Problem with grouping, uses Sort and GroupAggregate, HashAggregate is better(?)

(now with a more sensible subject)

I have been trying to change a many parameters on server versions
7.4.15, 8.1.4, 8.2.0 and 8.2.1. I still hope a have managed to keep
my head straigth and that i do not present to much faulty information.

The cost estimates generated by the different server versions differ.
I have a query which (as far as i can tell) have some strange differences
between 8.2.0 8.2.1. I can provide information about that if anyone want
it.

Generally these parameters are used.
default_statistics_target = 10
(4 selected columns is set to 1000)
(I have tested with 1000 as default value
but that did not have an impact)
(analyzed whenever value was changed)
shared_buffers = 64000 (512MB)
work_mem/sort_mem = variable, see different run's
effective_cache_size = 128000 (1G)
random_page_cost = 2
cpu_index_tuple_cost = 0.001
cpu_operator_cost = 0.025
cpu_tuple_cost = 0.01

I have tested with different values for random_page_cost and
cpu_*_cost but it have not made a difference.
I have tried with random_page cost between 1 and 8,
and cpu_*_cost with standard value and 50x bigger)

Query is:
explain
analyze
select
ur.id as ur_id,
ur.unit_ref,
ur.execution_time,
u.serial_number,
to_char(ur.start_date_time, 'YYYY-MM-DD'),
count(*) as num_test
from
uut_result as ur
inner join units as u
on ur.unit_ref=u.ref
inner join step_result as sr
on ur.id=sr.uut_result
where
ur.id between 174000 and 174000+999
group by
ur.id,
ur.unit_ref,
ur.execution_time,
u.serial_number,
ur.start_date_time
-- order by
-- ur.start_date_time
;
NB: order by clause is used in some results below.

=== Run 1:
Detect work_mem setting influence (See also Run 2)
- server version 8.2.1
- Query executed without "order by" clause
- work_mem = 8600;
QUERY PLAN
---------------------------------------------
GroupAggregate (cost=44857.70..47976.79 rows=95972 width=37) (actual time=1802.716..2017.337 rows=1000 loops=1)
-> Sort (cost=44857.70..45097.63 rows=95972 width=37) (actual time=1802.461..1892.743 rows=138810 loops=1)
Sort Key: ur.id, ur.unit_ref, ur.execution_time, u.serial_number, ur.start_date_time
-> Nested Loop (cost=0.00..36915.87 rows=95972 width=37) (actual time=0.063..268.186 rows=138810 loops=1)
-> Nested Loop (cost=0.00..5017.65 rows=981 width=37) (actual time=0.047..11.919 rows=1000 loops=1)
-> Index Scan using uut_result_pkey on uut_result ur (cost=0.00..1538.77 rows=1000 width=24) (actual time=0.029..1.727 rows=1000 loops=1)
Index Cond: ((id >= 174000) AND (id <= 174999))
-> Index Scan using units_pkey on units u (cost=0.00..3.47 rows=1 width=17) (actual time=0.006..0.007 rows=1 loops=1000)
Index Cond: (ur.unit_ref = u.ref)
-> Index Scan using uut_result_key on step_result sr (cost=0.00..30.82 rows=136 width=4) (actual time=0.011..0.125 rows=139 loops=1000)
Index Cond: (ur.id = sr.uut_result) Total runtime: 2021.833 ms
(12 rows)

=== Run 2:
Detect work_mem setting influence (See also Run 1)
- server version 8.2.1
- Query executed without "order by" clause
- work_mem = 8700;
QUERY PLAN
---------------------------------------------
HashAggregate (cost=38355.45..39795.03 rows=95972 width=37) (actual time=436.406..439.867 rows=1000 loops=1)
-> Nested Loop (cost=0.00..36915.87 rows=95972 width=37) (actual time=0.066..256.235 rows=138810 loops=1)
-> Nested Loop (cost=0.00..5017.65 rows=981 width=37) (actual time=0.049..10.858 rows=1000 loops=1)
-> Index Scan using uut_result_pkey on uut_result ur (cost=0.00..1538.77 rows=1000 width=24) (actual time=0.031..1.546 rows=1000 loops=1)
Index Cond: ((id >= 174000) AND (id <= 174999))
-> Index Scan using units_pkey on units u (cost=0.00..3.47 rows=1 width=17) (actual time=0.005..0.006 rows=1 loops=1000)
Index Cond: (ur.unit_ref = u.ref)
-> Index Scan using uut_result_key on step_result sr (cost=0.00..30.82 rows=136 width=4) (actual time=0.011..0.123 rows=139 loops=1000)
Index Cond: (ur.id = sr.uut_result) Total runtime: 441.193 ms (10 rows)

