{
  "date": "2026-11-11",
  "status": "Bounded and continuous MOR complete release pipelines qualify at one million rows against stock and previous SF. Small cases retain startup regressions. Continuous qualification uses the public SDK engine-record iterator and validates every operation kind. Earlier internal-tag timings remain historical diagnostics.",
  "platform": "Apple M4 (10 cores), macOS 15.3.1, Temurin 17",
  "libraries": "Release engine and Parquet; canonical Hudi 1.2.1 Flink 1.18 bundle",
  "query": "INSERT INTO output SELECT id,name,ts FROM h",
  "fixture": "Stock-written MOR Parquet base with 25% full-row event-time updates in Parquet logs; nullable VARCHAR and BIGINT ordering; immutable COW append target",
  "resources": {
    "parallelism": 1,
    "heap_gib": 8,
    "task_offheap_gib": 1,
    "decoder_batch_rows": 4096,
    "merge_budget_mib": 256
  },
  "warmups": 2,
  "trials": 5,
  "order": "Rotating stock/previous SF/native",
  "previous_sf": "Stock Hudi MOR source followed by native COW append sink",
  "native": "Java Hudi merging Arrow references over Rust Parquet decode, Arrow gather and native COW append sink",
  "timer": "SQL execute through committed job completion, including planning/admission, Java storage I/O, JNI, merging, Parquet encoding and commits. Fixture writes, EXPLAIN and stock aggregate readback are outside timing.",
  "validation": "Every trial verifies stock-read count, key sum, event-version sum and nullable name count. Native plan requires Java Hudi merger and native append sink without a payload transpose.",
  "measurements": [
    {
      "rows": 200000,
      "updates": 50000,
      "phase": "Initial Java Arrow reader",
      "qualification": "Correctness passed. Native remains slower than previous SF. Profiling and optimization pending.",
      "trial_ms": {
        "stock": [
          759.49925,
          710.953458,
          705.583208,
          673.299542,
          681.906125
        ],
        "previous_sf": [
          553.691875,
          669.787917,
          596.052541,
          612.182458,
          594.117375
        ],
        "native": [
          612.063416,
          627.781375,
          636.614417,
          676.320959,
          670.687125
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 705.583208,
          "min": 673.299542,
          "max": 759.49925
        },
        "previous_sf": {
          "median": 596.052541,
          "min": 553.691875,
          "max": 669.787917
        },
        "native": {
          "median": 636.614417,
          "min": 612.063416,
          "max": 676.320959
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 61,
        "external_build_processes": []
      }
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "phase": "Initial Java Arrow reader",
      "qualification": "Correctness passed. Native remains slower than previous SF. Profiling and optimization pending.",
      "trial_ms": {
        "stock": [
          1540.765,
          1773.91625,
          1810.3445,
          1800.860292,
          1960.241708
        ],
        "previous_sf": [
          1070.759,
          1013.329625,
          1120.760958,
          1159.590417,
          1482.965541
        ],
        "native": [
          1211.402417,
          1317.065584,
          1262.389542,
          1311.2355,
          1204.200666
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 1800.860292,
          "min": 1540.765,
          "max": 1960.241708
        },
        "previous_sf": {
          "median": 1120.760958,
          "min": 1013.329625,
          "max": 1482.965541
        },
        "native": {
          "median": 1262.389542,
          "min": 1204.200666,
          "max": 1317.065584
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 154,
        "external_build_processes": []
      }
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "phase": "ProjectedRowData wrapper and hoisted selection function",
      "qualification": "Correctness passed. Per-row projection wrappers remain slower than previous SF. Next change defers terminal projection to the Arrow batch gather.",
      "trial_ms": {
        "stock": [
          1662.764292,
          1969.176541,
          2026.9165,
          2178.946459,
          2141.825291
        ],
        "previous_sf": [
          1258.632417,
          1198.356,
          1427.687416,
          1282.896875,
          1424.768666
