilities.

Composed capabilities are only in scope when their "parent" capability is granted.

In Pact transaction messages, each signer can "scope" their signature to one or more capabilities. This restricts keyset guard operations on that signature: keysets demanding the scoped signature will only succeed while the ticket is held, or is in the process of being acquired -- keysets are often checked in order to grant a capability.

This "scoping" allows the signer to safely call untrusted code. For instance, in the Chainweb gas system, the "sender" signs the message to fund whatever gas costs are charged for the transaction. By signing the message, the sender has potentially allowed any code to debit from their account!

With that sender's signature has (GAS) added to it, it is scoped within gas payments in the coin contract only. Third-party code is prohibited from accessing that account during the transaction.

Signature capabilities are also a mechanism to install capabilities, but only if that capability is managed. "Vanilla" capabilities are just tickets to show before you try some protected operation, but managed capabilities are able to change the state of a capability as it is brought into and out of scope. The ticket metaphor breaks down here, as this is now a dynamic object that mediates whether capabilities are acquired.

If a signer attaches a managed capability to their signature list, the capability is "installed", which is not the same as "granted" or "acquired": if the capability's predicate function allows this signer to install the capability, the installed version will then govern any code needing the capability to unlock some protected operation, by means of a manager function.

A managed capability allows for safe interoperation with otherwise untrusted code. By signing with a managed capability, you are allowing some untrusted code to request grant of the capability; if the capability was not in the signature list, the untrusted code cannot request it.

If the capability manager function doesn't grant the request, the untrusted code fails to execute. The common usage of this is to grant a payment to third-party code, such that the third-party code can directly transfer on behalf of the user some amount of coin, but only up to the indicated amount.

TRANSFER allows for sender to approve any number of payments to receiver up to some amount. Once the amount is exceeded, the capability can no longer be brought into scope.

This allows third-party code to directly enact payments. Managed capabilities are an important feature to allow smart contracts to directly call some other trusted code in a tightly-constrained context.

A managed capability that does not specify a manager function is "auto-managed", meaning that after install, the capability can be granted exactly once for the given parameters. Further attempts will fail after the initial grant goes out of scope.

In the following example, the capability will have "one-shot" automatic management:

Guards and capabilities can be confusing: given we have guards like keysets, what do we need the capability concept for?

Guards allow us to define a rule that must be satisfied for the transaction to proceed. As such, they really are just a way to declare a pass-fail condition or predicate. The Pact guard system is flexible enough to express any rule you can code.

Capabilities allow us to declare how that rule is deployed to grant some authority. In doing so, they enumerate the critical rights that are extended to users of the smart contract, and "protect" code from being called incorrectly.

Note also that capabilities can only be granted inside the module code that declares them, whereas guards are simply data that can be tested anywhere. This is an important security property, as it ensures an attacker cannot elevate their privileges from outside the module code.

The only problem with the above code is it pushed the awareness of DEBIT into the transfer function, whereas separation of concerns would better have it housed in debit. What's more, we'd like to ensure that debit is always called in a "transfer" capacity, that is, that the corresponding credit occurs. Thus, the better way to model this is with two capabilities, with TRANSFER being a "no-guard" capability that simply encloses debit and credit calls:

Thus, TRANSFER protects debit and credit from being used independently, while DEBIT governs specifically the ability to debit, enforcing the guard, while CREDIT simply creates a "restricted" capability for credit.

Once capabilities are granted they are installed into the pact environment for the scope of the call to with-capability; once that form is exited, the capability is uninstalled. This scoping prevents duplicate testing of the predicate: capabilities that have already been acquired (or installed) and are in-scope are not re-evaluated, either by acquiring or requiring.

Since a defcap production both specifies a "domain" of capability instances, and implements the guard function, it has some surprising features. Since capability grant is cached in the environment, the function does not need to be called when invoked in with-capability or require-capability asks for some already-granted ticket.

As a result, defcaps cannot be executed directly, as arbitrary execution would violate the semantics described here. This is an important security property as it ensures that the granting code can only be called in approved contexts, inside the module.

Scoped signatures can be tested using the new env-sigs REPL function as follows:

Guards come in five flavors: keyset, keyset reference, module, pact, and user guards.

These are the classic pact keysets. Using the keyset type is the one instance where you can restrict a guard subtype, otherwise the guard type obscures the implementation type to prevent developers from engaging in guard-specific control flow, which would be against best practices. Again, it is better to switch to guard unless there is a specific need to use keysets.

