ila · Oct 30, 2026 · Oct 30, 2026 · Oct 30, 2026 · Nov 1, 2026 · Nov 1, 2026
diff --git a/.github/workflows/CodeQuality.yml b/.github/workflows/CodeQuality.yml
@@ -0,0 +1,52 @@
+name: Code Quality
+
+on:
+  workflow_call:
+
+jobs:
+  check:
+    name: ${{ matrix.name }}
+    runs-on: ${{ matrix.runner }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: Format Check
+            runner: ubuntu-22.04
+            pip_packages: cmake-format 'black==24.*' cxxheaderparser pcpp 'clang_format==11.0.1'
+            pip_flags: ''
+            tidy_binary: ''
+            target: format-check
+          - name: Tidy Check
+            runner: ubuntu-24.04
+            pip_packages: '"pybind11[global]"'
+            pip_flags: --break-system-packages
+            tidy_binary: clang-tidy-18
+            target: tidy-check
+    env:
+      CC: gcc
+      CXX: g++
+      GEN: ninja
+      TIDY_THREADS: 4
+      TIDY_BINARY: ${{ matrix.tidy_binary }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          fetch-depth: 0
+          submodules: recursive
+
+      - name: Install Python check tools
+        shell: bash
+        run: |
+          # Ninja and clang-tidy-18 are already on the hosted Ubuntu image.
+          # Avoid redundant apt installs that fail on the image's package mirrors.
+          sudo pip3 install ${{ matrix.pip_packages }} ${{ matrix.pip_flags }}
+
+      - name: Run check
+        run: |
+          if [ '${{ matrix.target }}' = 'tidy-check' ]; then
+            ninja --version
+            "$TIDY_BINARY" --version
+          fi
+          make ${{ matrix.target }}
diff --git a/.github/workflows/MainDistributionPipeline.yml b/.github/workflows/MainDistributionPipeline.yml
@@ -14,17 +14,12 @@ concurrency:
 jobs:
   duckdb-stable-build:
     name: Build extension binaries
-    uses: duckdb/extension-ci-tools/.github/workflows/_extension_distribution.yml@v1.5.4
+    uses: duckdb/extension-ci-tools/.github/workflows/_extension_distribution.yml@ebdcc1961780b0ce1272a0d7637dc58447c34885
     with:
-      duckdb_version: v1.5.4
-      ci_tools_version: v1.5.4
+      duckdb_version: v2.0.0
+      ci_tools_version: ebdcc1961780b0ce1272a0d7637dc58447c34885
       extension_name: privacy
 
   code-quality-check:
     name: Code Quality Check
-    uses: duckdb/extension-ci-tools/.github/workflows/_extension_code_quality.yml@v1.5.4
-    with:
-      duckdb_version: v1.5.4
-      ci_tools_version: v1.5.4
-      extension_name: privacy
-      format_checks: 'format;tidy'
+    uses: ./.github/workflows/CodeQuality.yml
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -3,10 +3,11 @@ cmake_minimum_required(VERSION 3.5)
 # Set extension name here
 set(TARGET_NAME privacy)
 
-# DuckDB's extension distribution supports vcpkg. As such, dependencies can be added in ./vcpkg.json and then
-# used in cmake with find_package. Feel free to remove or replace with other dependencies.
-# Note that it should also be removed from vcpkg.json to prevent needlessly installing it..
-#find_package(OpensSL REQUIRED)
+# DuckDB's extension distribution supports vcpkg. As such, dependencies can be
+# added in ./vcpkg.json and then used in cmake with find_package. Feel free to
+# remove or replace with other dependencies. Note that it should also be removed
+# from vcpkg.json to prevent needlessly installing it.. find_package(OpensSL
+# REQUIRED)
 
 set(EXTENSION_NAME ${TARGET_NAME}_extension)
 set(LOADABLE_EXTENSION_NAME ${TARGET_NAME}_loadable_extension)
@@ -15,15 +16,16 @@ project(${TARGET_NAME})
 include_directories(src/include)
 include_directories(benchmark/include)
 