=== Comment on Run 1 versus Run 2 (adjusted work_mem) ===
The difference in setup is value of work_mem. Bigger work_mem gave different
cost estimates and selected HashAggregate instead of GroupAggregate.
Result was a reduced runtime. I guess that is as expected.

(One remark, the switchover between different plans on version 8.1.5 was for
work_mem values of 6800 and 6900)

=== Run 3 (with order by clause):
Test "group by" and "order by" (See also Run 1 and Run 4)
- server version 8.2.1
- Query executed with "order by" clause
- work_mem = 8700
(tried values from 2000 to 128000 with same cost and plan as result)
QUERY PLAN
---------------------------------------------
GroupAggregate (cost=44857.70..47976.79 rows=95972 width=37) (actual time=1891.464..2114.462 rows=1000 loops=1)
-> Sort (cost=44857.70..45097.63 rows=95972 width=37) (actual time=1891.263..1982.137 rows=138810 loops=1)
Sort Key: ur.start_date_time, ur.id, ur.unit_ref, ur.execution_time, u.serial_number
-> Nested Loop (cost=0.00..36915.87 rows=95972 width=37) (actual time=0.064..264.358 rows=138810 loops=1)
-> Nested Loop (cost=0.00..5017.65 rows=981 width=37) (actual time=0.047..12.253 rows=1000 loops=1)
-> Index Scan using uut_result_pkey on uut_result ur (cost=0.00..1538.77 rows=1000 width=24) (actual time=0.029..1.743 rows=1000 loops=1)
Index Cond: ((id >= 174000) AND (id <= 174999))
-> Index Scan using units_pkey on units u (cost=0.00..3.47 rows=1 width=17) (actual time=0.006..0.007 rows=1 loops=1000)
Index Cond: (ur.unit_ref = u.ref)
-> Index Scan using uut_result_key on step_result sr (cost=0.00..30.82 rows=136 width=4) (actual time=0.011..0.124 rows=139 loops=1000)
Index Cond: (ur.id = sr.uut_result) Total runtime: 2118.986 ms
(12 rows)

=== Run 4 (with order by clause, on server 8.1.4):
Test "group by" and "order by" (See also Run 1 and Run 3)
- server version 8.1.4
- Query executed with "order by" clause
- work_mem = 6900
(same plan select for all work_mem values above 6900)
QUERY PLAN
------------------------------------------------------------
Sort (cost=46578.83..46820.66 rows=96734 width=37) (actual time=505.562..505.988 rows=1000 loops=1)
Sort Key: ur.start_date_time
-> HashAggregate (cost=37117.40..38568.41 rows=96734 width=37) (actual time=498.697..502.374 rows=1000 loops=1)
-> Nested Loop (cost=0.00..35666.39 rows=96734 width=37) (actual time=0.058..288.270 rows=138810 loops=1)
-> Nested Loop (cost=0.00..5342.20 rows=984 width=37) (actual time=0.042..11.773 rows=1000 loops=1)
-> Index Scan using uut_result_pkey on uut_result ur (cost=0.00..1626.46 rows=1003 width=24) (actual time=0.020..1.868 rows=1000 loops=1)
Index Cond: ((id >= 174000) AND (id <= 174999))
-> Index Scan using units_pkey on units u (cost=0.00..3.69 rows=1 width=17) (actual time=0.006..0.007 rows=1 loops=1000)
Index Cond: ("outer".unit_ref = u.ref)
-> Index Scan using uut_result_key on step_result sr (cost=0.00..29.09 rows=138 width=4) (actual time=0.006..0.146 rows=139 loops=1000)
Index Cond: ("outer".id = sr.uut_result) Total runtime: 507.452 ms
(12 rows)

=== Coemment on selected plan for 8.2.1 when using "order by" ===
Run 3 (8.2.1 with order by) selects same plan as Run1 (without order by).
It does hovever exist a better plan for Run3, and 8.1.5 manages to select
that plan (shown in Run 4).
Both versions (8.1.5 and 8.2.1) uses same plan until the uppermost Nested Loop.
The big difference is that 8.1.5 then will do HashAggregate, and then sort,
while 8.2.1 will does a sort and then a GroupAggregate.