Keysets can be installed into the environment with define-keyset, but if you wanted to store a reference to a defined keyset, you would need to use a string type. To make environment keysets interoperate with concrete keysets and other guards, we introduce the "keyset reference guard" which indicates that a defined keyset is used instead of a concrete keyset.

Module guards are a special guard that when enforced will fail unless:

This is for allowing a module or smart contract to autonomously "own" and manage some asset. As such it is operationally identical to how module table access is guarded: only module code or a transaction having module admin can directly write to a module tables, or upgrade the module, so there is no need to use a module guard for these in-module operations. A module guard is used to "project" module admin outside of the module (e.g. to own coins in an external ledger), or "inject" module admin into an internal database representation (e.g. to own an internally-managed asset alongside other non-module owners).

See Module Governance for more information about module admin management.

create-module-guard takes a string argument to allow naming the guard, to indicate the purpose or role of the guard.

Pact guards are a special guard that will only pass if called in the specific defpact execution in which the guard was created.

Imagine an escrow transaction where the funds need to be moved into an escrow account: if modeled as a two-step pact, the funds can go into a special account named after the pact id, guarded by a pact guard. This means that only code in a subsequent step of that particular pact execution (ie having the same pact ID) can pass the guard.

Pact guards turn pact executions into autonomous processes that can own assets, and is a powerful technique for trustless asset management within a multi-step operation.

User guards allow the user to design an arbitrary predicate function to enforce the guard, given some initial data. For instance, a user guard could be designed to require two separate keysets to be enforced:

User guards can seem similar to capabilities but are different, namely in that they can be stored in the database and passed around like plain data. Capabilities are in-module rights that can only be enforced within the declaring module, and offer scoping and the other benefits mentioned above. User guards are for implementing custom predicate logic that can't be expressed by other built-in guard types.

The following example shows how a "hash timelock" guard can be made, to implement atomic swaps.

Pact 3.7 introduces events which are emitted in the course of a transaction and included in the transaction receipt to allow for monitoring and proving via SPV that a particular event transpired.

In Pact, events are modeled as capabilities, for the following reasons:

Any capability can cause events to be emitted upon acquisition by using the @event metadata tag.

@event cannot be used alongside @managed, because ...

Managed capabilites emit events automatically with the parameters specified in acquisition (as opposed to install). From an eventing point of view, managed capabilities are those capabilities that can only "happen once". Whereas, a non-managed, eventing capability can fire events an arbitrary amount of times.

Use env-events to test for emitted events in repl scripts.

Before Pact 3.0, module upgrade and administration was governed by a defined keyset that is referenced in the module definition. With Pact 3.0, this string value can alternately be an unqualified bareword that references a defcap within the module body. This defcap is the module governance capability.

With the introduction of the governance capability syntax, Pact modules now support generalized module governance, allowing for module authors to design any governance scheme they wish. Examples include tallying a stakeholder vote on an upgrade hash, or enforcing more than one keyset.

To illustrate, let's consider a module governed by a keyset:

This indicates that if a user tried to upgrade the module, or directly write to the module tables, 'foo-keyset would be enforced on the transaction signature set.

This can be directly implemented in a governance capability as follows:

Note the capability can have whatever name desired; GOVERNANCE is a good idiomatic name however.

As a defcap, the governance function cannot be called directly by user code. It is automatically invoked in the following circumstances:

In these cases, the transaction is tested for elevated access to "module admin", defined as the grant of the module admin capability. This capability cannot be expressed in user code, so it cannot be installed, acquired, required or composed.

However, the implementing capability, here called GOVERNANCE, can be installed or acquired etc. If passed, this gets scoped like any normal capability, here over some protected code that only module admins can run.

The special module admin capability, once automatically installed in the cases described above, stays in scope for the rest of the calling transaction. This is unlike "user" capabilities, which can only be acquired in a fixed scope specified by the body of with-capability.

This may sound worrisome, but the rationale is that a governance capability once granted should not be based on some transient fact that can become false during a single transaction. This is important especially in module upgrades, which can change the governance capability itself: if the module admin was tested again this could cause the upgrade to fail, for instance when migrating data with direct table rights.

Also, this means that, when initially installing a module, the governance function is not invoked. This is different behavior than when a keyset is specified: the keyset must be defined and it is enforced, to ensure that the keyset actually exists.

Module governance is therefore more "risky" as it can mean that the module cannot be upgraded if there is a bug in the governance capability. Clearly, care must be taken when implementing module capabilities, and using the Pact formal verification system is highly recommended here.

