CEL expressions, executed by Truffle. TruffleCEL is an independent Common Expression Language parser, type checker, and evaluator for Java and Elide. Compile reusable policies, validation rules, and filters for GraalVM JVM or Native Image, with optional guest compilation for hot expressions. Production code has no cel-java dependency.
Get started · Documentation · Benchmarks · Contribute · Releases
Use Java 25 and Elide GraalVM 25.5.5-elide.28. TruffleCEL 0.13.0 uses
TruffleData 0.14.0. Add the thin core JAR and the matching runtime repository:
repositories {
mavenCentral()
maven("https://maven.elide.dev")
maven("https://maven.elide.dev/gvm-elide/25.5.5-elide.28")
}
dependencies {
implementation("dev.elide:trufflecel-core:0.13.0")
}import dev.trufflecel.Cel;
import java.util.List;
import java.util.Map;
var policy = Cel.compile("user.age >= 18 && 'admin' in user.roles");
boolean allowed = policy.test(Map.of(
"user", Map.of("age", 30, "roles", List.of("admin"))));Compile once and reuse across threads. Each evaluation has its own activation,
scopes, and operation budget. Map inputs are borrowed: keep their contents stable
until evaluation returns. Use Activation.snapshot(bindings) for reusable owned
inputs. Declare an environment with Cel.newBuilder() for typed variables,
custom functions, and explicit resource limits.
RE2/J is optional. For compatibility-mode regex support, add
com.google.re2j:re2j:1.8. Applications can omit it by selecting
TruffleRegex-only mode, subject to its explicit
context and supported-pattern contract.
| Maven coordinate | Purpose |
|---|---|
dev.elide:trufflecel-core |
Compiler, evaluator, Polyglot language and native TruffleData rows |
dev.elide:trufflecel-protobuf |
Protobuf inputs, wire decoding and canonical CEL AST interchange |
dev.elide:trufflecel-json |
Oracle Truffle JSON input conversion |
dev.elide:trufflecel-processor |
Build-time generation of reflection-free POJO, JavaBean and record adapters |
Use matching versions. Put the processor on the annotation processor path; generated
adapters depend on core. The direct Java API does not require a Polyglot context;
the registered language ID is cel. See integration for
ownership, numeric and presence semantics, and publishing
for Maven/Bazel consumption and runtime dependencies.
Start with interpreter-to-interpreter results under the same host runtime. Guest JIT is a separate choice for hot, reused expressions; warmed evaluation excludes compilation and warmup. The comparison guide explains the evidence, comparator settings, and remaining limits in response to upstream methodology feedback.
The 0.11.0 report covers six fixtures, including protobuf policy, materialized repeated fields, maps/regex, wide activations, and optimizer opportunities. cel-java 0.15.0 runs with both default and constant-folding/CSE planners. Every non-scalar fixture retains all 69 bindings. Figures use ordinary Map inputs and three forks per cell in one guarded round on the recorded Linux host; they do not establish universal rankings.
| Metric | TruffleCEL interpreter: lower / tied / higher | Guest JIT: lower / tied / higher |
|---|---|---|
| Hot evaluation time | 12 / 0 / 0 | 12 / 0 / 0 |
| Current-thread allocation per evaluation | 11 / 1 / 0 | 12 / 0 / 0 |
| Current whole-process RSS | 5 / 0 / 7 | 3 / 0 / 9 |
| Lifetime peak whole-process RSS | 6 / 0 / 6 | 3 / 0 / 9 |
Each mode is compared with the lower cel-java planner aggregate for that metric, fixture and runtime. Counts are descriptive, not significance tests. Two of the interpreter peak-RSS comparisons have overlapping fork ranges.
Hot evaluation with guest JIT
Guest compilation is forced before timing and its cost is excluded.
Allocation per evaluation
Current and peak process memory: TruffleCEL is higher in most comparisons
All cards and controls · Every plotted value · All values and ranges · 0.10.0 report
Input preparation changes the comparison. cel-java copies ordinary maps but reuses Guava ImmutableMaps. In the separate prepared-input controls, JVM scalar-wide cel-java allocation falls from 2,936 to 72 B/op, below TruffleCEL interpreter's 120 B/op. Caller preparation is excluded. Only two fixtures have these controls; ordinary-map counts do not apply to every input API.
Memory and startup remain tradeoffs. Allocation is current-thread bytes/op, not retained heap. RSS covers the whole process, including setup and compilation; the shared comparison binary is not a standalone deployment footprint. Startup/payback and retained-heap measurements use earlier TruffleData 0.9.0 artifacts, and are explicitly historical. Current-release startup/payback and separate deployment-size measurements remain open evidence gaps.
CI also runs CodSpeed simulation, allocation limits and native guest-compilation gates. CodSpeed instruction counts are distinct from JVM/JIT throughput. See methodology and historical reports.
All 2,771 cases in the pinned 32-file cel-spec corpus run without skips, through both a test fixture and the shipped protobuf adapter, and through frontend and canonical AST paths. An additional 431-case negative/edge table checks rejection phase and defined error categories; 368 cases also run against cel-java with both parsers and optimizer settings, with 22 documented divergences.
This is evidence for the tested contracts, not exhaustive CEL specification compliance. CI enforces 99.8% line and branch coverage, resource-limit and concurrency tests, and coverage-guided frontend/decoder fuzzing. Read compatibility and limitations, validation, and the threat model before exposing expressions to untrusted input.
Bazelisk selects the pinned Bazel and GraalVM toolchains. Local development needs no BuildBuddy credentials:
bazel run --config=local //bazel/gazelle:update
bazel run --config=local //tools:format
bazel test --config=local //...
bazel coverage --config=local //src/test/...
bazel run --config=local //tools:coverageThe default configuration uses BuildBuddy RBE; --config=local uses the host.
Native guest-JIT qualification runs on Linux. See Bazel setup,
contributor guidance, agent guidance, and
CI. Contributions of code, tests, documentation, and reproducible
performance investigations are welcome.
Support · Governance · Code of conduct · Security policy · OpenSSF evidence and remaining work
First-party TruffleCEL is dual-licensed under Apache-2.0 or
MIT, at your option (Apache-2.0 OR MIT). Contributors offer their
work under both alternatives and retain copyright; see the
contributor agreement. Third-party components
retain their own terms and notices. Report vulnerabilities privately using the
security process.
