Optimize JNI performance and fix Decimal validation by geekphilosophy · Pull Request #349 · cedar-policy/cedar-java

geekphilosophy · 2026-04-16T18:41:41Z

Improve JNI call performance ~3x with reduced overhead. Add JMH benchmark suite for tracking JNI performance. Fix Decimal validator to accept leading zeros in corpus integration tests.

Issue #, if available:

Description of changes:
Cedar-Java JMH Benchmark Results — After Performance Fixes

Date: 2026-04-15
Base: 1d141bc (main)
JDK: 21.0.10 (Corretto)
JMH: 1.37
OS: macOS (Darwin 25.3.0, aarch64)
Config: @WarmUp(iterations=3, time=1) @measurement(iterations=5, time=1) @fork(1)

Changes applied:

Remove thread::spawn per JNI call, use std::panic::catch_unwind instead (interface.rs)
Remove redundant j_input_str.push(' ') (interface.rs)
Eliminate double JSON deserialization (readTree+readValue → single readValue) (BasicAuthorizationEngine.java)
Cache JNI version check in static initializer instead of per-call (BasicAuthorizationEngine.java)
Single-pass escape detection in ValueDeserializer using node.size() + node.has() (ValueDeserializer.java)

Results (thorough run: 2 forks, 5 warmup, 20 measurement iterations):
Benchmark Mode Cnt Score Error Units
AuthorizationBenchmark.isAuthorized_medium avgt 40 122.606 ± 2.303 us/op
AuthorizationBenchmark.isAuthorized_small avgt 40 25.416 ± 0.837 us/op
AuthorizationBenchmark.validate_small avgt 40 159.865 ± 29.990 us/op

Comparison to baseline (1d141bc):
Baseline After Improvement
isAuthorized_small 85.4 ± 3.4 25.4 ± 0.8 ~3.4x faster
isAuthorized_medium 329.1 ± 258.4 122.6 ± 2.3 ~2.7x faster
validate_small 353.3 ± 146.0 159.9 ± 30.0 ~2.2x faster

Improve JNI call performance ~3x with reduced overhead. Add JMH benchmark suite for tracking JNI performance. Fix Decimal validator to accept leading zeros in corpus integration tests. Signed-off-by: Chris Simmons <simmonsc@amazon.com>

Rename snake_case methods to camelCase to comply with the MethodName checkstyle rule ('^[a-z][a-zA-Z0-9]*$'). Signed-off-by: Chris Simmons <simmonsc@amazon.com>

JMH generates padding fields and dead stores intentionally for cache-line isolation, which triggers false positives in SpotBugs. Signed-off-by: Chris Simmons <simmonsc@amazon.com>

mark-creamer-amazon · 2026-04-17T18:34:52Z

+        smallEntities.add(new Entity(resource, new HashMap<>(), new HashSet<>()));
+    }
+
+    private void setUpMediumScenario() throws Exception {


Nit: I think if we anticipate adding more scenarios, we might want to rename this (perhaps basicPhotoFlashScenario), but I think this is good as is

mark-creamer-amazon · 2026-04-17T19:11:57Z

+                if (node.size() == 1) {
+                    if (node.has(ENTITY_ESCAPE_SEQ)) {
+                        JsonNode val = node.get(ENTITY_ESCAPE_SEQ);
+                        if (val.isObject() && val.has("id") && val.has("type") && val.size() == 2) {


Nit: Looks like we can also shed the val.isObject() && because val.has(...) checks this already

Method that allows checking whether this node is JSON Object node and contains value for specified property.

muditchaudhary

Mostly nits except for decimal validation.

muditchaudhary · 2026-04-18T18:40:38Z

+                        JsonNode val = node.get(ENTITY_ESCAPE_SEQ);
+                        if (val.isObject() && val.has("id") && val.has("type") && val.size() == 2) {
+                            EntityIdentifier id = new EntityIdentifier(val.get("id").textValue());
+                            Optional<EntityTypeName> type = EntityTypeName.parse(val.get("type").textValue());


Not required for this PR, but something I remembered while reviewing it: we can further optimize the deserializer by caching results from pure functions that cross the FFI boundary like EntityTypeName.parse(_) to avoid redundant and expensive FFI roundtrip costs for the same inputs. Can probably extend this idea beyond deserializers.

muditchaudhary · 2026-04-18T19:32:32Z

+        mediumRequest = new AuthorizationRequest(principal, action, resource, new HashMap<>());
+
+        Set<Policy> policies = new HashSet<>();
+        policies.add(new Policy(


Nit: We can put the policies in a file and use PolicySet.parsePolicies() Example

Same for Entities using Entities.parse Example

muditchaudhary · 2026-04-18T19:47:37Z

-    let handle = thread::spawn(move || call_cedar_in_thread(j_call_str, j_input_str));

-    let result = match handle.join() {
+    let result = match panic::catch_unwind(|| call_cedar(&j_call_str, &j_input_str)) {


This is nice.