In the following code, a module can be upgraded based on a vote. An upgrade is designed as a Pact transaction, and its hash and code are distributed to stakeholders, who vote for the upgrade. Once the upgrade is sent in, the vote is tallied in the governance capability, and if a simple majority is found, the code is upgraded.

An interface, as defined in Pact, is a collection of models used for formal verification, constant definitions, and typed function signatures. When a module issues an implements, then that module is said to 'implement' said interface, and must provide an implementation . This allows for abstraction in a similar sense to Java's interfaces, Scala's traits, Haskell's typeclasses or OCaML's signatures. Multiple interfaces may be implemented in a given module, allowing for an expressive layering of behaviors.

Interfaces are declared using the interface keyword, and providing a name for the interface. Since interfaces cannot be upgraded, and no function implementations exist in an interface aside from constant data, there is no notion of governance that need be applied. Multiple interfaces may be implemented by a single module. If there are conflicting function names among multiple interfaces, then the two interfaces are incompatible, and the user must either inline the code they want, or redefine the interfaces to the point that the conflict is resolved.

Constants declared in an interface can be accessed directly by their fully qualified name namespace.interface.const, and so, they do not have the same naming constraints as function signatures.

Additionally, interfaces my make use of module declarations, admitting use of the use keyword, allowing interfaces to import members of other modules. This allows interface signatures to be defined in terms of table types defined in an imported module.

Formal verification is implemented at multiple levels within an interface in order to provide an extra level of security. Models may be declared either within the body of the interface or at the function level in the same way that one would declare them in a module, with the exception that not all models are applicable to an interface. Indeed, since there is no abstract notion of tables for interfaces, abstract table invariants cannot be declared. However, if an interface imports table schema and types from a module via the use keyword, then the interface can define body and function models that apply directly to the concrete table type. Otherwise, all properties are candidates for declaration in an interface.

When models are declared in an interface, they are appeneded to the list of models present in the implementing module at the level of declaration: body-level models are appended to body-level models, and function-level models are appended to function-level models. This allows users to extend the constraints of an interface with models applicable to specific business logic and implementation.

Declaring models shares the same syntax with modules:

Pact 3.7 introduces module references (also called "modrefs"), a new language feature that enables important use-cases that require polymorphism. For example, a Uniswap-like DEX allows users to specify pairs of tokens to allow trading between them. The fungible-v2 interface allows tokens to offer identical operations such as transfer-create, but without a way to abstract over different fungible-v2 implementations, a DEX smart contract would have to be upgraded for each pair with custom code for every operation.

With module references, the DEX can now accept pairs of modref values where each value references a concrete module that implements the fungible-v2 interface, giving it the ability to call fungible-v2 operations using those values.

To invoke the above function, the module names are directly referenced in code.

Module reference values are "normal Pact values" that can be stored in the database, referenced in events and returned from functions.

Modrefs provide polymorphism for use cases like the example above with an emphasis on interoperability. A modref is specified with one or more interfaces, allowing for values of that modref to reference modules that implement those interfaces.

In the calling example above, the modref a-token:module{fungible-v2} accepts a reference to the Kadena coin KDA token module, because coin implements fungible-v2. Of course there is nothing special about fungible-v2: modrefs can specify any defined interface and accept any module that implements said interface.

The polymorphism offered by modrefs resembles generics in Java or traits in Rust, and should not be confused with more object-oriented polymorphism like that found with Java classes or TypeScript types. Modules cannot "extend" one another, they can only offer operations that match some interface specification, and interfaces themselves cannot extend some other interface.

Modrefs introduce indirection which increases overall complexity, making the system harder to understand and reason about. Reach for modrefs when your code wants to offer flexible interoperation to other smart contracts, but if it's just your code, strive to use direct references whenever possible.

Modrefs are "late-binding", which means that the latest upgraded version of a module will be used when a module operation is invoked.

Consider a modref to a module stored in the database when the module is at version 1. Sometime later the module is upgraded to version 2. The modref in the database will refer to the upgraded version 2 of the module when read back in and used.

As described in the Dependency Management section, Pact direct references are not late-binding, so this modref behavior might be surprising.

In the common case of employing modrefs to allow foreign modules to operate with your code, this of course means that you should not assume that this code is safe: indeed, any modref call should be treated as untrusted code.

Specifically, modref invocation in the context of capability acquisition can result in unintended privilege escalation, in the common case of using require-capability to protect functions from being called directly.

Consider a module with a public function collect-data that is intended to allow foreign modules to provide some data, resulting in the one-time payment of a fee. The foreign modules implement data-collector which offers collect to get the data, and get-fee-recipient to identify the receiving account. The module code acquires the COLLECT capability, and uses this to prevent two delegate functions from being called directly. Unfortunately, with the wrong code, this seemingly benign code can be exploited by a malicious modref implementor.

