I spent about a week optimizing a query in our performance-testing environment, which has hardware similar to
production.
I was able to refactor the query and reduce the runtime from hours to about 40 seconds, through the use of CTEs and a
coupleof new indexes.
The database was rebuilt and refreshed with the very similar data from production, but now the query takes hours again.
In the query plan, it is clear that the row count estimates are WAY too low, even though the statistics are up to date.
Here's a sample query plan:
CTE Scan on stef (cost=164.98..165.00 rows=1 width=38)
CTE terms
-> Nested Loop (cost=0.00..62.40 rows=1 width=12)
-> Index Scan using term_idx1 on term t (cost=0.00..52.35 rows=1 width=12)
Index Cond: (partner_id = 497)
Filter: (recalculate_district_averages_yn AND (NOT is_deleted_yn))
-> Index Scan using growth_measurement_window_fk1 on growth_measurement_window gw (cost=0.00..10.04 rows=1
width=4)
Index Cond: (term_id = t.term_id)
Filter: (test_window_complete_yn AND (NOT is_deleted_yn) AND ((growth_window_type)::text =
'DISTRICT'::text))
CTE stef
-> Nested Loop (cost=0.00..102.58 rows=1 width=29)
Join Filter: ((ssef.student_id = terf.student_id) AND (ssef.grade_id = terf.grade_id))
-> Nested Loop (cost=0.00..18.80 rows=3 width=28)
-> CTE Scan on terms t (cost=0.00..0.02 rows=1 width=8)
-> Index Scan using student_school_enrollment_fact_idx2 on student_school_enrollment_fact ssef
(cost=0.00..18.74rows=3 width=20)
Index Cond: ((partner_id = t.partner_id) AND (term_id = t.term_id))
Filter: primary_yn
-> Index Scan using test_event_result_fact_idx3 on test_event_result_fact terf (cost=0.00..27.85 rows=4
width=25)
Index Cond: ((partner_id = t.partner_id) AND (term_id = t.term_id))
Filter: growth_event_yn
The estimates in the first CTE are correct, but in the second, the scan on student_school_enrollment_fact will return
about1.5 million rows, and the scan on test_event_result_fact actually returns about 1.1 million. The top level join
shouldreturn about 900K rows. I believe the fundamental issue is that the CTE stef outer nested loop should be a merge
joininstead, but I cannot figure out why the optimizer is estimating one row when it has the statistics to correctly
estimatethe count.
What would cause PG to so badly estimate the row counts?
I've already regenerated the indexes and re-analyzed the tables involved.
What else can I do to find out why it's running so slowly?