-# Option to disable approximate sum algorithm (use exact cascading instead)
-# By default, pac_sum/pac_avg use the approx algorithm (16-bit cascading counters)
+# Option to disable approximate sum algorithm (use exact cascading instead) By
+# default, pac_sum/pac_avg use the approx algorithm (16-bit cascading counters)
 # Set PAC_NOAPPROX=ON to use exact cascading instead
-option(PAC_NOAPPROX "Disable approximate sum algorithm, use exact cascading" OFF)
+option(PAC_NOAPPROX "Disable approximate sum algorithm, use exact cascading"
+       OFF)
 if(PAC_NOAPPROX)
-    add_definitions(-DPAC_EXACTSUM)
-    message(STATUS "privacy: Using exact cascading sum algorithm (PAC_NOAPPROX)")
+  add_definitions(-DPAC_EXACTSUM)
+  message(STATUS "privacy: Using exact cascading sum algorithm (PAC_NOAPPROX)")
 else()
-    message(STATUS "privacy: Using approximate sum algorithm (default)")
+  message(STATUS "privacy: Using approximate sum algorithm (default)")
 endif()
 
 # Extension sources
@@ -37,6 +39,7 @@ set(EXTENSION_SOURCES
     src/aggregates/as_count.cpp
     src/aggregates/as_min_max.cpp
     src/aggregates/as_sum.cpp
+    src/aggregates/as_sum_state_export.cpp
     src/aggregates/as_finalize.cpp
     src/aggregates/as_clip_sum.cpp
     src/aggregates/as_clip_min_max.cpp
@@ -48,6 +51,7 @@ set(EXTENSION_SOURCES
     src/compiler/privacy_mechanisms.cpp
     src/compiler/privacy_compiler_helpers.cpp
     src/query_processing/pac_avg_rewriter.cpp
+    src/query_processing/as_correlation_optimizer.cpp
     src/query_processing/privacy_expression_builder.cpp
     src/query_processing/privacy_plan_traversal.cpp
     src/query_processing/pac_projection_propagation.cpp
@@ -69,8 +73,8 @@ build_static_extension(${TARGET_NAME} ${EXTENSION_SOURCES})
 build_loadable_extension(${TARGET_NAME} " " ${EXTENSION_SOURCES})
 
 if(WIN32)
-    target_link_libraries(${EXTENSION_NAME} bcrypt)
-    target_link_libraries(${LOADABLE_EXTENSION_NAME} bcrypt)
+  target_link_libraries(${EXTENSION_NAME} bcrypt)
+  target_link_libraries(${LOADABLE_EXTENSION_NAME} bcrypt)
 endif()
 
 install(
@@ -79,76 +83,73 @@ install(
   LIBRARY DESTINATION "${INSTALL_LIB_DIR}"
   ARCHIVE DESTINATION "${INSTALL_LIB_DIR}")
 
-# Test runners and benchmarks - only build for native platforms, not WASM or Windows
-# (Windows lacks unistd.h and other POSIX features used by the benchmark code)
-set(PRIVACY_LINK_LIBS
-    duckdb_static
-    duckdb_generated_extension_loader
-    ${EXTENSION_NAME}
-    core_functions_extension
-    parquet_extension
-    duckdb_static)
-
-# jemalloc is only built on Linux 64-bit; link it when the extension was loaded
-# (can't use if(TARGET ...) here because the target isn't defined yet at this point)
-if(DUCKDB_EXTENSION_JEMALLOC_SHOULD_LINK)
-    list(APPEND PRIVACY_LINK_LIBS jemalloc_extension)
-endif()
+# Test runners and benchmarks - only build for native platforms, not WASM or
+# Windows (Windows lacks unistd.h and other POSIX features used by the benchmark
+# code)
 
