Steps to reproduce the issue:

1. Create a database, insert a record, update it in a transaction with no auto undo and rollback.

create table t1_blob (id integer, str1 blob);
create index t1_blob_idx1 on t1_blob (id);
insert into t1_blob values (0, 'abc');
commit;

set transaction no auto undo;
update t1_blob set id = 1, str1 = '123' where id = 0;
rollback;

2. Close the connection. Execute gstat -r to see that we have 2 record versions (total records: 1, total versions: 1), 2 blobs (Blobs: 2) and 2 index nodes (nodes: 2).
3. Connect again and execute select to trigger the garbage collection:

select * from t1_blob;

4. Close the connection. Execute gstat -r again and see that the uncommitted record version is deleted (total versions: 0), but the amount of blobs and index nodes remained unchanged (Blobs: 2, nodes: 2).

The issue starts in VIO_record. There is a call record->reset(format) which clears the REC_gc_active flag by mistake. After that VIO_gc_record reuses the record which is already active.