Direct JVM temporal bridge costs¶

These measurements accompany the verified table-function coverage. The admitted paths are correctness-verified UDF coverage under the accepted performance exception. They are slower in these standalone jobs; this is not an acceleration claim.

The complete artifact retains all warmup/measured trials, their order, immutable revisions, fixture/native hashes, configuration, scoped correctness counts and sampled host observations. The candidate is 5ee08ce3; the actual previous production revision is f5274a10. Native sources are unchanged, so both revisions use the same freshly built release/mimalloc DSO. Only new test/benchmark fixtures are copied into the previous worktree; its production Java sources remain unchanged.

Method and limits¶

Released Flink 1.18.1 and 2.2.1 use Temurin 17.0.17, one task, a 2 GiB heap and 1,024-row transpose batches on a Darwin arm64 machine with ten logical CPUs and 16 GiB memory. Each case uses two warmup pairs and five alternating measured pairs. Inputs contain the same values, types, NULLs, cardinalities and container widths. SQL setup and symmetric EXPLAIN precede the job timer; submission, callbacks, released conversion, owned writes, row source/blackhole sink and both perimeter transposes remain inside it. Cluster creation/teardown is outside it. Every candidate trial requires both transposes, no fallback and positive table-function input. The previous enabled route requires zero native substitutions/work for these new signatures.

The 262,144-row matrix includes boxed DATE with ARRAY/List, all eight additional temporal bridges with ARRAY, and JDBC DATE/TIMESTAMP with List. A separate 2,097,152-row JDBC TIMESTAMP/List case retains less favorable sustained-processing observations as well as short jobs. Tables show median seconds and full measured ranges; every warmup is retained in the artifact. The two stock columns deliberately retain the independent controls rather than hiding their drift.

Background macOS indexing/analysis consumed substantial sampled CPU. No other Java process was observed during timed series. The larger stock controls drift between JVMs and include outliers; these runs cannot isolate the effect of a conversion optimization. Host observations also have UTC gaps, so process wall duration is not used as throughput. Short-job startup/JIT costs and unfavorable complete-job results are retained. No optimization or performance-only opt-in is introduced by these coverage changes.

Flink 2.2¶

Case / source rows	Stock in candidate JVM	Candidate native	Stock in previous JVM	Previous enabled fallback
boxed DATE / array / 262,144	0.112 (0.108–0.123)	0.151 (0.146–0.155)	0.110 (0.101–0.114)	0.108 (0.104–0.114)
boxed DATE / list / 262,144	0.654 (0.647–0.664)	1.200 (1.195–1.207)	0.653 (0.650–0.669)	0.660 (0.647–0.663)
date-sql:array / 262,144	0.100 (0.099–0.111)	0.197 (0.196–0.200)	0.097 (0.095–0.099)	0.098 (0.094–0.107)
time-integer:array / 262,144	0.099 (0.095–0.108)	0.190 (0.189–0.194)	0.094 (0.093–0.103)	0.092 (0.091–0.103)
time-long:array / 262,144	0.107 (0.097–0.114)	0.192 (0.189–0.202)	0.098 (0.095–0.099)	0.096 (0.092–0.108)
time-sql:array / 262,144	0.105 (0.102–0.113)	0.198 (0.196–0.208)	0.099 (0.098–0.102)	0.101 (0.097–0.108)
timestamp-sql:array / 262,144	0.114 (0.112–0.128)	0.205 (0.203–0.208)	0.108 (0.105–0.109)	0.108 (0.105–0.109)
ltz-integer:array / 262,144	0.100 (0.097–0.101)	0.202 (0.197–0.209)	0.100 (0.096–0.102)	0.097 (0.095–0.101)
ltz-long:array / 262,144	0.100 (0.098–0.102)	0.201 (0.199–0.207)	0.100 (0.097–0.107)	0.100 (0.097–0.102)
ltz-sql:array / 262,144	0.101 (0.098–0.104)	0.198 (0.194–0.203)	0.097 (0.094–0.099)	0.100 (0.096–0.101)
date-sql:list / 262,144	0.622 (0.621–0.632)	2.026 (2.014–2.038)	0.625 (0.623–0.627)	0.626 (0.624–0.886)
timestamp-sql:list / 262,144	0.640 (0.632–0.671)	2.053 (2.037–2.071)	0.641 (0.638–0.652)	0.636 (0.632–0.647)
timestamp-sql:list / 2,097,152	12.117 (12.103–14.893)	23.871 (23.589–27.927)	6.119 (6.086–6.321)	6.142 (6.009–6.181)

