Materialized views in PostgreSQL use the rule system like views do, but persist the results in a table-like form. The main differences between:
CREATE MATERIALIZED VIEW mymatview AS SELECT * FROM mytab;
and:
CREATE TABLE mymatview AS SELECT * FROM mytab;
are that the materialized view cannot subsequently be directly updated and that the query used to create the materialized view is stored in exactly the same way that a view's query is stored, so that fresh data can be generated for the materialized view with:
REFRESH MATERIALIZED VIEW mymatview;
The information about a materialized view in the PostgreSQL system catalogs is exactly the same as it is for a table or view. So for the parser, a materialized view is a relation, just like a table or a view. When a materialized view is referenced in a query, the data is returned directly from the materialized view, like from a table; the rule is only used for populating the materialized view.
While access to the data stored in a materialized view is often much faster than accessing the underlying tables directly or through a view, the data is not always current; yet sometimes current data is not needed. Consider a table which records sales:
CREATE TABLE invoice ( invoice_no integer PRIMARY KEY, seller_no integer, -- ID of salesperson invoice_date date, -- date of sale invoice_amt numeric(13,2) -- amount of sale );
If people want to be able to quickly graph historical sales data, they might want to summarize, and they may not care about the incomplete data for the current date:
CREATE MATERIALIZED VIEW sales_summary AS SELECT seller_no, invoice_date, sum(invoice_amt)::numeric(13,2) as sales_amt FROM invoice WHERE invoice_date < CURRENT_DATE GROUP BY seller_no, invoice_date; CREATE UNIQUE INDEX sales_summary_seller ON sales_summary (seller_no, invoice_date);
This materialized view might be useful for displaying a graph in the dashboard created for salespeople. A job could be scheduled to update the statistics each night using this SQL statement:
REFRESH MATERIALIZED VIEW sales_summary;
Another use for a materialized view is to allow faster access to data brought across from a remote system through a foreign data wrapper. A simple example using file_fdw
is below, with timings, but since this is using cache on the local system the performance difference compared to access to a remote system would usually be greater than shown here. Notice we are also exploiting the ability to put an index on the materialized view, whereas file_fdw
does not support indexes; this advantage might not apply for other sorts of foreign data access.
Setup:
CREATE EXTENSION file_fdw; CREATE SERVER local_file FOREIGN DATA WRAPPER file_fdw; CREATE FOREIGN TABLE words (word text NOT NULL) SERVER local_file OPTIONS (filename '/usr/share/dict/words'); CREATE MATERIALIZED VIEW wrd AS SELECT * FROM words; CREATE UNIQUE INDEX wrd_word ON wrd (word); CREATE EXTENSION pg_trgm; CREATE INDEX wrd_trgm ON wrd USING gist (word gist_trgm_ops); VACUUM ANALYZE wrd;
Now let's spell-check a word. Using file_fdw
directly:
SELECT count(*) FROM words WHERE word = 'caterpiler'; count ------- 0 (1 row)
With EXPLAIN ANALYZE
, we see:
Aggregate (cost=21763.99..21764.00 rows=1 width=0) (actual time=188.180..188.181 rows=1 loops=1) -> Foreign Scan on words (cost=0.00..21761.41 rows=1032 width=0) (actual time=188.177..188.177 rows=0 loops=1) Filter: (word = 'caterpiler'::text) Rows Removed by Filter: 479829 Foreign File: /usr/share/dict/words Foreign File Size: 4953699 Planning time: 0.118 ms Execution time: 188.273 ms
If the materialized view is used instead, the query is much faster:
Aggregate (cost=4.44..4.45 rows=1 width=0) (actual time=0.042..0.042 rows=1 loops=1) -> Index Only Scan using wrd_word on wrd (cost=0.42..4.44 rows=1 width=0) (actual time=0.039..0.039 rows=0 loops=1) Index Cond: (word = 'caterpiler'::text) Heap Fetches: 0 Planning time: 0.164 ms Execution time: 0.117 ms
Either way, the word is spelled wrong, so let's look for what we might have wanted. Again using file_fdw
and pg_trgm
:
SELECT word FROM words ORDER BY word <-> 'caterpiler' LIMIT 10; word --------------- cater caterpillar Caterpillar caterpillars caterpillar's Caterpillar's caterer caterer's caters catered (10 rows)
Limit (cost=11583.61..11583.64 rows=10 width=32) (actual time=1431.591..1431.594 rows=10 loops=1) -> Sort (cost=11583.61..11804.76 rows=88459 width=32) (actual time=1431.589..1431.591 rows=10 loops=1) Sort Key: ((word <-> 'caterpiler'::text)) Sort Method: top-N heapsort Memory: 25kB -> Foreign Scan on words (cost=0.00..9672.05 rows=88459 width=32) (actual time=0.057..1286.455 rows=479829 loops=1) Foreign File: /usr/share/dict/words Foreign File Size: 4953699 Planning time: 0.128 ms Execution time: 1431.679 ms
Using the materialized view:
Limit (cost=0.29..1.06 rows=10 width=10) (actual time=187.222..188.257 rows=10 loops=1) -> Index Scan using wrd_trgm on wrd (cost=0.29..37020.87 rows=479829 width=10) (actual time=187.219..188.252 rows=10 loops=1) Order By: (word <-> 'caterpiler'::text) Planning time: 0.196 ms Execution time: 198.640 ms
If you can tolerate periodic update of the remote data to the local database, the performance benefit can be substantial.
If you see anything in the documentation that is not correct, does not match your experience with the particular feature or requires further clarification, please use this form to report a documentation issue.