Do not use partial indexes for compression

They refer only to a subset of the table.
timescale · Nov 6, 2023 · c34fd0b · c34fd0b
1 parent 2422f51
commit c34fd0b
Show file tree

Hide file tree

Showing 4 changed files with 58 additions and 1 deletion.
diff --git a/.unreleased/fix_partial_index b/.unreleased/fix_partial_index
@@ -0,0 +1 @@
+Fixes: #6280 Potential data loss when compressing a table with a partial index that matches compression order.
diff --git a/tsl/src/compression/compression.c b/tsl/src/compression/compression.c
@@ -291,6 +291,17 @@ compress_chunk(Oid in_table, Oid out_table, const ColumnCompressionInfo **column
 			Oid index_oid = lfirst_oid(lc);
 			Relation index_rel = index_open(index_oid, AccessShareLock);
 			IndexInfo *index_info = BuildIndexInfo(index_rel);
+
+			if (index_info->ii_Predicate != 0)
+			{
+				/*
+				 * Can't use partial indexes for compression because they refer
+				 * only to a subset of all rows.
+				 */
+				index_close(index_rel, AccessShareLock);
+				continue;
+			}
+
 			int previous_direction = NoMovementScanDirection;
 			int current_direction = NoMovementScanDirection;
 

diff --git a/tsl/test/expected/compression_indexscan.out b/tsl/test/expected/compression_indexscan.out
@@ -912,6 +912,43 @@ SELECT decompress_chunk(show_chunks('tab1'));
 (4 rows)
 
 DROP INDEX idx_asc_null_first;
+-- Can't use partial indexes for compression because they refer only to a subset of the table.
+create index predicate on tab1(id, c1, time nulls first) where c2 = 0;
+select count(*) from tab1;
+ count 
+-------
+ 62400
+(1 row)
+
+select compress_chunk(show_chunks('tab1'));
+INFO:  compress_chunk_tuplesort_start
+INFO:  compress_chunk_tuplesort_start
+INFO:  compress_chunk_tuplesort_start
+INFO:  compress_chunk_tuplesort_start
+             compress_chunk             
+----------------------------------------
+ _timescaledb_internal._hyper_1_1_chunk
+ _timescaledb_internal._hyper_1_2_chunk
+ _timescaledb_internal._hyper_1_3_chunk
+ _timescaledb_internal._hyper_1_4_chunk
+(4 rows)
+
+select decompress_chunk(show_chunks('tab1'));
+            decompress_chunk            
+----------------------------------------
+ _timescaledb_internal._hyper_1_1_chunk
+ _timescaledb_internal._hyper_1_2_chunk
+ _timescaledb_internal._hyper_1_3_chunk
+ _timescaledb_internal._hyper_1_4_chunk
+(4 rows)
+
+select count(*) from tab1;
+ count 
+-------
+ 62400
+(1 row)
+
+drop index predicate;
 --Tear down
 DROP TABLE tab1;
 DROP TABLE tab2;

diff --git a/tsl/test/sql/compression_indexscan.sql b/tsl/test/sql/compression_indexscan.sql
@@ -251,8 +251,16 @@ SELECT compress_chunk(show_chunks('tab1'));
 SELECT decompress_chunk(show_chunks('tab1'));
 DROP INDEX idx_asc_null_first;
 
+-- Can't use partial indexes for compression because they refer only to a subset of the table.
+create index predicate on tab1(id, c1, time nulls first) where c2 = 0;
+select count(*) from tab1;
+select compress_chunk(show_chunks('tab1'));
+select decompress_chunk(show_chunks('tab1'));
+select count(*) from tab1;
+drop index predicate;
+
 --Tear down
 DROP TABLE tab1;
 DROP TABLE tab2;
 DROP TABLE tab3;
-SET timescaledb.show_compression_path_info = 'off';
+SET timescaledb.show_compression_path_info = 'off';
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		Fixes: #6280 Potential data loss when compressing a table with a partial index that matches compression order.