Repository navigation
v4.12.27
Version 4.12.27, 10/30/2026
The lock-step twin of Java 4.12.27. A dead-letter copy keeps its record's key (Fixed 1), a failed Schema Registry request is
retried as Confluent's client retries one (Changed 1), and the schema caches follow the Confluent client's, which retires
schema.registry.cache.ttl (Changed 2). One RustSec lockfile bump is included: libz-sys (Security 1). The Java release's
dependency changes (Tomcat 11.0.27, Spring Boot 4.1.2 and the library round) and the mini-scheduler's job time zone have no
counterpart in this engine. Upgrade notes: a lookup against a registry that never answers now fails after four attempts,
and max.retries: 0 keeps the single attempt (Changed 1). With schema.registry.cache.ttl unset, a newly registered latest
schema version is picked up within 60 seconds instead of 30 minutes; a value still set is passed on as latest.cache.ttl.sec,
with a warning (Changed 2). A dead-letter reprocessor reads the original key from the record (Fixed 1).
Security
- libz-sys 1.1.31 (Increment 192). The
auditgate failed on 2026-10-30 on RUSTSEC-2026-0418 (GHSA-8p3x-q3q3-q3v5),
published on 2026-10-28 againstlibz-sys1.1.29 and patched in 1.1.31: its build script logs the compiler command it
probes with, and withcc1.6.0 or later that command carries the whole build environment, so any secret in it lands in
the build script's output undertarget/.libz-sysreaches the workspace throughrdkafka-sys(minimalist-kafkaand
the two Kafka examples). The lockedccis 1.2.67, so this workspace's builds were not exposed. The lockfile moves that one
crate. Upgrade note: none - a lockfile bump; an application that buildsmercury-minimalist-kafkawith its own lockfile
andcc1.6.0 or later updateslibz-sysitself (cargo update -p libz-sys).
Changed
-
A failed Schema Registry request is retried, as the Java engine's Confluent client retries one (Increment 191). The
registry client made one attempt per lookup, so a registry that was briefly unreachable dead-lettered the next record of every
schema id due a lookup, and failed a publish with asubjectheader. A request answered 408, 429, 500, 502, 503 or 504, or
failed by the network, is now tried again up tomax.retriestimes; before retryn(from 0) the client waits a random time
from 1 ms tomin(retries.wait.ms × 2^n, retries.max.wait.ms), the policy of Confluent'sRetryExecutor, with its keys and
defaults (3, 1000 and 20000) inschema-registry.yml. A not-found id and a refused credential still fail at once. Upgrade
note: a lookup against a registry that never answers now fails after four attempts and at most seven seconds of waiting, each
attempt still bounded by its 10-second timeout;max.retries: 0inschema-registry.ymlkeeps the single attempt. -
The schema caches follow the Confluent client's, and
schema.registry.cache.ttlis retired (Increment 193). The
twin of the Java engine's change (mercury-composable #598), which found that the Java codec's own cache in front of
Confluent's registry client saved no registry request: that client keeps every schema fetched by id for the life of
the process and caches a subject'slatestversion forlatest.cache.ttl.sec(60 seconds). This engine has no
Confluent client, so its caches now do what that client does: a schema by id is cached for the life of the process
(bounded to 1000 ids) instead of 30 minutes after each fetch, andlatestforlatest.cache.ttl.secin
schema-registry.yml(default 60; 0 = not cached). The pinned-version cache is unchanged. Upgrade note: with
neither key set, a newly registeredlatestversion is picked up within 60 seconds instead of 30 minutes, and a
registry outage can affect only a schema id's first lookup after a start. An application that sets
schema.registry.cache.ttl(orschema.registry.consumer.cache.ttl) keeps its value forlatest: it is passed on
aslatest.cache.ttl.sec, with a warning at startup, unless the template sets that key; move it to
schema-registry.ymlaslatest.cache.ttl.sec(seconds).
Fixed
- The dead-letter copy keeps the record key (Increment 191). The twin of the Java engine's fix (mercury-composable #597). The
flow adapter wrote a failed record to itsdlq-topicwithout its key, so a reprocessor lost the key and a keyed record landed
on a random dead-letter partition.KafkaRequestPublisher::publish_keyedtakes the key, and the dead-letter write passes the
original record's key bytes. Upgrade note: none - a keyed record's dead-letter copy now lands on the partition its key
hashes to (murmur2, as the Java producer hashes it).
Documentation
- The Minimalist Kafka guide catches up with the Java guide's application-builder corrections (Increment 191).
simple.kafka.notificationpublishes without a key, so per-entity ordering takes an explicitpartitionheader (the guide had
implied keyed publishes); the dead-letter copy keeps the key; and a registry outage dead-letters a record once the client's
retries are spent - for the first record of each schema id after a start (Changed 2) - so everyschema.enabledbinding
should have adlq-topic. The configuration reference, the bundled
schema-registry.ymland the port spec (§7 item 13) list the retry keys.