We are trying to create OEBS-analog functionality in Postgresql. Let's say we have a form constructor and need to store form results in database (e.g. email bodies). In Oracle you could use a table with 150~ columns (and some mapping stored elsewhere) to store each field in separate column. But in contrast to Oracle we would like to store all the form in postgresql xml field.
The example of the tree is
<row xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
<object_id>2</object_id>
<pack_form_id>23</pack_form_id>
<prod_form_id>34</prod_form_id>
</row>
We would like to search through this field.
Test table contains 400k rows and the following select executes in 90 seconds:
select *
from params
where (xpath('//prod_form_id/text()'::text, xmlvalue))[1]::text::int=34
So I created this index:
create index prod_form_idx
ON params using btree(
((xpath('//prod_form_id/text()'::text, xmlvalue))[1]::text::int)
);
And it made no difference. Still 90 seconds execution. EXPLAIN plan show this:
Bitmap Heap Scan on params (cost=40.29..6366.44 rows=2063 width=292)
Recheck Cond: ((((xpath('//prod_form_id/text()'::text, xmlvalue, '{}'::text[]))[1])::text)::integer = 34)
-> Bitmap Index Scan on prod_form_idx (cost=0.00..39.78 rows=2063 width=0)
Index Cond: ((((xpath('//prod_form_id/text()'::text, xmlvalue, '{}'::text[]))[1])::text)::integer = 34)
I am not the great plan interpreter so I suppose this means that index is being used. The question is: where's all the speed? And what can i do in order to optimize this kind of queries?
Well, at least the index is used. You get a bitmap index scan instead of a normal index scan though, which means the xpath() function will be called lots of times.
Let's do a little check :
CREATE TABLE foo ( id serial primary key, x xml, h hstore );
insert into foo (x,h) select XMLPARSE( CONTENT '<row xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance">
<object_id>2</object_id>
<pack_form_id>' || n || '</pack_form_id>
<prod_form_id>34</prod_form_id>
</row>' ),
('object_id=>2,prod_form_id=>34,pack_form_id=>'||n)::hstore
FROM generate_series( 1,100000 ) n;
test=> EXPLAIN ANALYZE SELECT count(*) FROM foo;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------
Aggregate (cost=4821.00..4821.01 rows=1 width=0) (actual time=24.694..24.694 rows=1 loops=1)
-> Seq Scan on foo (cost=0.00..4571.00 rows=100000 width=0) (actual time=0.006..13.996 rows=100000 loops=1)
Total runtime: 24.730 ms
test=> explain analyze select * from foo where (h->'pack_form_id')='123';
QUERY PLAN
----------------------------------------------------------------------------------------------------
Seq Scan on foo (cost=0.00..5571.00 rows=500 width=68) (actual time=0.075..48.763 rows=1 loops=1)
Filter: ((h -> 'pack_form_id'::text) = '123'::text)
Total runtime: 36.808 ms
test=> explain analyze select * from foo where ((xpath('//pack_form_id/text()'::text, x))[1]::text) = '123';
QUERY PLAN
------------------------------------------------------------------------------------------------------
Seq Scan on foo (cost=0.00..5071.00 rows=500 width=68) (actual time=4.271..3368.838 rows=1 loops=1)
Filter: (((xpath('//pack_form_id/text()'::text, x, '{}'::text[]))[1])::text = '123'::text)
Total runtime: 3368.865 ms
As we can see,
scanning the whole table with count(*) takes 25 ms
extracting one key/value from a hstore adds a small extra cost, about 0.12 µs/row
extracting one key/value from a xml using xpath adds a huge cost, about 33 µs/row
Conclusions :
xml is slow (but everyone knows that)
if you want to put a flexible key/value store in a column, use hstore
Also since your xml data is pretty big it will be toasted (compressed and stored out of the main table). This makes the rows in the main table much smaller, hence more rows per page, which reduces the efficiency of bitmap scans since all rows on a page have to be rechecked.
You can fix this though. For some reason the xpath() function (which is very slow, since it handles xml) has the same cost (1 unit) as say, the integer operator "+"...
update pg_proc set procost=1000 where proname='xpath';
You may need to tweak the cost value. When given the right info, the planner knows xpath is slow and will avoid a bitmap index scan, using an index scan instead, which doesn't need rechecking the condition for all rows on a page.
Note that this does not solve the row estimates problem. Since you can't ANALYZE the inside of the xml (or hstore) you get default estimates for the number of rows (here, 500). So, the planner may be completely wrong and choose a catastrophic plan if some joins are involved. The only solution to this is to use proper columns.
