- Extract the tenant-profile fetch lock stripe count to a named constant (PR #15744)
- Trim the stale global-lock sentence from the per-tenant lock comment (PR #15744)
- Rename the rate-limit onCreate callback to onMiss and document its idempotency requirement (PR #15744)
- Reuse a single tenant profile local in update(TenantProfileUpdateResult) (PR #15744)
- Make the transport callback thread pool size configurable via transport.callback_thread_pool_size (PR #15744)
- Add a parameterized test locking the TransportLimitsType enum-to-profile-field mapping (PR #15744)
- Add device/gateway rate-limit coverage asserting update(tenantId) reaches tracked entities (PR #15744)
- Add a same-tenant fetch-dedup test and pin the cross-tenant test to distinct stripes (PR #15744)
Construct CoapServer and the LwM2M bootstrap server inside the init try block so a failure in the constructor or build() is cleaned up by the existing catch. Guard CoAP shutdown() against a null server. Add a DTLS-enabled CoAP test that covers the dtlsSessionsExecutor shutdown branch.
The defective code lives in common/transport/transport-api and is shared by all
transports (MQTT, HTTP, CoAP, LwM2M, SNMP); the production incident happened to
surface on MQTT.
On a cold tenant-profile cache (e.g. after a cache clear + restart), a device
reconnect storm could serialize the whole transport instance behind tenant-profile
resolution, saturating the callback pool and stalling the node for ~15 minutes.
Two compounding causes are addressed:
- DefaultTransportTenantProfileCache held a single process-wide ReentrantLock across
the synchronous cross-service getEntityProfile round-trip, so every tenant-profile
cache miss in the whole process was serialized one-at-a-time. Replace it with a
bounded set of per-tenant locks (Guava Striped) so different tenants resolve
concurrently while concurrent misses for the same tenant are still de-duplicated.
- DefaultTransportRateLimitService performed that blocking fetch inside
ConcurrentHashMap.computeIfAbsent's mapping function, holding a CHM bin lock across
the remote round-trip. Pre-fetch the tenant profile before computeIfAbsent so no bin
lock is held across I/O.
Also de-duplicate the four near-identical getXRateLimits methods into one generic
helper, move the per-type rate-limit getters onto the TransportLimitsType enum, and
avoid fetching the tenant profile four times in update(TenantId).
Upgrade LangChain4j to 1.15.1-TB1 and migrate GitHub Models to the
official OpenAI integration (OpenAiOfficialChatModel), enabling JSON
Schema (strict) structured output for GitHub Models and exposing JSON
Schema as a response format for Anthropic and Amazon Bedrock.
- Rename model capability flags to supportsSchemalessJsonOutput() and
supportsJsonSchemaOutput(), declared explicitly per provider
- Push response-format capability checks into TbResponseFormat and
simplify response-format handling in TbAiNode
- Remove obsolete manual JSON escaping for GitHub Models, which now
double-escaped requests under the official OpenAI SDK
- Drop the now-obsolete opennlp-tools version pin (the upgraded
LangChain4j fork ships opennlp-tools >= 2.5.9)
- Update the AI rule node UI to offer response formats per provider,
refresh available model lists, and migrate toggle options to @if
- Expand configurer tests to verify per-provider model configuration
SimpleCalculatedFieldState.formatResult cast the double result down to
int via TbUtils.toInt when "Decimals by default" was 0 (the UI default).
BigDecimal.intValue() returns only the low-order 32 bits, so a sum
above ~2.1B wrapped to a negative number (e.g. 3,980,173,734 ->
-314,793,562).
Add TbUtils.toLong(double) alongside toInt (toInt is left untouched to
preserve TBEL script behavior), switch the CF precision=0 path to it,
and add a Long branch in createResultJson so the JSON node is emitted
as a numeric long
- Group tomcat, commons-lang3 version properties under spring-boot.version
- Drop thymeleaf override (PE-only dependency, not present in CE)
- Drop lz4 plumbing: kafka-clients 3.9.2 and cassandra-all 5.0.7 now transitively ship at.yawk.lz4:lz4-java, making the Dec 2025 CVE hack obsolete
AbstractTbQueueConsumerTemplate.poll() returned emptyList() immediately
when partitions was empty, bypassing both doPoll() and the secondary
sleep guard (which is also skipped for backends that report
isLongPollingSupported()==true, e.g. Kafka). The result was a
permanent CPU-burning loop on consumers whose partition assignment
ended up empty after a rebalance cascade — observed on 26
ie-downlink-consumer threads (~244% total CPU) until container restart.
Route the empty-partition path through sleepAndReturnEmpty() so the
caller honors durationInMillis regardless of long-polling support.
Moved NioEventLoopGroup allocations into the try block so that a
constructor failure for the second group no longer leaks the first.
Channel close failures during cleanup now attach via addSuppressed
instead of replacing the original BindException. Narrowed the outer
catch from Throwable to Exception, removing the brittle (Error) cast
that would have masked any direct Throwable subclass.
Creates missing system images from application/src/main/data/resources/images
during LTS patch startup, mirroring the upgrade-path loadSystemResources logic.
Existing system images in the DB are left untouched.
Prevents UnrecognizedPropertyException during rolling upgrades when a
newer node writes a cached entity with an added field and an older node
reads it back. The Redis-backed TbJsonRedisSerializer now uses
JacksonUtil.IGNORE_UNKNOWN_PROPERTIES_JSON_MAPPER instead of the strict
OBJECT_MAPPER used by JacksonUtil.fromBytes.
The alreadyProcessed() method used separate get() and put() calls on the
local cache, allowing concurrent threads in the notification executor pool
to both read null and bypass deduplication, creating duplicate notifications.
Replace with a single compute() call that atomically checks and updates
the cache entry, preventing the race between concurrent trigger processing.
Also fix: discard external cache timestamps that are more than 1 hour in
the future (clock skew protection), and avoid reading back from the SOFT
ref local cache when writing to external cache (GC could null it out).
Viacheslav Klimov
Andrii Landiak
Oleksandra Matviienko
dashevchenko
Dmytro Skarzhynets
Volodymyr Babak
Sergey Matvienko