The problem with the above code is that the with-capability call happens before the calls to the modref operations, such that while the foreign module code is executing, the COLLECT capability is in scope. While this is true, pay-fee (and store-data as well) can be called from anywhere.

As such, a malicious coder could provide a modref whose code directly calls data-market.pay-fee as many times as they like in the seemingly innocent calls to collect or get-fee-recipient. They could also call data-market.store-data and wreak havoc that way. Once a capability is in scope, the protections provided by require-capability are not available.

Fortunately, this is easily avoided by keeping modref calls out of scope of the sensitive capability.

Now, the modref calls have safely returned before the capability is acquired. A malicious implementation has no way to invoke the sensitive code.

Modules and interfaces thus need to be referenced directly, which is simply accomplished by issuing their name in code.

Using a module reference in a function is accomplished by specifying the type of the module reference argument, and using the dereference operator :: to invoke a member function of the interfaces specified in the type.

Here we cover various aspects of Pact's approach to computation.

Pact is turing-incomplete, in that there is no recursion (recursion is detected before execution and results in an error) and no ability to loop indefinitely. Pact does support operation on list structures via map, fold and filter, but since there is no ability to define infinite lists, these are necessarily bounded.

Turing-incompleteness allows Pact module loading to resolve all references in advance, meaning that instead of addressing functions in a lookup table, the function definition is directly injected (or "inlined") into the callsite. This is an example of the performance advantages of a Turing-incomplete language.

Pact allows variable declarations in let expressions and bindings. Variables are immutable: they cannot be re-assigned, or modified in-place.

A common variable declaration occurs in the with-read function, assigning variables to column values by name. The bind function offers this same functionality for objects.

Module-global constant values can be declared with defconst.

Pact code can be explicitly typed, and is always strongly-typed under the hood as the native functions perform strict type checking as indicated in their documented type signatures.

Pact's supported types are:

Pact is designed to maximize the performance of transaction execution, penalizing queries and module definition in favor of fast recording of business events on the blockchain. Some tips for fast execution are:

Design transactions so they can be executed with a single function call.

When calling module functions in transactions, use reference syntax instead of importing the module with use. When defining modules that reference other module functions, use is fine, as those references will be inlined at module definition time.

A transaction can encode values directly into the transactional code:

or it can read values from the message JSON payload:

The latter will execute slightly faster, as there is less code to interpret at transaction time.

With table schemas, Pact will be strongly typed for most use cases, but functions that do not use the database might still need types. Use the typecheck REPL function to add the necessary types. There is a small cost for type enforcement at runtime, and too many type signatures can harm readability. However types can help document an API, so this is a judgement call.

Pact supports conditionals via if, bounded looping, and of course function application.

"If" should never be used to enforce business logic invariants: instead, enforce is the right choice, which will fail the transaction.

Indeed, failure is the only non-local exit allowed by Pact. This reflects Pact's emphasis on totality.

Note that enforce-one (added in Pact 2.3) allows for testing a list of enforcements such that if any pass, the whole expression passes. This is the sole example in Pact of "exception catching" in that a failed enforcement simply results in the next test being executed, short-circuiting on success.

The built-in keyset functions keys-all, keys-any, keys-2 are hardcoded in the interpreter to execute quickly. Custom keysets require runtime resolution which is slower.

Pact includes the functional-programming "greatest hits": map, fold and filter. These all employ partial application, where the list item is appended onto the application arguments in order to serially execute the function.

Pact also has compose, which allows "chaining" applications in a functional style.

In certain contexts Pact can guarantee that computation is "pure", which simply means that the database state will not be modified. Currently, enforce, enforce-one and keyset predicate evaluation are all executed in a pure context. defconst memoization is also pure.

Pact's use of LISP syntax is intended to make the code reflect its runtime representation directly, allowing contract authors focus directly on program execution. Pact code is stored in human-readable form on the ledger, such that the code can be directly verified, but the use of LISP-style s-expression syntax allows this code to execute quickly.

Pact expects code to arrive in a message with a JSON payload and signatures. Message data is read using read-msg and related functions. While signatures are not directly readable or writable, they are evaluated as part of keyset predicate enforcement.

Values returned from Pact transactions are expected to be directly represented as JSON values.

When reading values from a message via read-msg, Pact coerces JSON types as follows:

Integer values are represented as objects and read using read-integer.