 if(NOT EMSCRIPTEN AND NOT WIN32)
-    # Build a single unified test executable that invokes all tests via src/test/test_runner.cpp
-    add_executable(pac_test_runner
-        src/test/test_runner.cpp
-        src/test/test_plan_traversal.cpp
-        src/test/test_schema_metadata.cpp
-        src/test/test_compiler_functions.cpp
-        src/test/test_privacy_parser.cpp)
-    target_link_libraries(pac_test_runner ${PRIVACY_LINK_LIBS})
-
-    # TPCH + AS benchmark runner
-    add_executable(as_tpch_benchmark benchmark/tpch/as_tpch_benchmark.cpp)
-    target_link_libraries(as_tpch_benchmark ${PRIVACY_LINK_LIBS})
-
-    # WIP rebuttal experiment: four-way MEPS balanced repeated replication benchmark
-    add_executable(meps_brr_poc benchmark/meps_brr/meps_brr_poc.cpp)
-    target_link_libraries(meps_brr_poc ${PRIVACY_LINK_LIBS})
-
-    # WIP rebuttal experiment: standard joined SSB queries and 64 sampled worlds
-    add_executable(as_ssb_benchmark benchmark/ssb/as_ssb_benchmark.cpp)
-    target_link_libraries(as_ssb_benchmark ${PRIVACY_LINK_LIBS})
-
-    # ClickBench + AS benchmark runner
-    add_executable(as_clickbench_benchmark benchmark/clickbench/as_clickbench_benchmark.cpp)
-    target_link_libraries(as_clickbench_benchmark ${PRIVACY_LINK_LIBS})
-
-    # TPC-DS automatic-rewriter coverage and end-to-end performance
-    add_executable(as_tpcds_benchmark benchmark/tpcds/as_tpcds_benchmark.cpp)
-    target_link_libraries(as_tpcds_benchmark ${PRIVACY_LINK_LIBS})
-
-    # Unified DP utility benchmark runner
-    set(TPCH_STOCK_DP_QUERY_NAMES q01 q05 q06 q14 q19)
-    set(TPCH_STOCK_DP_QUERY_FILES)
-    foreach(QUERY_NAME IN LISTS TPCH_STOCK_DP_QUERY_NAMES)
-        string(TOUPPER ${QUERY_NAME} QUERY_NAME_UPPER)
-        set(QUERY_FILE "${CMAKE_CURRENT_SOURCE_DIR}/benchmark/dp/tpch_queries/${QUERY_NAME}.sql")
-        list(APPEND TPCH_STOCK_DP_QUERY_FILES "${QUERY_FILE}")
-        file(READ
-            "${QUERY_FILE}"
-            "TPCH_${QUERY_NAME_UPPER}_SQL")
-    endforeach()
-    set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS
-        ${TPCH_STOCK_DP_QUERY_FILES})
-    set(TPCH_QUERY_GENERATED_DIR "${CMAKE_CURRENT_BINARY_DIR}/generated")
-    file(MAKE_DIRECTORY "${TPCH_QUERY_GENERATED_DIR}")
-    configure_file(
-        benchmark/include/tpch_stock_queries.hpp.in
-        "${TPCH_QUERY_GENERATED_DIR}/tpch_stock_queries.hpp"
-        @ONLY)
-    add_executable(dp_benchmark_runner benchmark/dp/dp_benchmark_runner.cpp)
-    target_include_directories(dp_benchmark_runner PRIVATE "${TPCH_QUERY_GENERATED_DIR}")
-    target_link_libraries(dp_benchmark_runner ${PRIVACY_LINK_LIBS})
-
-    # IMDb/JOB benchmark runner
-    add_executable(imdb_benchmark benchmark/imdb/imdb_benchmark.cpp)
-    target_link_libraries(imdb_benchmark ${PRIVACY_LINK_LIBS})
