Pax: fix delete bloom crash

contrib/pax_storage/expected/delete_bloom_stats.out

@@ -0,0 +1,60 @@
+-- Regression test for issue #1749:
+-- PAX DELETE crashes with SIGSEGV when bloomfilter_columns are not a
+-- subset of minmax_columns. The stats refresh inside
+-- DeleteWithVisibilityMap must project every column it reads.
+-- Case 1: bloomfilter column (payload) is NOT in minmax_columns.
+-- Pre-fix: segment crashed on DELETE.
+drop table if exists pax_delete_bloom_crash;
+NOTICE:  table "pax_delete_bloom_crash" does not exist, skipping
+create table pax_delete_bloom_crash (id int, k int, payload text)
+using pax
+with (minmax_columns = 'id', bloomfilter_columns = 'payload');
+NOTICE:  Table doesn't have 'DISTRIBUTED BY' clause -- Using column named 'id' as the Apache Cloudberry data distribution key for this table.
+HINT:  The 'DISTRIBUTED BY' clause determines the distribution of data. Make sure column(s) chosen are the optimal data distribution key to minimize skew.
+insert into pax_delete_bloom_crash
+select i, i % 10, 'payload-' || i::text
+from generate_series(1, 10000) as i;
+delete from pax_delete_bloom_crash where id between 1 and 100;
+select count(*) from pax_delete_bloom_crash;
+ count 
+-------
+  9900
+(1 row)
+
+drop table pax_delete_bloom_crash;
+-- Case 2: bloomfilter only, no minmax columns.
+drop table if exists pax_delete_bf_only;
+NOTICE:  table "pax_delete_bf_only" does not exist, skipping
+create table pax_delete_bf_only (id int, payload text)
+using pax
+with (bloomfilter_columns = 'payload');
+NOTICE:  Table doesn't have 'DISTRIBUTED BY' clause -- Using column named 'id' as the Apache Cloudberry data distribution key for this table.
+HINT:  The 'DISTRIBUTED BY' clause determines the distribution of data. Make sure column(s) chosen are the optimal data distribution key to minimize skew.
+insert into pax_delete_bf_only
+select i, 'payload-' || i::text from generate_series(1, 5000) as i;
+delete from pax_delete_bf_only where id between 1 and 50;
+select count(*) from pax_delete_bf_only;
+ count 
+-------
+  4950
+(1 row)
+
+drop table pax_delete_bf_only;
+-- Case 3: minmax and bloomfilter columns overlap but neither is a subset.
+drop table if exists pax_delete_mm_bf_mixed;
+NOTICE:  table "pax_delete_mm_bf_mixed" does not exist, skipping
+create table pax_delete_mm_bf_mixed (id int, k int, payload text)
+using pax
+with (minmax_columns = 'id,payload', bloomfilter_columns = 'k,payload');
+NOTICE:  Table doesn't have 'DISTRIBUTED BY' clause -- Using column named 'id' as the Apache Cloudberry data distribution key for this table.
+HINT:  The 'DISTRIBUTED BY' clause determines the distribution of data. Make sure column(s) chosen are the optimal data distribution key to minimize skew.
+insert into pax_delete_mm_bf_mixed
+select i, i % 7, 'p-' || i::text from generate_series(1, 5000) as i;
+delete from pax_delete_mm_bf_mixed where id between 1 and 50;
+select count(*) from pax_delete_mm_bf_mixed;
+ count 
+-------
+  4950
+(1 row)
+
+drop table pax_delete_mm_bf_mixed;

contrib/pax_storage/pax_schedule

@@ -6,6 +6,7 @@ test: alter_distributed
 test: toast toast_failed detoast ddl numeric types filter
 test: update
 test: statistics_bloom_filter
+test: delete_bloom_stats
 