Related
I am using the extension
CREATE EXTENSION btree_gin;
I have an index that looks like this...
create index boundaries2 on rets USING GIN(source, isonlastsync, status, (geoinfo::jsonb->'boundaries'), ctcvalidto, searchablePrice, ctcSortOrder);
before I started messing with it, the index looked like this, with the same results that I'm about to share, so it seems minor variations in the index definition don't make a difference:
create index boundaries on rets USING GIN((geoinfo::jsonb->'boundaries'), source, status, isonlastsync, ctcvalidto, searchablePrice, ctcSortOrder);
I give pgsql 11 this query:
explain analyze select id from rets where ((geoinfo::jsonb->'boundaries' ?| array['High School: Torrey Pines']) AND source='SDMLS'
AND searchablePrice>=800000 AND searchablePrice<=1200000 AND YrBlt>=2000 AND EstSF>=2300
AND Beds>=3 AND FB>=2 AND ctcSortOrder>'2019-07-05 16:02:54 UTC' AND Status IN ('ACTIVE','BACK ON MARKET')
AND ctcvalidto='9999-12-31 23:59:59 UTC' AND isonlastsync='true') order by LstDate desc, ctcSortOrder desc LIMIT 3000;
with result...
Limit (cost=120.06..120.06 rows=1 width=23) (actual time=472.849..472.850 rows=1 loops=1)
-> Sort (cost=120.06..120.06 rows=1 width=23) (actual time=472.847..472.848 rows=1 loops=1)
Sort Key: lstdate DESC, ctcsortorder DESC
Sort Method: quicksort Memory: 25kB
-> Bitmap Heap Scan on rets (cost=116.00..120.05 rows=1 width=23) (actual time=472.748..472.841 rows=1 loops=1)
Recheck Cond: ((source = 'SDMLS'::text) AND (((geoinfo)::jsonb -> 'boundaries'::text) ?| '{"High School: Torrey Pines"}'::text[]) AND (ctcvalidto = '9999-12-31 23:59:59+00'::timestamp with time zone) AND (searchableprice >= 800000) AND (searchableprice <= 1200000) AND (ctcsortorder > '2019-07-05 16:02:54+00'::timestamp with time zone))
Rows Removed by Index Recheck: 93
Filter: (isonlastsync AND (yrblt >= 2000) AND (estsf >= 2300) AND (beds >= 3) AND (fb >= 2) AND (status = ANY ('{ACTIVE,"BACK ON MARKET"}'::text[])))
Rows Removed by Filter: 10
Heap Blocks: exact=102
-> Bitmap Index Scan on boundaries2 (cost=0.00..116.00 rows=1 width=0) (actual time=471.762..471.762 rows=104 loops=1)
Index Cond: ((source = 'SDMLS'::text) AND (((geoinfo)::jsonb -> 'boundaries'::text) ?| '{"High School: Torrey Pines"}'::text[]) AND (ctcvalidto = '9999-12-31 23:59:59+00'::timestamp with time zone) AND (searchableprice >= 800000) AND (searchableprice <= 1200000) AND (ctcsortorder > '2019-07-05 16:02:54+00'::timestamp with time zone))
Planning Time: 0.333 ms
Execution Time: 474.311 ms
(14 rows)
The Question
Why are the columns status and isonlastsync not used by the Bitmap Index Scan on boundaries2?
It can do so if it predicts that filtering out those columns will be faster. This is usually the case if cardinality on columns is very low and you will fetch large enough portion of all rows; this is true for boolean like isonlastsync and usually true for status columns with just a few distinct values.
Rows Removed by Filter: 10 this is very little to filter out, either because your table does not hold large number of rows or most of them fit into condition you specified for those two columns. You might try generating more data in that table or selecting rows with rare status.
I suggest doing partial indexes (with WHERE condition), at least for boolean value and remove those two columns to make this index a bit more lightweight.
I cannot tell you why, but I can help you optimize the query.
You should not use a multi-column GIN index, but a GIN index on only the jsonb expression and a b-tree index on the other columns.
The order of columns matters: put the oned used in an equality condition first, with the most selective in the beginning. Next put the column with the must selective inequality or IN conditions. For the following columns, the order doesn't matter, as they will only act as filters in the index scan.
Make sure that the indexes are cached in RAM.
I'd expect you to be faster that way.
I think you're asking yourself the wrong question. As Lukasz answered already, PostgreSQL may find inefficient to check all columns in the index. The problem here is that your index is too big on disk.
Probably by trying to make this SQL faster you added as many columns as possible to the index, but it is backfiring.