I have tried different combinations for statistics_target, cpu_*_cost,
work_mem and random page cost without finding a solution.

Anyone with an idea on what to do? Feel free to suggest one of the above
parameters, i might have overlooked some combination.

I am a little unsure on how much extra information is necessery, but i
will provide some:

The three tables are
units List of produced items
uut_Result Summary of test result
step_result Individuel tests results
The system is a production test log. (there are a lot of units which
does not have an entry in uut_result).

Table "public.units"
Column | Type | Modifiers
------------------+-----------------------+-----------------------------
------------------+-----------------------+-----------------------------
------------------+-----------------------+----------
ref | integer | not null default nextval(('public.units_ref_seq'::text)::regclass)
serial_number | character varying(30) | not null
product_ref | integer | not null
week | integer | not null
status | integer | not null
comment | text |
last_user | text | default "current_user"()
last_date | date | default ('now'::text)::date
product_info_ref | integer | not null
Indexes:
"units_pkey" PRIMARY KEY, btree (ref)
"units_no_sno_idx" UNIQUE, btree (product_ref, week) WHERE serial_number::text = ''::text
"units_serial_number_idx" UNIQUE, btree (serial_number, product_info_ref) WHERE serial_number::text <> ''::text
"units_product_ref_key" btree (product_ref)
Triggers:
ct_unit_update_log AFTER UPDATE ON units FOR EACH ROW EXECUTE PROCEDURE cf_unit_update_log()
ct_units_update_product_info_ref BEFORE INSERT OR UPDATE ON units FOR EACH ROW EXECUTE PROCEDURE cf_units_update_product_info_ref()
select count(*) from units => 292 676 rows

Table "public.uut_result"
Column | Type | Modifiers
-------------------+-----------------------------+----------------------
-------------------+-----------------------------+----------------------
-------------------+-----------------------------+--------
id | integer | not null
uut_serial_number | text |
unit_ref | integer |
order_unit_ref | integer |
user_login_name | text |
start_date_time | timestamp without time zone |
execution_time | double precision |
uut_status | text |
uut_error_code | integer |
uut_error_message | text |
last_user | text | default "current_user"()
last_timestamp | timestamp with time zone | default ('now'::text)::timestamp(6) with time zone
test_name | text |
teststation_name | text |
teststation_ref | integer |
process_step_ref | integer |
Indexes:
"uut_result_pkey" PRIMARY KEY, btree (id)
"uut_result_start_date_time_idx" btree (start_date_time)
"uut_result_test_name" btree (test_name)
Triggers:
ct_set_process_step_ref BEFORE INSERT OR UPDATE ON uut_result FOR EACH ROW EXECUTE PROCEDURE cf_set_process_step_ref() select count(*) from uut_result => 180 111 rows

Table "public.step_result"
Column | Type | Modifiers
--------------------+------------------+-----------
id | integer | not null
uut_result | integer |
step_parent | integer |
step_name | text |
step_extra_info | text |
step_type | text |
status | text |
report_text | text |
error_code | integer |
error_message | text |
module_time | double precision |
total_time | double precision |
num_loops | integer |
num_passed | integer |
num_failed | integer |
ending_loop_index | integer |
loop_index | integer |
interactive_exenum | integer |
step_group | text |
step_index | integer |
order_number | integer |
pass_fail | integer |
numeric_value | double precision |
high_limit | double precision |
low_limit | double precision |
comp_operator | text |
string_value | text |
string_limit | text |
button_pressed | integer |
response | text |
exit_code | integer |
num_limits_in_file | integer |
num_rows_in_file | integer |
num_limits_applied | integer |
sequence_name | text |
sequence_file_path | text |
Indexes:
"step_result_pkey" PRIMARY KEY, btree (id)
"step_parent_key" btree (step_parent)
"temp_index_idx" btree (sequence_file_path)
"uut_result_key" btree (uut_result)
select count(*) from step_result => 17 624 657 rows

Best regards
Rolf Østvik