Current behavior
CI and the upstream suite already run both Flink lines. What remains is the release half: the
release workflow publishes one line. The Flink 1.18 benchmark is already recorded in #198.
Shipped
- The CI Java matrix runs the full suite on both lines, and uploads per-line portable SQL audit
outcomes.
- All nine tagged connector and format module suites run on both lines.
- Lake module CI runs Paimon Parquet and ORC on both lines; Delta stays on 2.2.
- Image integration builds the optimized payload and runs the Session-cluster smoke test on both
lines, including bin/check-artifacts.sh --flink-line.
- The upstream suite matrix runs
runtime, formats, parquet, orc, kafka, paimon and
state on both lines, with delta excluded on 1.18 and a separate Delta 3.3 / Flink 1.18 host
audit job. Each run records result totals and separates native execution, expected fallback and
cases with no execution contract.
- The loader is tested on both lines with real planner installation.
The optimized release+mimalloc Flink 1.18 Nexmark comparison shipped in #198 and is linked from
docs/benchmarks.md: recorded method and results.
It keeps the unchanged rowwise generator, blackhole sink and both transposes. All ten measured
queries are slower than stock (0.28–0.88x); the existing Kafka headline table is unchanged.
Remaining
Publication of the 1.18 line stays blocked on #188 regardless of this issue; passing the CI and
upstream matrix does not establish real-cluster cross-line upgrade support.
Done when
The release workflow publishes and validates both lines' artifact sets. The required release-build
1.18 benchmark and its methodology are already recorded.
Split from #182.
Current behavior
CI and the upstream suite already run both Flink lines. What remains is the release half: the
release workflow publishes one line. The Flink 1.18 benchmark is already recorded in #198.
Shipped
outcomes.
lines, including
bin/check-artifacts.sh --flink-line.runtime,formats,parquet,orc,kafka,paimonandstateon both lines, withdeltaexcluded on 1.18 and a separate Delta 3.3 / Flink 1.18 hostaudit job. Each run records result totals and separates native execution, expected fallback and
cases with no execution contract.
The optimized release+mimalloc Flink 1.18 Nexmark comparison shipped in #198 and is linked from
docs/benchmarks.md: recorded method and results.It keeps the unchanged rowwise generator, blackhole sink and both transposes. All ten measured
queries are slower than stock (0.28–0.88x); the existing Kafka headline table is unchanged.
Remaining
line-qualified archives and image tags, per-line source and Javadoc artifacts, and
bin/check-artifacts.shagainst each line's artifact set. The workflow currently performs onedeployment and assembles unsuffixed archives.
docs/releasing.mdwith the completed automated dual-line publication process. Theexperimental build/package commands and benchmark methodology are already documented.
Publication of the 1.18 line stays blocked on #188 regardless of this issue; passing the CI and
upstream matrix does not establish real-cluster cross-line upgrade support.
Done when
The release workflow publishes and validates both lines' artifact sets. The required release-build
1.18 benchmark and its methodology are already recorded.
Split from #182.