The trick is to realize how much data PostgreSQL has to read to find your records. If your index contains too much data, it will have to read a lot. Also, be aware that low cardinality columns don't play well with BTree and common index types; generally you want to avoid indexing them.
To have an index as small as possible and it's quick to do lookups you have to find the column with more cardinality, or better, the column that will return less rows for your query. My guess is "ctcSortOrder". This will be the first column of your index.
Now look, after filtering by the 1st column, which column has now the most cardinality or will filter out most rows. I have no idea on your data, but "source" looks like a good candidate.
Try to avoid jsonb searches unless they are the primary source of the cardinality, and keep the index as a Btree. BTree is several times faster.
And like Lukasz suggested, look on partial indexes. For example add "WHERE Status IN ('ACTIVE','BACK ON MARKET') AND isonlastsync='true'" as these two may be common for all your searches.
Bottom line is, having a simpler, smaller index is faster than having all columns indexed. And the order of the columns does matter a lot. Stick with BTree unless there is a good reason (lots of cardinality in non-btree compatible types).
If your table is huge (>10M rows) consider table partitioning, for example by ctcSortOrder. But I don't think this is your case.
I have one de-normalized table with 40+ columns (~ 1.5 million rows, 1 Gb).
CREATE TABLE tbl1 (
...
division_id integer,
division_name varchar(10),
...
);
I need to speed up query
SELECT DISTINCT division_name, division_id
FROM table
ORDER BY division_name;
Query return only ~250 rows, but very slow cause size of table.
I have tried to create index:
create index idx1 on tbl1 (division_name, division_id)
But current execution plan:
explain analyze SELECT Distinct division_name, division_id FROM tbl1 ORDER BY 1;
QUERY PLAN
-----------------------------------------------------------------
Sort (cost=143135.77..143197.64 rows=24748 width=74) (actual time=1925.697..1925.723 rows=294 loops=1)
Sort Key: division_name
Sort Method: quicksort Memory: 74kB
-> HashAggregate (cost=141082.30..141329.78 rows=24748 width=74) (actual time=1923.853..1923.974 rows=294 loops=1)
Group Key: division_name, division_id
-> Seq Scan on tbl1 (cost=0.00..132866.20 rows=1643220 width=74) (actual time=0.069..703.008 rows=1643220 loops=1)
Planning time: 0.311 ms
Execution time: 1925.883 ms
Any suggestion why index does not work or how I can speed up query in other way?
Server Postgresql 9.6.
p.s. Yes, table has 40+ columns and de-normalized, but I know all pros and cons for with decision.
Update1
#a_horse_with_no_name suggest to use vacuum analyze instead of analyze to update table statistic. Now query plain is:
QUERY PLAN
------------------------
Unique (cost=0.55..115753.43 rows=25208 width=74) (actual time=0.165..921.426 rows=294 loops=1)
-> Index Only Scan using idx1 on tbl1 (cost=0.55..107538.21 rows=1643044 width=74) (actual time=0.162..593.322 rows=1643220 loops=1)
Heap Fetches: 0
Much better!
The index will probably only help if PostgreSQL chooses an “index only scan”, that means that it does not have to look at the table data at all.
Normally PostgreSQL has to check the table data (“heap”) to see if a row is visible for the current transaction, because visibility information is not stored in the index.
If, however, the table does not change much and has recently been VACUUMed, PostgreSQL knows that most of the pages consist only of items visible for everyone (there is a “visibility map” to keep track of that information), and then it might be cheaper to scan the index.
Try running VACUUM on the table and see if that causes an index only scan to be used.
Other than that, there is no way to speed up such a query.
When searching on Google for join table indexes, I got this question.
Now, I believe that it is giving some false information in the accepted answer, or I do not understand how everything works.
Given the following tables (running on PostGreSQL 9.4):
CREATE TABLE "albums" ("album_id" serial PRIMARY KEY, "album_name" text)
CREATE TABLE "artists" ("artist_id" serial PRIMARY KEY, "artist_name" text)
CREATE TABLE "albums_artists" ("album_id" integer REFERENCES "albums", "artist_id" integer REFERENCES "artists")
I was trying to replicate the scenario from the question mentioned above, by creating first an index on both of the columns of the albums_artists table and then one index for each column (without keeping the index on both columns).
I would have been expecting very different results when using the EXPLAIN command for a normal, traditional select like the following one:
SELECT "artists".* FROM "test"."artists"
INNER JOIN "test"."albums_artists" ON ("albums_artists"."artist_id" = "artists"."artist_id")
WHERE ("albums_artists"."album_id" = 1)
However, when actually running explain on it, I get exactly the same result for each of the cases (with one index on each column vs. one index on both columns).
