{
  "date": "2026-10-24",
  "previous_revision": "9ee3080f",
  "previous_revision_note": "Integer COLLECT candidate retained by PR479; this ARRAY_AGG increment merges after it. Identical benchmark/input source is copied to that checkout.",
  "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, released Flink handler and RAW LinkedList accumulators; DISTINCT uses managed DataViews. Production changes are planner/version admission and existing output-boundary classification."
  },
  "configuration": {
    "flink": "2.2.1",
    "rows": 50000,
    "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": "Every eighth input value isNULL; key7mod8groups contain onlyNULL elements under defaultRESPECTNULLS.",
    "warmups_per_engine": 2,
    "trials_per_engine": 3,
    "test_heap_gib": 8,
    "jvm_accumulator_cache_entries": 0,
    "key_prefix_note": "Default16 prefix chars unused forINTkeys.",
    "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",
    "value_types": [
      "BIGINT",
      "INT via CAST(v AS INT) before grouping"
    ],
    "array_length_note": "The8-value input domain does not cap ARRAY length: respect-NULLs arrays grow to781/782 elements at64groups and12/13at4096groups."
  },
  "trials": [
    {
      "phase": "previous",
      "groups": 64,
      "shape": "array_agg",
      "element_type": "BIGINT",
      "flink_seconds": [
        0.552738849,
        0.507870178,
        0.663671289
      ],
      "native_seconds": [
        0.517952953,
        0.656764508,
        0.557445921
      ]
    },
    {
      "phase": "previous",
      "groups": 64,
      "shape": "array_agg_int",
      "element_type": "INT",
      "flink_seconds": [
        0.558824948,
        0.528121578,
        0.55175267
      ],
      "native_seconds": [
        0.53539401,
        0.654646231,
        0.67115133
      ]
    },
    {
      "phase": "previous",
      "groups": 4096,
      "shape": "array_agg",
      "element_type": "BIGINT",
      "flink_seconds": [
        0.189795898,
        0.171819386,
        0.102443112
      ],
      "native_seconds": [
        0.114341029,
        0.180525437,
        0.098210255
      ]
    },
    {
      "phase": "previous",
      "groups": 4096,
      "shape": "array_agg_int",
      "element_type": "INT",
      "flink_seconds": [
        0.157129123,
        0.099901388,
        0.085930588
      ],
      "native_seconds": [
        0.179144672,
        0.154373263,
        0.208797195
      ]
    },
    {
      "phase": "final",
      "groups": 64,
      "shape": "array_agg",
      "element_type": "BIGINT",
      "flink_seconds": [
        0.612707224,
        0.590128554,
        0.530636916
      ],
      "native_seconds": [
        0.615159676,
        0.682652485,
        0.695289212
      ]
    },
    {
      "phase": "final",
      "groups": 64,
      "shape": "array_agg_int",
      "element_type": "INT",
      "flink_seconds": [
        0.631507186,
        0.530444687,
        0.526918269
      ],
      "native_seconds": [
        0.659266675,
        0.630828152,
        0.575246956
      ]
    },
    {
      "phase": "final",
      "groups": 4096,
      "shape": "array_agg",
      "element_type": "BIGINT",
      "flink_seconds": [
        0.179081935,
        0.170332845,
        0.177614215
      ],
      "native_seconds": [
        0.223838682,
        0.187168255,
        0.109108191
      ]
    },
    {
      "phase": "final",
      "groups": 4096,
      "shape": "array_agg_int",
      "element_type": "INT",
      "flink_seconds": [
        0.094891903,
        0.128949309,
        0.099210414
      ],
      "native_seconds": [
        0.200834703,
        0.112457002,
        0.104144998
      ]
    }
  ],
  "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": {
      "array_agg_int-keys4096-native": {
        "total_samples": 25319,
        "inclusive_method_samples": {