        ],
        "native": [
          1303.684125,
          1473.504083,
          1517.795958,
          1713.758959,
          1556.079375
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 2026.9165,
          "min": 1662.764292,
          "max": 2178.946459
        },
        "previous_sf": {
          "median": 1282.896875,
          "min": 1198.356,
          "max": 1427.687416
        },
        "native": {
          "median": 1517.795958,
          "min": 1303.684125,
          "max": 1713.758959
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 170,
        "external_build_processes": []
      }
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "phase": "Terminal projection deferred to Arrow batch gathering",
      "qualification": "Correctness passed. Native median remains 14.2% slower than previous SF. Next change reuses stock fetch-thread decoding/merging and logical-row offsets, with transient base cursors.",
      "trial_ms": {
        "stock": [
          1538.488333,
          1442.145791,
          1566.281959,
          1452.367542,
          1492.751209
        ],
        "previous_sf": [
          925.378417,
          936.600709,
          1030.924375,
          942.76425,
          1019.725834
        ],
        "native": [
          1165.519167,
          1032.06675,
          1076.529542,
          1068.344292,
          1135.462334
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 1492.751209,
          "min": 1442.145791,
          "max": 1566.281959
        },
        "previous_sf": {
          "median": 942.76425,
          "min": 925.378417,
          "max": 1030.924375
        },
        "native": {
          "median": 1076.529542,
          "min": 1032.06675,
          "max": 1165.519167
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 169,
        "external_build_processes": []
      }
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "phase": "Original fetch-thread decoding/merging, logical-row offsets and transient base cursors",
      "qualification": "Correctness and complete-job comparison passed versus stock and previous SF. All 21 warmup/trial outputs matched stock aggregate readback. Sampled host had no competing builds.",
      "trial_ms": {
        "stock": [
          1631.913125,
          1848.654208,
          1811.815459,
          1883.697458,
          1789.649084
        ],
        "previous_sf": [
          957.676625,
          1094.629458,
          1325.646584,
          1137.40675,
          1349.519292
        ],
        "native": [
          1047.42825,
          1078.449042,
          1031.734333,
          1077.033875,
          1121.89175
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 1811.815459,
          "min": 1631.913125,
          "max": 1883.697458
        },
        "previous_sf": {
          "median": 1137.40675,
          "min": 957.676625,
          "max": 1349.519292
        },
        "native": {
          "median": 1077.033875,
          "min": 1031.734333,
          "max": 1121.89175
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 154,
        "external_build_processes": []
      }
    },
    {
      "rows": 200000,
      "updates": 50000,
      "phase": "Original fetch-thread decoding/merging, logical-row offsets and transient base cursors",
      "workload": "bounded MOR to COW append",
      "qualification": "All 21 warmup/trial outputs verified. No competing builds sampled. Report small/overlapping differences and previous-SF regressions explicitly.",
      "trial_ms": {
        "stock": [
          697.9435,
          808.675125,
          825.498792,
          839.450333,
          903.044708
        ],
        "previous_sf": [
          649.669208,
          622.582417,
          702.748209,
          630.5885,
          778.323709
        ],
        "native": [
          730.642417,
          563.712167,
          693.711792,
          606.692083,
          776.075375
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 825.498792,
          "min": 697.9435,
          "max": 903.044708
        },
        "previous_sf": {
          "median": 649.669208,
          "min": 622.582417,
          "max": 778.323709
        },
        "native": {
          "median": 693.711792,
          "min": 563.712167,
          "max": 776.075375
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 63,
        "external_build_processes": []
      }