 test: filter_tree filter_tree_arithmetic
 test: filter_tree_root_quals

contrib/pax_storage/sql/delete_bloom_stats.sql

@@ -0,0 +1,48 @@
+-- Regression test for issue #1749:
+-- PAX DELETE crashes with SIGSEGV when bloomfilter_columns are not a
+-- subset of minmax_columns. The stats refresh inside
+-- DeleteWithVisibilityMap must project every column it reads.
+
+-- Case 1: bloomfilter column (payload) is NOT in minmax_columns.
+-- Pre-fix: segment crashed on DELETE.
+drop table if exists pax_delete_bloom_crash;
+create table pax_delete_bloom_crash (id int, k int, payload text)
+using pax
+with (minmax_columns = 'id', bloomfilter_columns = 'payload');
+
+insert into pax_delete_bloom_crash
+select i, i % 10, 'payload-' || i::text
+from generate_series(1, 10000) as i;
+
+delete from pax_delete_bloom_crash where id between 1 and 100;
+select count(*) from pax_delete_bloom_crash;
+
+drop table pax_delete_bloom_crash;
+
+-- Case 2: bloomfilter only, no minmax columns.
+drop table if exists pax_delete_bf_only;
+create table pax_delete_bf_only (id int, payload text)
+using pax
+with (bloomfilter_columns = 'payload');
+
+insert into pax_delete_bf_only
+select i, 'payload-' || i::text from generate_series(1, 5000) as i;
+
+delete from pax_delete_bf_only where id between 1 and 50;
+select count(*) from pax_delete_bf_only;
+
+drop table pax_delete_bf_only;
+
+-- Case 3: minmax and bloomfilter columns overlap but neither is a subset.
+drop table if exists pax_delete_mm_bf_mixed;
+create table pax_delete_mm_bf_mixed (id int, k int, payload text)
+using pax
+with (minmax_columns = 'id,payload', bloomfilter_columns = 'k,payload');
+
+insert into pax_delete_mm_bf_mixed
+select i, i % 7, 'p-' || i::text from generate_series(1, 5000) as i;
+
+delete from pax_delete_mm_bf_mixed where id between 1 and 50;
+select count(*) from pax_delete_mm_bf_mixed;
+
+drop table pax_delete_mm_bf_mixed;

contrib/pax_storage/src/cpp/storage/pax.cc

@@ -27,8 +27,11 @@
 
 #include "storage/pax.h"
 
+#include <algorithm>
+#include <iterator>
 #include <map>
 #include <utility>
+#include <vector>
 
 #include "access/pax_visimap.h"
 #include "access/paxc_rel_options.h"
@@ -595,14 +598,23 @@ void TableDeleter::DeleteWithVisibilityMap(
     return;
   }
   std::vector<int> min_max_col_idxs;
+  std::vector<int> bf_col_idxs;
+  std::vector<int> stats_proj_col_idxs;
   auto stats_updater_projection = std::make_shared<PaxFilter>();
 
   std::unique_ptr<Bitmap8> visi_bitmap;
   auto catalog_update = pax::PaxCatalogUpdater::Begin(rel_);
   auto rel_path = cbdb::BuildPaxDirectoryPath(rel_->rd_node, rel_->rd_backend);
 
   min_max_col_idxs = cbdb::GetMinMaxColumnIndexes(rel_);
-  stats_updater_projection->SetColumnProjection(min_max_col_idxs,
+  bf_col_idxs = cbdb::GetBloomFilterColumnIndexes(rel_);
+
+  // Projection must cover minmax ∪ bloomfilter columns; otherwise
+  // AddRow reads uninitialized slot values for bf columns (issue #1749).
+  std::set_union(min_max_col_idxs.begin(), min_max_col_idxs.end(),
+                 bf_col_idxs.begin(), bf_col_idxs.end(),
+                 std::back_inserter(stats_proj_col_idxs));
+  stats_updater_projection->SetColumnProjection(stats_proj_col_idxs,
                                                 rel_->rd_att->natts);
   do {
     auto it = iterator->Next();
@@ -675,7 +687,7 @@ void TableDeleter::DeleteWithVisibilityMap(
     UpdateStatsInAuxTable(
         catalog_update, micro_partition_metadata,
         std::make_shared<Bitmap8>(visi_bitmap->Raw()), min_max_col_idxs,
-        cbdb::GetBloomFilterColumnIndexes(rel_), stats_updater_projection);
+        bf_col_idxs, stats_updater_projection);
 
     // write pg_pax_blocks_oid
     catalog_update.UpdateVisimap(block_id, visimap_file_name);