The Common Expression Language (CEL) is a non-Turing complete language designed for simplicity, speed, safety, and portability. CEL's C-like syntax looks nearly identical to equivalent expressions in C++, Go, Java, and TypeScript.

// Check whether a resource name starts with a group name.
resource.name.startsWith("/groups/" + auth.claims.group)

// Determine whether the request is in the permitted time window.
request.time - resource.age < duration("24h")

// Check whether all resource names in a list match a given filter.
auth.claims.email_verified && resources.all(r, r.startsWith(auth.claims.email))

A CEL "program" is a single expression. The examples have been tagged as java, go, and typescript within the markdown to showcase the commonality of the syntax.

CEL is ideal for lightweight expression evaluation when a fully sandboxed scripting language is too resource intensive. To get started, try the Codelab.

A dashboard that shows results of cel-go conformance tests can be found here.

• Overview
  • Environment Setup
  • Parse and Check
  • Macros
  • Evaluate
  • Errors
• Examples
• Install
• Common Questions
• License

Determine the variables and functions you want to provide to CEL. Parse and check an expression to make sure it's valid. Then evaluate the output AST against some input. Checking is optional, but strongly encouraged.

Let's expose name and group variables to CEL using the cel.Declarations environment option:

import "github.com/google/cel-go/cel"

env, err := cel.NewEnv(
    cel.Variable("name", cel.StringType),
    cel.Variable("group", cel.StringType),
)

That's it. The environment is ready to be used for parsing and type-checking. CEL supports all the usual primitive types in addition to lists, maps, as well as first-class support for JSON and Protocol Buffers.

The parsing phase indicates whether the expression is syntactically valid and expands any macros present within the environment. Parsing and checking are more computationally expensive than evaluation, and it is recommended that expressions be parsed and checked ahead of time.

The parse and check phases are combined for convenience into the Compile step:

ast, issues := env.Compile(`name.startsWith("/groups/" + group)`)
if issues != nil && issues.Err() != nil {
    log.Fatalf("type-check error: %s", issues.Err())
}
prg, err := env.Program(ast)
if err != nil {
    log.Fatalf("program construction error: %s", err)
}

The cel.Program generated at the end of parse and check is stateless, thread-safe, and cachable.

Type-checking in an optional, but strongly encouraged, step that can reject some semantically invalid expressions using static analysis. Additionally, the check produces metadata which can improve function invocation performance and object field selection at evaluation-time.

Macros are optional but enabled by default. Macros were introduced to support optional CEL features that might not be desired in all use cases without the syntactic burden and complexity such features might desire if they were part of the core CEL syntax. Macros are expanded at parse time and their expansions are type-checked at check time.

For example, when macros are enabled it is possible to support bounded iteration / fold operators. The macros all, exists, exists_one, filter, and map are particularly useful for evaluating a single predicate against list and map values.

// Ensure all tweets are less than 140 chars
tweets.all(t, t.size() <= 140)

The has macro is useful for unifying field presence testing logic across protobuf types and dynamic (JSON-like) types.

// Test whether the field is a non-default value if proto-based, or defined
// in the JSON case.
has(message.field)

Both cases traditionally require special syntax at the language level, but these features are exposed via macros in CEL.

Now, evaluate for fun and profit. The evaluation is thread-safe and side-effect free. Many different inputs can be sent to the same cel.Program and if fields are present in the input, but not referenced in the expression, they are ignored.

// The `out` var contains the output of a successful evaluation.
// The `details' var would contain intermediate evaluation state if enabled as
// a cel.ProgramOption. This can be useful for visualizing how the `out` value
// was arrive at.
out, details, err := prg.Eval(map[string]interface{}{
    "name": "/groups/acme.co/documents/secret-stuff",
    "group": "acme.co"})
fmt.Println(out) // 'true'

What if name hadn't been supplied? CEL is designed for this case. In distributed apps it is not uncommon to have edge caches and central services. If possible, evaluation should happen at the edge, but it isn't always possible to know the full state required for all values and functions present in the CEL expression.

To improve the odds of successful evaluation with partial state, CEL uses commutative logical operators &&, ||. If an error or unknown value (not the same thing) is encountered on the left-hand side, the right hand side is evaluated also to determine the outcome. While it is possible to implement evaluation with partial state without this feature, this method was chosen because it aligns with the semantics of SQL evaluation and because it's more robust to evaluation against dynamic data types such as JSON inputs.

In the following truth-table, the symbols <x> and <y> represent error or unknown values, with the ? indicating that the branch is not taken due to short-circuiting. When the result is <x, y> this means that the both args are possibly relevant to the result.

In the cases where unknowns are expected, cel.EvalOptions(cel.OptTrackState) should be enabled. The details value returned by Eval() will contain the intermediate evaluation values and can be provided to the interpreter.Prune function to generate a residual expression. e.g.:

// Residual when `name` omitted:
name.startsWith("/groups/acme.co")

This technique can be useful when there are variables that are expensive to compute unless they are absolutely needed. This functionality will be the focus of many future improvements, so keep an eye out for more goodness here!

Parse and check errors have friendly error messages with pointers to where the issues occur in source:

ERROR: <input>:1:40: undefined field 'undefined'
    | TestAllTypes{single_int32: 1, undefined: 2}
    | .......................................^`,

Both the parsed and checked expressions contain source position information about each node that appears in the output AST. This information can be used to determine error locations at evaluation time as well.

CEL-Go supports modules and uses semantic versioning. For more info see the Go Modules docs.

And of course, there is always the option to build from source directly.

JavaScript and Lua are rich languages that require sandboxing to execute safely. Sandboxing is costly and factors into the "what will I let users evaluate?" question heavily when the answer is anything more than O(n) complexity.

CEL evaluates linearly with respect to the size of the expression and the input being evaluated when macros are disabled. The only functions beyond the built-ins that may be invoked are provided by the host environment. While extension functions may be more complex, this is a choice by the application embedding CEL.

But, why not WASM? WASM is an excellent choice for certain applications and is far superior to embedded JavaScript and Lua, but it does not have support for garbage collection and non-primitive object types require semi-expensive calls across modules. In most cases CEL will be faster and just as portable for its intended use case, though for node.js and web-based execution CEL too may offer a WASM evaluator with direct to WASM compilation.

Checking is an optional, but strongly suggested, step in CEL expression validation. It is sufficient in some cases to simply Parse and rely on the runtime bindings and error handling to do the right thing.

• See the CEL Spec for the specification and conformance test suite.
• Ask for support on the CEL Go Discuss Google group.

• See GoDoc to learn how to integrate CEL into services written in Go.
• See the CEL C++ toolchain (under development) for information about how to integrate CEL evaluation into other environments.

• See CONTRIBUTING.md to get started.
• Use GitHub Issues to request features or report bugs.

A handful of tests rely on Bazel. In particular dynamic proto support at check time and the conformance test driver require Bazel to coordinate the test inputs:

bazel test ...

Released under the Apache License.

Disclaimer: This is not an official Google product.