    },
    {
      "rows": 200000,
      "updates": 50000,
      "phase": "Original fetch-thread decoding/merging, logical-row offsets and transient base cursors",
      "workload": "continuous Calc to row collector",
      "qualification": "Historical data-only diagnostic before the base/log row-kind regression was exposed. Counts, values and deletes passed, but internal UPDATE_AFTER tags differed from the stock engine-record INSERT tags. Superseded by the public-engine-iterator measurements with complete kind checks.",
      "trial_ms": {
        "stock": [
          1254.048625,
          1302.465292,
          1269.578875,
          1239.659458,
          1187.220541
        ],
        "previous_sf": [
          1161.471833,
          1147.809125,
          1227.665666,
          1410.049125,
          1235.881083
        ],
        "native": [
          1153.366375,
          1187.714541,
          1362.591208,
          1254.419333,
          1203.276166
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 1254.048625,
          "min": 1187.220541,
          "max": 1302.465292
        },
        "previous_sf": {
          "median": 1227.665666,
          "min": 1147.809125,
          "max": 1410.049125
        },
        "native": {
          "median": 1203.276166,
          "min": 1153.366375,
          "max": 1362.591208
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 47,
        "external_build_processes": []
      },
      "deletes": 2000,
      "query": "SELECT id+1,name,ts FROM h",
      "previous_sf": "Stock Hudi source -> Row-to-Arrow -> native Calc -> Arrow-to-row -> collector",
      "native": "Java Hudi merging Arrow references with Rust Parquet decode/gather -> native Calc -> Arrow-to-row -> collector",
      "timer": "SQL execute through consumption of all rows from the completed initial commit range, including planning, admission, discovery, I/O, JNI, merge and final row conversion. Fixture writes, EXPLAIN, cancellation and task termination are outside timing.",
      "validation": "Each trial checks key sum, version sum, non-null name count, name hash sum and exact delete count. Native plan requires Java Hudi and NativeCalc without a RowData-to-Arrow edge."
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "phase": "Original fetch-thread decoding/merging, logical-row offsets and transient base cursors",
      "workload": "continuous Calc to row collector",
      "qualification": "Historical data-only diagnostic before the base/log row-kind regression was exposed. Counts, values and deletes passed, but internal UPDATE_AFTER tags differed from the stock engine-record INSERT tags. Superseded by the public-engine-iterator measurements with complete kind checks.",
      "trial_ms": {
        "stock": [
          4066.695542,
          3971.291,
          3967.016667,
          4111.747916,
          4173.719458
        ],
        "previous_sf": [
          3873.697875,
          3773.955917,
          3951.049875,
          4381.731625,
          4011.466167
        ],
        "native": [
          4098.106416,
          3964.219209,
          3733.753542,
          3777.44875,
          4143.207917
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 4066.695542,
          "min": 3967.016667,
          "max": 4173.719458
        },
        "previous_sf": {
          "median": 3951.049875,
          "min": 3773.955917,
          "max": 4381.731625
        },
        "native": {
          "median": 3964.219209,
          "min": 3733.753542,
          "max": 4143.207917
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 110,
        "external_build_processes": []
      },
      "deletes": 10000,
      "query": "SELECT id+1,name,ts FROM h",
      "previous_sf": "Stock Hudi source -> Row-to-Arrow -> native Calc -> Arrow-to-row -> collector",
      "native": "Java Hudi merging Arrow references with Rust Parquet decode/gather -> native Calc -> Arrow-to-row -> collector",
      "timer": "SQL execute through consumption of all rows from the completed initial commit range, including planning, admission, discovery, I/O, JNI, merge and final row conversion. Fixture writes, EXPLAIN, cancellation and task termination are outside timing.",
      "validation": "Each trial checks key sum, version sum, non-null name count, name hash sum and exact delete count. Native plan requires Java Hudi and NativeCalc without a RowData-to-Arrow edge."