Pact is designed to be used in a confidentiality-preserving environment, where messages are only visible to a subset of participants. This has significant implications for smart contract execution.

An entity is a business participant that is able or not able to see a confidential message. An entity might be a company, a group within a company, or an individual.

Pact smart contracts operate on messages organized by a blockchain, and serve to produce a database of record, containing results of transactional executions. In a confidential environment, different entities execute different transactions, meaning the resulting databases are now disjoint.

This does not affect Pact execution; however, database data can no longer enact a "two-sided transaction", meaning we need a new concept to handle enacting a single transaction over multiple disjoint datasets.

An important feature for confidentiality in Pact is the ability to orchestrate disjoint transactions in sequence to be executed by targeted entities. This is described in the next section.

"Pacts" are multi-stage sequential transactions that are defined as a single body of code called a pact. Defining a multi-step interaction as a pact ensures that transaction participants will enact an agreed sequence of operations, and offers a special "execution scope" that can be used to create and manage data resources only during the lifetime of a given multi-stage interaction.

Pacts are a form of coroutine, which is a function that has multiple exit and re-entry points. Pacts are composed of steps such that only a single step is executed in a given blockchain transaction. Steps can only be executed in strict sequential order.

A pact is defined with arguments, similarly to function definition. However, arguments values are only evaluated in the execution of the initial step, after which those values are available unchanged to subsequent steps. To share new values with subsequent steps, a step can yield values which the subsequent step can recover using the special resume binding form.

Pacts are comprised of steps that can only execute in strict sequence. Any enforcement of who can execute a step happens within the code of the step expression. All steps are "manually" initiated by some participant in the transaction with CONTINUATION commands sent into the blockchain.

In pacts, a rollback expression is specified to indicate that the pact can be "cancelled" at this step with a participant sending in a CANCEL message before the next step is executed. Once the last step of a pact has been executed, the pact will be finished and cannot be rolled back. Failures in public steps are no different than a failure in a non-pact transaction: all changes are rolled back. Pacts can therefore only be canceled explicitly and should be modeled to offer all necessary cancel options.

A step can yield values to the following step using yield and resume. This is an unforgeable value, as it is maintained within the blockchain pact scope.

Every time a pact is initiated, it is given a unique ID which is retrievable using the pact-id function, which will return the ID of the currently executing pact, or fail if not running within a pact scope. This mechanism can thus be used to guard access to resources, analogous to the use of keysets and signatures. One typical use of this is to create escrow accounts that can only be used within the context of a given pact, eliminating the need for a trusted third party for many use-cases.

Pacts can be tested in repl scripts using the env-entity, env-step and pact-state repl functions to simulate pact executions.

It is also possible to simulate pact execution in the pact server API by formatting continuation Request yaml files into API requests with a cont payload.

Pact supports a number of features to manage a module's dependencies on other Pact modules.

Once loaded, a Pact module is associated with a hash computed from the module's source code text. This module hash uniquely identifies the version of the module. Hashes are base64url-encoded BLAKE2 256-bit hashes. Module hashes can be examined with describe-module:

The use special form allows a module hash to be specified, in order to pin the dependency version. When used within a module declaration, it introduces the dependency hash value into the module's hash. This allows a "dependency-only" upgrade to push the upgrade to the module version.

When a module is loaded, all references to foreign modules are resolved, and their code is directly inlined. At this point, upstream definitions are permanent: the only way to upgrade dependencies is to reload the original module.

This permanence is great for user code: once a module is loaded, an upstream provider cannot change what code is executed within. However, this creates a big problem for upstream developers, as they cannot upgrade the downstream code themselves in order to address an exploit, or to introduce new features.

A trade-off is needed to balance these opposing interests. Pact offers the ability for upstream code to break downstream dependent code at runtime. Table access is guarded to enforce that the module hash of the inlined dependency either matches the runtime version, or is in a set of "blessed" hashes, as specified by bless in the module declaration:

Dependencies with these hashes will continue to function after the module is loaded. Unrecognized hashes will cause the transaction to fail. However, "pure" code that does not access the database is unaffected. This prevents a "leftpad situation" where trivial utility functions can harm downstream code stability.

Upstream providers can use the bless mechanism to phase in an important upgrade, by renaming the upgraded module to indicate the new version, and replacing the old module with a new, empty module that only blesses the last version (and whatever earlier versions desired). New clients will fail to import the "v1" code, requiring them to use the new version, while existing users can continue to use the old version, presumably up to some advertised time limit. The "empty" module can offer migration functions to handle migrating user data to the new module, for the user to self-upgrade in the time window.