Nit: call_cedar_in_thread is probably dead code now.

- Fix Decimal validation to correctly handle leading zeros and enforce range bounds [-922337203685477.5808, 922337203685477.5807] instead of the previous format-only regex check - Add DecimalTests covering boundary values, leading zeros, and out-of-range rejection - Remove unused EscapeType enum from ValueDeserializer - Remove redundant val.isObject() check in entity deserialization - Move benchmark policies and entities to resource files, use PolicySet.parsePolicies() and Entities.parse() instead of inline construction Signed-off-by: Chris Simmons <simmonsc@amazon.com>

Add --locked flag to cargo install for cargo-zigbuild to prevent dependency resolution from pulling cargo-platform 0.3.3 which requires rustc 1.91. Also fix trailing whitespace flagged by checkstyle. Signed-off-by: Chris Simmons <simmonsc@amazon.com>

Tighten Decimal regex to require digits before and after the dot, rejecting invalid forms like .1, 1., and -.0. Simplify range validation using BigDecimal. Remove trim() to reject whitespace inputs. Add edge case tests. Fix JMH resource path so benchmarks run. Signed-off-by: Chris Simmons <simmonsc@amazon.com>

geekphilosophy force-pushed the main branch from 6ebb7ce to 37a6e85 Compare April 16, 2026 18:47

Chris Simmons added 3 commits April 17, 2026 08:44

Fix checkstyle violations in JMH benchmark method names

c26754c

Rename snake_case methods to camelCase to comply with the MethodName checkstyle rule ('^[a-z][a-zA-Z0-9]*$'). Signed-off-by: Chris Simmons <simmonsc@amazon.com>

Exclude JMH-generated code from SpotBugs analysis

67c69da

JMH generates padding fields and dead stores intentionally for cache-line isolation, which triggers false positives in SpotBugs. Signed-off-by: Chris Simmons <simmonsc@amazon.com>

geekphilosophy force-pushed the main branch from 96aadf0 to 67c69da Compare April 17, 2026 15:44

mark-creamer-amazon reviewed Apr 17, 2026

View reviewed changes

Comment thread CedarJava/src/main/java/com/cedarpolicy/serializer/ValueDeserializer.java

muditchaudhary requested changes Apr 18, 2026

View reviewed changes

Chris Simmons added 2 commits April 18, 2026 15:52

muditchaudhary self-requested a review April 19, 2026 19:38

muditchaudhary reviewed Apr 19, 2026

View reviewed changes

Comment thread CedarJava/src/test/java/com/cedarpolicy/DecimalTests.java

Comment thread CedarJava/src/main/java/com/cedarpolicy/value/Decimal.java Outdated

Comment thread CedarJava/src/main/java/com/cedarpolicy/value/Decimal.java Outdated

muditchaudhary approved these changes Apr 20, 2026

View reviewed changes

mark-creamer-amazon approved these changes Apr 21, 2026

View reviewed changes

muditchaudhary merged commit edffaac into cedar-policy:main Apr 21, 2026
4 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Optimize JNI performance and fix Decimal validation#349

Optimize JNI performance and fix Decimal validation#349
muditchaudhary merged 6 commits intocedar-policy:mainfrom
geekphilosophy:main

geekphilosophy commented Apr 16, 2026

Uh oh!

mark-creamer-amazon Apr 17, 2026

Uh oh!

mark-creamer-amazon Apr 17, 2026 •

edited

Loading

Uh oh!

Uh oh!

muditchaudhary left a comment

Uh oh!

Uh oh!

Uh oh!

muditchaudhary Apr 18, 2026

Uh oh!

Uh oh!

Uh oh!

muditchaudhary Apr 18, 2026

Uh oh!

muditchaudhary Apr 18, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

geekphilosophy commented Apr 16, 2026

Description of changes: Cedar-Java JMH Benchmark Results — After Performance Fixes

Uh oh!

mark-creamer-amazon Apr 17, 2026

Choose a reason for hiding this comment

Uh oh!

mark-creamer-amazon Apr 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

muditchaudhary left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

muditchaudhary Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

muditchaudhary Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

muditchaudhary Apr 18, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Description of changes:
Cedar-Java JMH Benchmark Results — After Performance Fixes

mark-creamer-amazon Apr 17, 2026 •

edited

Loading