Repository navigation
4.8.2: Bounded Native Responses
What Changes For Your Application?
Your Java handlers, service classes and configuration stay the same. Update the
platform parent or REST dependency to 4.8.2, then rebuild your application.
Use cache 0.9.2 and Dubbo 0.9.2 when those libraries are part of your service.
<dependency>
<groupId>com.reactor</groupId>
<artifactId>rust-java-rest</artifactId>
<version>4.8.2</version>
</dependency>Smaller Response Setup
An untouched RawResponse.json(body) no longer creates a mutable header map.
Reading or changing its headers still works. This is not response caching: each
request keeps its own body and encoded headers.
return RawResponse.json(jsonBytes);Repeated bodyless JNI handlers also reuse a bounded per-worker response-size
hint. The initial hint is capped at 16 KiB. It does not enlarge queued request
frames, keep business responses, or change response-size limits.
Dubbo Completion Ownership
The shared Rust runtime releases a completion queue slot when it dequeues a
result. The response-body permit remains held through the Java callback. This
avoids rejecting a follow-up RPC because its previous result still occupied a
queue slot, without removing backpressure or allowing unbounded response memory.
Upgrade Native Files Together
The JNI contracts remain REST 29, Dubbo 7, Spring Dubbo 3, Redis 6 and Glowroot 6.
Use this release's packaged native library. If you set an explicit native path,
replace that DLL/SO/dylib too; updating only the Java dependency is insufficient.
The clean native source revision is
26ee9b9f518077af1fc0817c8f37a96fd1a78533.
Verification Scope
The EmaTech cache detail/segment c16/c64 comparison used three fresh JVMs per arm.
RPS changed by +0.03% to +6.55%; p99 changed by -9.97% to +8.27%. Median peak RSS
changed from 79.66 to 79.41 MiB. No unexpected request errors or OOM occurred.
These are workload-specific results, not a universal speed or memory guarantee.
The full REST suite executed 586 tests: 585 passed and one allocation-counter
check was unsupported on OpenJ9. That check separately passed on EmaTech HotSpot
21: 336 bytes across 100,000 operations, below the 8,192-byte bound. It does not
measure native allocations or prove OpenJ9 zero allocation.
Release assets retain exact compatibility and soak receipts. Compilation,
runtime compatibility and performance qualification are separate evidence.
Memory/latency settings and platform defaults are not enlarged by this patch.
The 30-minute EmaTech soak completed 17,430,899 requests across 24 load windows
without request errors, OOM or restarts. Peak RSS was 68.36 MiB and peak sampled
cgroup memory was 109.64 MiB under a 128 MiB limit. The last two idle intervals
were approximately flat. This is a 30-minute observation, not a multi-day leak
guarantee. Published patch JARs are bound separately by class/resource hashes.