Flink 1.18¶

Case / source rows	Stock in candidate JVM	Candidate native	Stock in previous JVM	Previous enabled fallback
boxed DATE / array / 262,144	0.115 (0.112–0.119)	0.153 (0.150–0.153)	0.110 (0.107–0.112)	0.109 (0.102–0.120)
boxed DATE / list / 262,144	0.647 (0.645–0.668)	1.175 (1.167–1.189)	0.663 (0.648–0.683)	0.665 (0.648–0.704)
date-sql:array / 262,144	0.100 (0.094–0.105)	0.198 (0.194–0.207)	0.097 (0.094–0.107)	0.097 (0.094–0.108)
time-integer:array / 262,144	0.094 (0.091–0.102)	0.193 (0.190–0.202)	0.094 (0.090–0.101)	0.097 (0.089–0.102)
time-long:array / 262,144	0.095 (0.092–0.100)	0.196 (0.192–0.204)	0.105 (0.096–0.106)	0.096 (0.094–0.105)
time-sql:array / 262,144	0.099 (0.099–0.108)	0.199 (0.195–0.200)	0.101 (0.096–0.107)	0.100 (0.096–0.106)
timestamp-sql:array / 262,144	0.109 (0.108–0.116)	0.206 (0.206–0.210)	0.109 (0.106–0.110)	0.108 (0.107–0.114)
ltz-integer:array / 262,144	0.097 (0.096–0.106)	0.199 (0.199–0.209)	0.097 (0.092–0.107)	0.095 (0.094–0.098)
ltz-long:array / 262,144	0.096 (0.094–0.098)	0.203 (0.202–0.203)	0.103 (0.098–0.108)	0.104 (0.096–0.107)
ltz-sql:array / 262,144	0.100 (0.098–0.107)	0.204 (0.202–0.212)	0.098 (0.096–0.099)	0.099 (0.098–0.107)
date-sql:list / 262,144	0.632 (0.626–0.634)	1.990 (1.971–1.996)	0.636 (0.634–0.639)	0.631 (0.626–0.634)
timestamp-sql:list / 262,144	0.639 (0.635–0.652)	2.010 (1.999–2.018)	0.696 (0.685–0.701)	0.702 (0.684–0.712)
timestamp-sql:list / 2,097,152	16.908 (12.123–57.638)	24.214 (23.209–73.133)	12.145 (12.137–12.157)	12.152 (12.145–12.166)

Reproduction¶

Prepare the runtime classes with mvn -pl streamfusion-runtime -Pbench test-compile, then run:

SF_BENCHMARK=true mvn -pl streamfusion-runtime -Pbench -Dnative.build.skip=true \
  -Dtest=BoxedDateTableFunctionBenchmark,AlternateTemporalTableFunctionBenchmark \
  -Dsf.testForks=1 '-Dsf.benchmarkJvmArgs=-Xmx2g -Dstreamfusion.exchange.zeroCopyLocal=false' \
  -Dboundary.rows=262144 -Dboundary.warmups=2 -Dboundary.runs=5 \
  surefire:test@default-test

Add ,flink-1.18 to the Maven profile for Flink 1.18.1 and rebuild its classes before timing. For the larger case select AlternateTemporalTableFunctionBenchmark, set -Dtemporal.bridge.benchcases=timestamp-sql:list and -Dboundary.rows=2097152. For previous production, check out f5274a10, copy only the four new fixtures named in the artifact, prepare that revision's classes with the same profile, and set -Dboundary.expectNative=false. Reuse the immutable native library through native.resource.directory/java.library.path when working in a separate checkout; verify its hash and that the production native sources agree. Do not compile or run other tests concurrently with measurements. The retained revision hashes define the measured workload; additional bridge cases added later require their own named measurements.

Additional boxed interval bridges¶

The interval artifact retains all trials for Integer months and Long milliseconds on both releases. It measures immutable fa31713d (integrated as 7c866310) against unchanged f5274a10, with the same released DSO and complete-job boundaries. Each of the 18 selected correctness controls passes per line. The following 262,144-row jobs use ARRAY containers; medians and full measured ranges are seconds.

Flink / case	Stock in candidate JVM	Candidate native	Stock in previous JVM	Previous enabled fallback
2.2 / yearmonth-integer:array	0.100 (0.091–0.102)	0.202 (0.194–0.208)	0.097 (0.085–0.100)	0.095 (0.084–0.098)
2.2 / daytime-long:array	0.100 (0.091–0.109)	0.189 (0.183–0.193)	0.097 (0.090–0.104)	0.098 (0.093–0.100)
1.18 / yearmonth-integer:array	0.098 (0.095–0.099)	0.195 (0.191–0.196)	0.096 (0.096–0.097)	0.096 (0.090–0.098)
1.18 / daytime-long:array	0.091 (0.089–0.101)	0.178 (0.177–0.187)	0.098 (0.090–0.099)	0.098 (0.088–0.104)

Reproduce with the same benchmark command, selecting AlternateTemporalTableFunctionBenchmark and -Dtemporal.bridge.benchcases=yearmonth-integer:array;daytime-long:array (quote the property argument in a shell). Add ,flink-1.18 for that line. These slower native costs remain accepted UDF coverage, not an acceleration claim.