+  # Build a single unified test executable that invokes all tests via
+  # src/test/test_runner.cpp
+  add_executable(
+    pac_test_runner
+    src/test/test_runner.cpp src/test/test_plan_traversal.cpp
+    src/test/test_schema_metadata.cpp src/test/test_compiler_functions.cpp
+    src/test/test_privacy_parser.cpp)
+
+  # TPCH + AS benchmark runner
+  add_executable(as_tpch_benchmark benchmark/tpch/as_tpch_benchmark.cpp)
+
+  # WIP rebuttal experiment: four-way MEPS balanced repeated replication
+  # benchmark
+  add_executable(meps_brr_poc benchmark/meps_brr/meps_brr_poc.cpp)
+
+  # WIP rebuttal experiment: standard joined SSB queries and 64 sampled worlds
+  add_executable(as_ssb_benchmark benchmark/ssb/as_ssb_benchmark.cpp)
+
+  # ClickBench + AS benchmark runner
+  add_executable(as_clickbench_benchmark
+                 benchmark/clickbench/as_clickbench_benchmark.cpp)
+
+  # TPC-DS automatic-rewriter coverage and end-to-end performance
+  add_executable(as_tpcds_benchmark benchmark/tpcds/as_tpcds_benchmark.cpp)
+
+  # Unified DP utility benchmark runner
+  set(TPCH_STOCK_DP_QUERY_NAMES q01 q05 q06 q14 q19)
+  set(TPCH_STOCK_DP_QUERY_FILES)
+  foreach(QUERY_NAME IN LISTS TPCH_STOCK_DP_QUERY_NAMES)
+    string(TOUPPER ${QUERY_NAME} QUERY_NAME_UPPER)
+    set(QUERY_FILE
+        "${CMAKE_CURRENT_SOURCE_DIR}/benchmark/dp/tpch_queries/${QUERY_NAME}.sql"
+    )
+    list(APPEND TPCH_STOCK_DP_QUERY_FILES "${QUERY_FILE}")
+    file(READ "${QUERY_FILE}" "TPCH_${QUERY_NAME_UPPER}_SQL")
+  endforeach()
+  set_property(
+    DIRECTORY
+    APPEND
+    PROPERTY CMAKE_CONFIGURE_DEPENDS ${TPCH_STOCK_DP_QUERY_FILES})
+  set(TPCH_QUERY_GENERATED_DIR "${CMAKE_CURRENT_BINARY_DIR}/generated")
+  file(MAKE_DIRECTORY "${TPCH_QUERY_GENERATED_DIR}")
+  configure_file(benchmark/include/tpch_stock_queries.hpp.in
+                 "${TPCH_QUERY_GENERATED_DIR}/tpch_stock_queries.hpp" @ONLY)
+  add_executable(dp_benchmark_runner benchmark/dp/dp_benchmark_runner.cpp)
+  target_include_directories(dp_benchmark_runner
+                             PRIVATE "${TPCH_QUERY_GENERATED_DIR}")
+
+  # IMDb/JOB benchmark runner
+  add_executable(imdb_benchmark benchmark/imdb/imdb_benchmark.cpp)
+
+  foreach(
+    PRIVACY_RUNNER IN
+    ITEMS pac_test_runner
+          as_tpch_benchmark
+          meps_brr_poc
+          as_ssb_benchmark
+          as_clickbench_benchmark
+          as_tpcds_benchmark
+          dp_benchmark_runner
+          imdb_benchmark)
+    target_link_libraries(${PRIVACY_RUNNER} PRIVATE duckdb_static)
+    link_extension_libraries(${PRIVACY_RUNNER} PRIVATE)
+  endforeach()
 endif()
diff --git a/Makefile b/Makefile
@@ -3,6 +3,7 @@ PROJ_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 # Configuration of extension
 EXT_NAME=privacy
 EXT_CONFIG=${PROJ_DIR}extension_config.cmake
+OVERRIDE_GIT_DESCRIBE ?= v2.0.0
 