I've been reading the documentation on PostGreSQL about indexing and it doesn't make any sense on the results that I am getting:
Hash Join (cost=15.05..42.07 rows=11 width=36) (actual time=0.024..0.025 rows=1 loops=1)
Hash Cond: (artists.artist_id = albums_artists.artist_id)
-> Seq Scan on artists (cost=0.00..22.30 rows=1230 width=36) (actual time=0.006..0.006 rows=1 loops=1)
-> Hash (cost=14.91..14.91 rows=11 width=4) (actual time=0.009..0.009 rows=1 loops=1)
Buckets: 1024 Batches: 1 Memory Usage: 1kB
-> Bitmap Heap Scan on albums_artists (cost=4.24..14.91 rows=11 width=4) (actual time=0.008..0.009 rows=1 loops=1)
Recheck Cond: (album_id = 1)
Heap Blocks: exact=1
-> Bitmap Index Scan on albums_artists_album_id_index (cost=0.00..4.24 rows=11 width=0) (actual time=0.005..0.005 rows=1 loops=1)
Index Cond: (album_id = 1)
I would expect to not get an Index scan at the last step when using an index composed by 2 different columns (since I am only using one of them in the WHERE clause).
I was about to open a bug in an ORM library that adds one index for both columns for join tables, but now I am not so sure. Can anyone help me understand why is the behavior similar in the two cases and what would actually be the difference, if there is any?
add a NOT NULL constraint on the key columns (a tuple with NULLs would make no sense here)
add a PRIMARY KEY (forcing a UNIQUE index on the two keyfields)
As a suport for FK lookups : add a compound index for the PK fields in reversed order
after creating/adding PKs and indexes, you may want to ANALYZE the table (only key columns have statistics)
CREATE TABLE albums_artists
( album_id integer NOT NULL REFERENCES albums (album_id)
, artist_id integer NOT NULL REFERENCES artists (artist_id)
, PRIMARY KEY (album_id, artist_id)
);
CREATE UNIQUE INDEX ON albums_artists (artist_id, album_id);
The reason behind the observed behaviour is the fact that the planner/optimiser is information based, driven by heuristics. Without any information about the fraction of rows that will actually be needed given the conditions, or the fraction of rows that actually maches (in case of a JOIN), the planner makes a guess: (for example: 10% for a range query). For a small query, a hash join will always be a winning scenario, it does imply fetching all tuples from both tables, but the join itself is very efficient.
For columns that are part of a key or index, statistics will be collected, so the planner can make more realistic estimates about the amount of rows involved. Ald that will often result in an indexed plan, since that could need fewer pages to be fetched.
Foreign keys are a very special case; since the planner will know that all the values from the referring table will be present in the referred table. (that is 100%, assuming NOT NULL)
In a Postgres 9.4 database, I created a GIN trigram index on a table called 'persons' that contains 1514 rows like the following:
CREATE INDEX persons_index_name_1 ON persons
USING gin (lower(name) gin_trgm_ops);
and a query that looks for similar names as follows:
select name, last_name from persons where lower(name) % 'thename'
So, I first issued a query with a name I knew beforehand that would have similar matches, so the explain analyze showed that the index I created was used in this query:
select name, last_name from persons where lower(name) % 'george'
And the results were the expected:
-> Bitmap Heap Scan on persons (cost=52.01..58.72 rows=2 width=26) (actual time=0.054..0.065 rows=1 loops=1)
Recheck Cond: (lower((name)::text) % 'george'::text)
Rows Removed by Index Recheck: 2
Heap Blocks: exact=1
-> Bitmap Index Scan on persons_index_name_1 (cost=0.00..52.01 rows=2 width=0) (actual time=0.032..0.032 rows=3 loops=1)
Index Cond: (lower((name)::text) % 'george'::text)
...
Execution time: 1.382 ms"
So, out of curiosity, I wanted to see if the index was used when the thename parameter contained a name that didn't exist at all in the table:
select name, last_name from persons where lower(name) % 'noname'
But I saw that in this case that the index was not used at all and the execution time was way higher:
-> Seq Scan on persons (cost=0.00..63.72 rows=2 width=26) (actual time=6.494..6.494 rows=0 loops=1)
Filter: (lower((name)::text) % 'noname'::text)
Rows Removed by Filter: 1514
...
Execution time: 7.387 ms
As a test, I tried the same with a GIST index and in both cases, the index was used and the execution time was like the first case above.
I went ahead and recreated the table but this time inserting 10014 rows; and I saw that in both cases above, the GIN index was used and the execution time was the best for those cases.