Primitive temporal bridges and custom scalar inference¶

The follow-up artifact records all warmup/measured trials for six primitive argument/Row-result bridges and two custom nested-scalar inference forms. It compares immutable 5fa5ceae with pre-goal canonical f5274a10 using identical fixture sources and the same release/mimalloc DSO. Previous production sources remain unchanged. Timed jobs use the same resources, alternating pairs, row source/blackhole and both transposes as above. The combined correctness selection passes 252 checks on 2.2.1 and 251 plus the existing ARRAY_SORT N/A on 1.18.1, without failures/errors.

These 262,144-row observations retain startup/JIT effects and sampled background activity. Native UDF costs are slower here; these are admitted under the accepted UDF performance exception. Independent stock controls remain separate rather than presenting their drift as an isolated effect. All measurements and unfavorable trials are retained.

Flink / case	Stock in candidate JVM	Candidate native	Stock in previous JVM	Previous enabled fallback
2.2 / inferred explicit	0.373 (0.371–0.375)	0.566 (0.560–0.569)	0.371 (0.368–0.375)	0.374 (0.367–0.374)
2.2 / inferred delegating	0.376 (0.366–0.377)	0.566 (0.563–0.585)	0.376 (0.375–0.398)	0.375 (0.374–0.375)
2.2 / primitive time-int	0.090 (0.084–0.093)	0.171 (0.165–0.173)	0.087 (0.081–0.092)	0.088 (0.085–0.098)
2.2 / primitive time-primitive-long	0.088 (0.081–0.091)	0.172 (0.168–0.174)	0.090 (0.081–0.092)	0.089 (0.080–0.091)
2.2 / primitive ltz-int	0.082 (0.079–0.096)	0.172 (0.162–0.178)	0.087 (0.078–0.091)	0.082 (0.080–0.091)
2.2 / primitive ltz-primitive-long	0.084 (0.083–0.089)	0.170 (0.169–0.178)	0.084 (0.079–0.091)	0.086 (0.081–0.098)
2.2 / primitive yearmonth-int	0.091 (0.081–0.094)	0.164 (0.163–0.166)	0.083 (0.082–0.090)	0.093 (0.086–0.111)
2.2 / primitive daytime-primitive-long	0.085 (0.082–0.094)	0.166 (0.164–0.172)	0.084 (0.082–0.091)	0.082 (0.080–0.092)
1.18 / inferred explicit	0.382 (0.375–0.384)	0.568 (0.565–0.575)	0.370 (0.365–0.372)	0.368 (0.366–0.372)
1.18 / inferred delegating	0.383 (0.379–0.392)	0.567 (0.564–0.568)	0.373 (0.368–0.375)	0.372 (0.370–0.374)
1.18 / primitive time-int	0.088 (0.086–0.090)	0.167 (0.167–0.170)	0.085 (0.083–0.089)	0.088 (0.084–0.093)
1.18 / primitive time-primitive-long	0.090 (0.081–0.091)	0.169 (0.166–0.169)	0.089 (0.080–0.090)	0.088 (0.080–0.090)
1.18 / primitive ltz-int	0.090 (0.088–0.091)	0.172 (0.162–0.174)	0.089 (0.086–0.089)	0.081 (0.079–0.090)
1.18 / primitive ltz-primitive-long	0.088 (0.082–0.091)	0.170 (0.162–0.174)	0.089 (0.081–0.090)	0.090 (0.089–0.091)
1.18 / primitive yearmonth-int	0.091 (0.080–0.101)	0.161 (0.159–0.168)	0.091 (0.081–0.092)	0.090 (0.080–0.091)
1.18 / primitive daytime-primitive-long	0.083 (0.080–0.093)	0.168 (0.162–0.172)	0.090 (0.088–0.097)	0.092 (0.089–0.094)

Reproduce with the same -Pbench command and -Dtest=PrimitiveTemporalTableFunctionBenchmark,InferredNestedScalarBenchmark. Add ,flink-1.18 for that line. For the baseline copy only the six test fixtures listed in the artifact and use -Dboundary.expectNative=false; leave production sources unchanged.

A subsequent NOT NULL metadata check adds three controls without production changes: matching root-NULL sink failure, released non-null-child behavior and successful filtered non-null roots. The final selection passes 255 checks on 2.2.1 and 254 plus the existing ARRAY_SORT N/A on 1.18.1. The artifact keeps the original measurement revision and records extended validation separately; the timed fixture bodies are unchanged.

The follow-up measurement used a local pre-rebase commit. Production sources are identical to the rebased revision recorded in the artifact. For byte-for-byte fixture reproduction, check out that rebased revision and apply the measured-fixture patch before copying fixtures into the baseline and compiling. It removes only the later NOT NULL controls; verify the recorded fixture hashes. This preserves public reproduction after rebasing onto the parent coverage merge without changing measured production code.

The existing ARRAY/MAP inference controls additionally require positive native Calc work for delegating default converters, while retaining the internal MapData fallback. Their complete 38-test regression selection is validated separately on each released line; this changes test expectations and coverage documentation, not production code or timed fixture bodies.

The fixture reversal is stored as JSON to retain exact unified-diff context without introducing trailing whitespace into the documentation sources. Decode its patch field to a temporary file, then use git apply and verify the fixture SHA-256 from the artifact.