 # Include the Makefile from extension-ci-tools
-include extension-ci-tools/makefiles/duckdb_extension.Makefile
+include extension-ci-tools/makefiles/duckdb_extension.Makefile
diff --git a/README.md b/README.md
@@ -7,7 +7,7 @@
 
 Both modes share the same DDL and require no changes to query syntax — the extension rewrites aggregate query plans transparently.
 
-This works on DuckDB v1.5 and beyond. See https://duckdb.org/install to install. Or if you do not want to install anything: this extension is also distributed in WASM, so you can run the examples also in https://shell.duckdb.org in a browser.
+This branch targets DuckDB v2.0.0. See https://duckdb.org/install to install. Or if you do not want to install anything: this extension is also distributed in WASM, so you can run the examples also in https://shell.duckdb.org in a browser.
 
 ## Install
 
@@ -218,6 +218,19 @@ This gives a formal (ε,δ)-DP guarantee at the **user level** — neighboring d
 | `dp_delta` | `1e-6` | dp_elastic | Failure probability δ (0 = pure DP) |
 | `dp_sum_bound` | required | dp_elastic | Clipping bound for `SUM`/`AVG` inputs: values are clipped to `[-bound, bound]` |
 
+For PAC, a fixed `privacy_seed` repeats the random draws when the query, data,
+settings, execution order, and initial p-tracking state are unchanged on the same
+build. Stream allocation resets for every query, including prepared executions,
+so unrelated queries do not shift later seeded draws. Streams advance across cells:
+identical aggregate counters in different groups still receive independent draws.
+
+For reproducible tests, use `SET threads = 1` and a stable input order. Parallel
+hash aggregation can emit groups in a different order even at a fixed thread count,
+so bit-identical per-group results are not guaranteed across parallel executions,
+row/batch reordering, or builds. P-tracking updates its calibration state after
+releases; comparisons must start from the same state (the RNG regression tests use
+`SET pac_ptracking = false`).
+
 ## SQL Reference
 
 ### Supported Aggregates and Operators

diff --git a/benchmark/meps_brr/ahrq2016/as_queries/q01_office.sql b/benchmark/meps_brr/ahrq2016/as_queries/q01_office.sql
@@ -2,24 +2,24 @@
 SELECT as_sum(p.mask_lo, p.weight) AS event_count_lo,
        as_sum(p.mask_hi, p.weight) AS event_count_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.self), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_self_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_self_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.self), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_self_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_self_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.private), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_private_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_private_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.private), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_private_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_private_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.medicare), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicare_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicare_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.medicare), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicare_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicare_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.medicaid), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicaid_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicaid_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.medicaid), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicaid_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicaid_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.other), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_other_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_other_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.other), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_other_hi
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_other_hi
 FROM persons_as p JOIN events e USING (person_id)
 WHERE e.kind = 'OB';
diff --git a/benchmark/meps_brr/ahrq2016/as_queries/q02_office_physician.sql b/benchmark/meps_brr/ahrq2016/as_queries/q02_office_physician.sql
@@ -2,24 +2,24 @@
 SELECT as_sum(p.mask_lo, p.weight) AS event_count_lo,
        as_sum(p.mask_hi, p.weight) AS event_count_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.self), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_self_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_self_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.self), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_self_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_self_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.private), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_private_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_private_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.private), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_private_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_private_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.medicare), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicare_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicare_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.medicare), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicare_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicare_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.medicaid), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicaid_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicaid_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.medicaid), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_medicaid_hi,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_medicaid_hi,
        list_transform(list_zip(as_sum(p.mask_lo, p.weight * e.other), as_sum(p.mask_lo, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_other_lo,
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_other_lo,
        list_transform(list_zip(as_sum(p.mask_hi, p.weight * e.other), as_sum(p.mask_hi, p.weight)),
-    x -> x[1] / nullif(x[2], 0)) AS mean_other_hi
+    lambda x: x[1] / nullif(x[2], 0)) AS mean_other_hi
 FROM persons_as p JOIN events e USING (person_id)
 WHERE e.kind = 'OB' AND e.physician = 1;