Why is a GIN index is not used when the query above returns no results in a table with not so much rows (1514 in my case)?
Trigram indexes are case insensitive, test with:
select 'case' <-> 'CASE' AS ci1
, 'case' % 'CASE' AS ci2
, 'CASE' <-> 'CASE' AS c1
, 'CASE' % 'CASE' AS c2;
So you might as well just:
CREATE INDEX persons_index_name_1 ON persons USING gin (name gin_trgm_ops);
And:
select name, last_name from persons where name % 'thename';
As to your actual question, for small tables an index look-up might not pay. That's exactly what your added tests demonstrate. And establishing that nothing matches can be more expensive than finding some matches.
Aside from that, your cost setting and / or table statistics may not be at their respective optimum to let Postgres pick the most adequate query plans.
The expected cost numbers translate to much higher actual cost for the sequential scan than for the bitmap index scan. You may be overestimating the cost of index scans as compared to sequential scans. random_page_cost (and cpu_index_tuple_cost) may be set too high and effective_cache_size too low.
Keep PostgreSQL from sometimes choosing a bad query plan
Below is a postgres query that seems to be taking far longer than I would expect. The field_instances table is indexed on both form_instance_id and field_id, and the form_instances table is indexed on workflow_state. So I thought it would be a fast query, but it takes forever. Can anybody help me interpret the query plan and what kinds of indexes to add to speed it up? Thanks.
explain analyze
select form_id,form_instance_id,answer,field_id
from form_instances,field_instances
where workflow_state = 'DRqueued'
and form_instance_id = form_instances.id
and field_id = 'Book_EstimatedDueDate';
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Hash Join (cost=8733.85..95692.90 rows=9277 width=29) (actual time=2550.000..15430.000 rows=11431 loops=1)
Hash Cond: (field_instances.form_instance_id = form_instances.id)
-> Bitmap Heap Scan on field_instances (cost=2681.11..89071.72 rows=47567 width=25) (actual time=850.000..13690.000 rows=51726 loops=1)
Recheck Cond: ((field_id)::text = 'Book_EstimatedDueDate'::text)
-> Bitmap Index Scan on index_field_instances_on_field_id (cost=0.00..2669.22 rows=47567 width=0) (actual time=830.000..830.000 rows=51729 loops=1)
Index Cond: ((field_id)::text = 'Book_EstimatedDueDate'::text)
-> Hash (cost=5911.34..5911.34 rows=11312 width=8) (actual time=1590.000..1590.000 rows=11431 loops=1)
-> Bitmap Heap Scan on form_instances (cost=511.94..5911.34 rows=11312 width=8) (actual time=720.000..1570.000 rows=11431 loops=1)
Recheck Cond: ((workflow_state)::text = 'DRqueued'::text)
-> Bitmap Index Scan on index_form_instances_on_workflow_state (cost=0.00..509.11 rows=11312 width=0) (actual time=650.000..650.000 rows=11509 loops=1)
Index Cond: ((workflow_state)::text = 'DRqueued'::text)
Total runtime: 15430.000 ms
(12 rows)
When you say The field_instances table is indexed on both form_instance_id and field_id you mean that there are separate indexes on form_instance_id and field_id on that table?
Try dropping the index on form_instance_id and put a concatenated index on (form_instance_id, field_id).
An index works by giving you a quick lookup that tells you where the rows are that match your index. It then has to fetch through those rows to do what you want. So you always want your index to be as specific as possible. If you put two indexes on the table, you'll have two different ways to do a lookup, but a query will usually only take advantage of one of them. If you put a concatenated index on the table, you'll be able to look up on the first field in the index, the first two fields, etc efficiently. (So a concatenated index on (a, b) gives you fast lookups on a, even faster lookups on both a and b, but doesn't help you look things up on b)
Right now it is figuring out all possible things in form_instances that have the right state. It separately figures out all of the field_instances that have the right field id. It then does a hash join. For this makes a lookup hash from one result set, and scans the other for matches.
With my suggestion it should figure out all possible form_instances of interest. It will then go to the index, and figure out all of the field_instances that match on both the form instance and field id, and then it will find exactly the results of interest. Because the index is more specific, the database will have fewer rows of data to deal with to process your query.
http://explain.depesz.com is a fantastic online tool that helps you identify the hot spots visually. I pasted your results into the tool and got this analysis: http://explain.depesz.com/s/VIk
It's hard to tell anything specifically without seeing your tables and indexes, however.
Need to know the data you have in your table however just from looking at the sql and column names I would recommend
do you really need an index on workflow_state assuming elements within it can't be very unique - this might not improve select but will insert or update...
try making field_id check the first condition in your where statement