    },
    {
      "rows": 200000,
      "updates": 50000,
      "deletes": 2000,
      "phase": "Public Hudi engine-record iterator with original source row kinds; fetch-thread Arrow decoding/merge/gather",
      "workload": "continuous Calc to row collector",
      "query": "SELECT id+1,name,ts FROM h",
      "qualification": "All 21 warmup/trial outputs passed all data, delete and positive-operation-kind checks. Million-row median improves both baselines. Small-run regressions and overlapping ranges are retained.",
      "validation": "Key sum, event-version sum, non-null name count, name hash, exact delete count and zero unexpected positive/retract kinds. Every non-delete is INSERT, matching released SDK Source V2. Native source/Calc plan and absence of a RowData-to-Arrow edge required.",
      "trial_ms": {
        "stock": [
          1042.131958,
          1024.354584,
          993.84375,
          1053.104292,
          1023.671583
        ],
        "previous_sf": [
          995.763291,
          1100.208375,
          1096.192084,
          988.965208,
          978.823833
        ],
        "native": [
          1196.48825,
          1122.436042,
          1034.080334,
          1066.519417,
          1002.535209
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 1024.354584,
          "min": 993.84375,
          "max": 1053.104292
        },
        "previous_sf": {
          "median": 995.763291,
          "min": 978.823833,
          "max": 1100.208375
        },
        "native": {
          "median": 1066.519417,
          "min": 1002.535209,
          "max": 1196.48825
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 38,
        "external_jvm_or_build_processes": [],
        "background_system_peak_cpu": {
          "mds_stores": 33.8,
          "mediaanalysisd": 111.4
        }
      }
    },
    {
      "rows": 1000000,
      "updates": 250000,
      "deletes": 10000,
      "phase": "Public Hudi engine-record iterator with original source row kinds; fetch-thread Arrow decoding/merge/gather",
      "workload": "continuous Calc to row collector",
      "query": "SELECT id+1,name,ts FROM h",
      "qualification": "All 21 warmup/trial outputs passed all data, delete and positive-operation-kind checks. Million-row median improves both baselines. Small-run regressions and overlapping ranges are retained.",
      "validation": "Key sum, event-version sum, non-null name count, name hash, exact delete count and zero unexpected positive/retract kinds. Every non-delete is INSERT, matching released SDK Source V2. Native source/Calc plan and absence of a RowData-to-Arrow edge required.",
      "trial_ms": {
        "stock": [
          3285.416917,
          3174.416125,
          3439.534042,
          3333.109792,
          3532.291792
        ],
        "previous_sf": [
          3115.59725,
          3098.897958,
          3420.946875,
          3746.090416,
          3406.786042
        ],
        "native": [
          3018.856,
          3053.586625,
          3200.648958,
          3327.78625,
          3403.165417
        ]
      },
      "summary_ms": {
        "stock": {
          "median": 3333.109792,
          "min": 3174.416125,
          "max": 3532.291792
        },
        "previous_sf": {
          "median": 3406.786042,
          "min": 3098.897958,
          "max": 3746.090416
        },
        "native": {
          "median": 3200.648958,
          "min": 3018.856,
          "max": 3403.165417
        }
      },
      "host": {
        "exit_code": 0,
        "samples": 88,
        "external_jvm_or_build_processes": [],
        "background_system_peak_cpu": {
          "mds_stores": 63.0,
          "mediaanalysisd": 118.3
        }
      }
    }
  ],
  "profiling": {
    "workload": "Continuous MOR source -> Calc -> row collector, 200k rows, 25% updates, 1% key-only deletes",
    "qualification": "JFR profile diagnostic, excluded from uninstrumented timing claims",
    "trial_ms": {
      "stock": [
        1302.725292,
        1321.773875,
        1371.899917,
        1280.574375,
        1199.108041
      ],
      "previous_sf": [
        1180.224458,
        1174.976959,
        1276.418666,
        1220.735292,
        1158.133667
      ],
      "native": [
        1205.364,
        1197.310958,
        1480.537167,
        1266.642458,
        1152.290709
      ]
    },
    "host": {
      "exit_code": 0,
      "samples": 47,
      "external_build_processes": []
    },
    "hudi_allocation_samples": 119,
    "sampled_allocation_mib": {
      "Cursor.project": 111.5,
      "reader_row_views": 34.19,
      "Cursor.constructor": 26.26,
      "selection_mapping": 21.63
    },
    "finding": "Per-row projection builds new column arrays, field mappings and VectorizedColumnBatch wrappers. Reuse Flink ProjectedRowData with the existing schema mapping and full-batch vector wrappers. Hoist the output-source mapping lambda outside the row loop."
  }
}