          "NativeColumnarJvmGroupOperator.ingest": 1047,
          "NativeColumnarJvmGroupOperator.append": 255,
          "ArrayAggFunction.getValue": 158,
          "ArrayWriter.doWrite": 55,
          "ArrayDataSerializer.copy": 114,
          "GroupAggFunction.processElement": 902,
          "LinkedList.toArray": 156
        },
        "top_self_samples": {
          "_raw_spin_unlock_irqrestore_[k]": 1857,
          "IndexSetIterator::advance_and_next": 561,
          "PhaseChaitin::Split": 448,
          "/usr/lib/x86_64-linux-gnu/libc.so.6": 395,
          "PhaseChaitin::gather_lrg_masks": 352,
          "PhaseLive::compute": 329,
          "G1ParScanThreadState::trim_queue_to_threshold": 323,
          "PhaseChaitin::interfere_with_live": 295,
          "PhaseIdealLoop::build_loop_late": 240,
          "PhaseChaitin::elide_copy": 237
        }
      },
      "array_agg-keys64-native": {
        "total_samples": 15403,
        "inclusive_method_samples": {
          "NativeColumnarJvmGroupOperator.ingest": 5979,
          "NativeColumnarJvmGroupOperator.append": 1326,
          "ArrayAggFunction.getValue": 3999,
          "ArrayWriter.doWrite": 569,
          "ArrayDataSerializer.copy": 1288,
          "GroupAggFunction.processElement": 5899,
          "LinkedList.toArray": 3996
        },
        "top_self_samples": {
          "java/util/LinkedList.toArray_[i]": 3940,
          "java/lang/Long.valueOf_[i]": 639,
          "_raw_spin_unlock_irqrestore_[k]": 598,
          "org/apache/flink/table/data/GenericRowData.<init>_[i]": 405,
          "IndexSetIterator::advance_and_next": 198,
          "/usr/lib/x86_64-linux-gnu/libc.so.6": 180,
          "G1ParScanThreadState::trim_queue_to_threshold": 176,
          "PhaseChaitin::Split": 171,
          "clear_page_erms_[k]": 170,
          "jlong_disjoint_arraycopy": 158
        }
      }
    },
    "limitations": "inclusive counts overlap; not allocation counts/copied bytes, and profiled times do not enter trial tables"
  },
  "scope": {
    "admitted": "Exact released2.2 ARRAY_AGG overINT/BIGINT, scalar inputs andnon-character output fields; defaultRESPECTNULLS/IGNORENULLS,DISTINCT/FILTER/global/mixedCOLLECT; verifiednullableprimitivearrayTopN/UNNEST consumers.",
    "retained_fallback": [
      "richer ARRAY_AGG element types",
      "character grouping/output fields",
      "mini-batch/async-state/approximate/ordered/specialized forms",
      "sensitive/unverified represented state or writer boundaries",
      "updating MAP/MULTISET retained payload extension#437"
    ],
    "not_applicable": "ARRAY_AGG absent onreleasedFlink1.18.1;compat helperreturnsfalse andruntimechecks explicitlyskipthere."
  },
  "conclusion": "All native medians exceed paired stock; BIGINT medians also exceed previous fallback while INT medians improve previous fallback. Ranges overlap and small-array jobs are startup dominated. No acceleration claim. Two profiles identify linked-list result snapshots, copying/boxing and startup/compilation; no safe small shortcut was identified. Accepted collection tradeoff; further performance in#239.",
  "validation": {
    "flink2.2.1": {
      "passed": 219,
      "skipped": 0
    },
    "flink1.18.1": {
      "passed": 185,
      "skipped": 34
    },
    "coverage": [
      "consumed NativeJvmGroup and per-key ordered change histories",
      "INT/BIGINT, duplicates, NULL and validNOTNULL elements viaCOALESCE",
      "RESPECT/IGNORENULLS and allNULL/global/emptyinput results",
      "DISTINCT/FILTER/mixedCOLLECT/scalars/formatting",
      "heap/nativeRocksDB2nonemptyprefixrestores andmemoryrelease",
      "consumed native producer+TopN/UNNEST",
      "richer/character schemas anddisabledgroupfallback",
      "ARRAY_AGG explicitN/A on1.18",
      "existing scalarUDAF,COLLECT,MULTISETcast,UNNEST andsink/producerguards"
    ]
  }
}