{
  "date": "2026-10-24",
  "previous_revision": "bbde17c4",
  "previous_revision_note": "The MULTISET increment tree is retained as an ancestor of PR477; this COLLECT increment merges after it. Benchmark/input code is identical in both trees.",
  "hardware": {
    "cpu": "Intel Core i7-12650H",
    "os": "Linux x86-64 / WSL2",
    "jdk": "OpenJDK17.0.20.1"
  },
  "native": {
    "profile": "release",
    "features": [
      "mimalloc",
      "rocksdb-state"
    ],
    "sha256": "b574491f08427c8b128107d6286cb403421a0bafbdba7a0e5cdd8fea30562d16",
    "production_rust_changes": false,
    "operator_note": "Existing columnar Java/JVM grouped runtime, generated released Flink handler and managed MapViews; production changes are planner admission and representation-boundary gates."
  },
  "configuration": {
    "flink": "2.2.1",
    "rows": 200000,
    "parallelism": 1,
    "key_type": "INT",
    "value_type": "BIGINT",
    "groups": [
      64,
      4096
    ],
    "distinct_values_per_non_null_group": 8,
    "key": "row modulo groups",
    "value": "(row/groups) modulo8; NULL when row modulo8=7",
    "null_note": "Groups whose key modulo8=7 receive onlyNULL values and emit empty multisets. Other groups cycle through8values.",
    "warmups_per_engine": 2,
    "trials_per_engine": 3,
    "test_heap_gib": 8,
    "jvm_accumulator_cache_entries": 0,
    "key_prefix_note": "Default16 prefix characters are unused for INT keys.",
    "boundary": "planning, submission, runtime Row source, keyed state, both row/Arrow transposes and blackhole row sink; cluster setup and initial explain excluded",
    "schedule": "alternating stock/native per JVM; separate JVMs per cardinality and phase; no concurrent tests, builds or benchmarks during measurements"
  },
  "trials": [
    {
      "phase": "previous",
      "groups": 64,
      "flink_seconds": [
        0.476706771,
        0.45490514,
        0.316043047
      ],
      "native_seconds": [
        0.333299805,
        0.412596248,
        0.325387721
      ]
    },
    {
      "phase": "previous",
      "groups": 4096,
      "flink_seconds": [
        0.421953327,
        0.42436506,
        0.458834734
      ],
      "native_seconds": [
        0.383962644,
        0.337086671,
        0.32316423
      ]
    },
    {
      "phase": "final",
      "groups": 64,
      "flink_seconds": [
        0.44061837,
        0.428994847,
        0.311923938
      ],
      "native_seconds": [
        0.415787529,
        0.344032549,
        0.329836623
      ]
    },
    {
      "phase": "final",
      "groups": 4096,
      "flink_seconds": [
        0.422888683,
        0.395371025,
        0.3864995
      ],
      "native_seconds": [
        0.472422155,
        0.374743265,
        0.382119947
      ]
    }
  ],
  "cpu_profiles": {
    "passes": 2,
    "closed": true,
    "event": "cpu",
    "interval": "1ms",
    "target_seconds_per_capture": 8,
    "profiler": "async-profiler4.6",
    "boundary": "one unprofiled native-only complete-job warmup; profile native-only jobs until the8second target completes, including planning/compilation/allJVMthreads; stop after currentjob",
    "captures": {
      "collect-keys4096-native": {
        "total_samples": 16940,
        "inclusive_method_samples": {
          "NativeColumnarJvmGroupOperator.ingest": 4276,
          "NativeColumnarJvmGroupOperator.append": 1071,
          "CollectAggFunction.getValue": 913,
          "MapWriter.doWrite": 366,
          "MapDataSerializer.toBinaryMap": 290,
          "GroupAggFunction.processElement": 4034
        },
        "top_self_samples": {
          "org/apache/flink/core/memory/MemorySegment.equalTo_[i]": 739,
          "_raw_spin_unlock_irqrestore_[k]": 694,
          "java/util/HashMap.putVal_[i]": 430,
          "IndexSetIterator::advance_and_next": 278,
          "G1ParScanThreadState::trim_queue_to_threshold": 273,
          "clear_page_erms_[k]": 264,
          "java/util/HashMap.valuesToArray_[i]": 255,
          "/usr/lib/x86_64-linux-gnu/libc.so.6": 221,
          "PhaseChaitin::Split": 216,
          "org/apache/flink/table/data/binary/MurmurHashUtils.hashBytesByInt_[i]": 215
        }
      },
      "collect-keys64-native": {
        "total_samples": 19173,
        "inclusive_method_samples": {
          "NativeColumnarJvmGroupOperator.ingest": 4362,
          "NativeColumnarJvmGroupOperator.append": 1264,
          "CollectAggFunction.getValue": 636,
          "MapWriter.doWrite": 460,
          "MapDataSerializer.toBinaryMap": 388,
          "GroupAggFunction.processElement": 4027
        },
        "top_self_samples": {
          "_raw_spin_unlock_irqrestore_[k]": 839,
          "IndexSetIterator::advance_and_next": 398,
          "java/util/HashMap.putVal_[i]": 369,
          "/usr/lib/x86_64-linux-gnu/libc.so.6": 283,
          "PhaseChaitin::Split": 283,
          "clear_page_erms_[k]": 257,
          "PhaseChaitin::gather_lrg_masks": 215,
          "java/util/HashMap.valuesToArray_[i]": 212,
          "java/util/HashMap.keysToArray_[i]": 198,
          "tech/streamfusion/arrow/writers/ArrowFieldWriter.write_[i]": 179
        }
      }
    },
    "limitations": "inclusive counts overlap; not allocation counts/copied bytes, and profiled times do not enter trial tables"
  },
  "scope": {
    "admitted": "INT/BIGINT COLLECT in synchronous single-phase grouping with scalar inputs and non-character output fields; DISTINCT/FILTER/global forms and verified final row-sink/fused Calc",
    "retained_fallback": [
      "other COLLECT element types",
      "character grouping/output fields",
      "nested UDAF results",
      "ARRAY_AGG",
      "mini-batch/async-state/approximate/ordered/specialized forms",
      "unverified state/exchange/ordering/native-writer boundaries",
      "updating MAP/MULTISET retained payload extension#437"
    ]
  },
  "conclusion": "Native medians beat paired stock but exceed prior fallback; ranges overlap and jobs are short. No sustained acceleration claim. Two profiling/optimization passes found grouped processing, map snapshots, output writing and startup costs; no safe small optimization was identified. Accepted bounded collection tradeoff; further performance work belongs to the native SQL backlog#239.",
  "validation": {
    "flink2.2.1": {
      "passed": 203,
      "skipped": 0
    },
    "flink1.18.1": {
      "passed": 185,
      "skipped": 18
    },
    "coverage": [
      "consumed NativeJvmGroup",
      "INT/BIGINT, NULLs/extremes, multiplicities, zero/negative counts, group deletion/reinsertion",
      "DISTINCT/FILTER/mixed COUNT/global grouping",
      "fused formatting",
      "empty global/grouped input with consumed NativeFilter and planned empty aggregate",
      "allinputRowKinds",
      "two nonempty-prefix restores on heap/nativeRocksDB, managed DataViews and native memory release",
      "unsupported element/character schema, disabled JVM group and updating Top-N fallback",
      "existing scalar UDAFs, MULTISET casts, generated UNNEST and sink/producers regressions"
    ]
  }
}