VACUUM FULL vs dropped columns

While doing some testing I noticed this, which seems somewhat perverse:

create table t();
insert into t select from generate_series(1,10000);
select 'alter table t ' || string_agg(' add column c'||x::text||' int
default ' ||x::text,',')
from generate_series(1,1000) x \gexec

create table t_dropped();
insert into t_dropped select from generate_series(1,10000);
select 'alter table t_dropped ' || string_agg(' add column
c'||x::text||' int default ' ||x::text,',')
from generate_series(1,1000)  x \gexec

alter table t_dropped drop column c900;

select pg_total_relation_size('t') as size_t,
pg_total_relation_size('t_dropped') as size_t_dropped;

  size_t  | size_t_dropped
----------+----------------
 40960000 |       40960000
(1 row)


vacuum full t;
vacuum full t_dropped;

select pg_total_relation_size('t') as size_t,
pg_total_relation_size('t_dropped') as size_t_dropped;

  size_t  | size_t_dropped
----------+----------------
 40960000 |       81920000
(1 row)


Why does VACUUM FULL cause the size of this table with a single
dropped column (1 out of 1000) cause the table size to double?

cheers

andrew


-- 
Andrew Dunstan                https://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services