8

I have a table that contains data that is extracted from text documents. The data is stored in a column called "CONTENT" for which I have created this index using GIN:

CREATE INDEX "File_contentIndex"
  ON "File"
  USING gin
  (setweight(to_tsvector('english'::regconfig
           , COALESCE("CONTENT", ''::character varying)::text), 'C'::"char"));

I use the following query to perform a full text search on the table:

SELECT "ITEMID",
  ts_rank(setweight(to_tsvector('english', coalesce("CONTENT",'')), 'C') , 
  plainto_tsquery('english', 'searchTerm')) AS "RANK"
FROM "File"
WHERE setweight(to_tsvector('english', coalesce("CONTENT",'')), 'C') 
  @@ plainto_tsquery('english', 'searchTerm')
ORDER BY "RANK" DESC
LIMIT 5;

The File table contains 250 000 rows and each "CONTENT" entry consists of one random word and a text string that is the same for all rows.

Now, when I search for a random word (1 hit in the whole table) the query runs very fast (<100 ms). However, when I search for a word that is present in all the rows, the query runs extremely slow (10 minutes or more).

EXPLAIN ANALYZE shows that for the 1-hit search a Bitmap Index Scan followed by a Bitmap Heap Scan is performed. For the slow search a Seq Scan is performed instead, which is what is taking so long.

Granted, it is not realistic to have the same data in all rows. But since I can't control the text documents that's uploaded by the users, nor the searches they perform it is possible that a similar scenario arises (search on terms with very high occurrence in DB). How can I increase the performance of my search query for such a scenario?

Running PostgreSQL 9.3.4

Query plans from EXPLAIN ANALYZE:

Quick search (1 hit in DB)

"Limit  (cost=2802.89..2802.90 rows=5 width=26) (actual time=0.037..0.037 rows=1 loops=1)"
"  ->  Sort  (cost=2802.89..2806.15 rows=1305 width=26) (actual time=0.037..0.037 rows=1 loops=1)"
"        Sort Key: (ts_rank(setweight(to_tsvector('english'::regconfig, (COALESCE("CONTENT", ''::character varying))::text), 'C'::"char"), '''wfecg'''::tsquery))"
"        Sort Method: quicksort  Memory: 25kB"
"        ->  Bitmap Heap Scan on "File"  (cost=38.12..2781.21 rows=1305 width=26) (actual time=0.030..0.031 rows=1 loops=1)"
"              Recheck Cond: (setweight(to_tsvector('english'::regconfig, (COALESCE("CONTENT", ''::character varying))::text), 'C'::"char") @@ '''wfecg'''::tsquery)"
"              ->  Bitmap Index Scan on "File_contentIndex"  (cost=0.00..37.79 rows=1305 width=0) (actual time=0.012..0.012 rows=1 loops=1)"
"                    Index Cond: (setweight(to_tsvector('english'::regconfig, (COALESCE("CONTENT", ''::character varying))::text), 'C'::"char") @@ '''wfecg'''::tsquery)"
"Total runtime: 0.069 ms"

Slow search (250k hits in DB)

"Limit  (cost=14876.82..14876.84 rows=5 width=26) (actual time=519667.404..519667.405 rows=5 loops=1)"
"  ->  Sort  (cost=14876.82..15529.37 rows=261017 width=26) (actual time=519667.402..519667.402 rows=5 loops=1)"
"        Sort Key: (ts_rank(setweight(to_tsvector('english'::regconfig, (COALESCE("CONTENT", ''::character varying))::text), 'C'::"char"), '''cyberspace'''::tsquery))"
"        Sort Method: top-N heapsort  Memory: 25kB"
"        ->  Seq Scan on "File"  (cost=0.00..10541.43 rows=261017 width=26) (actual time=2.097..519465.953 rows=261011 loops=1)"
"              Filter: (setweight(to_tsvector('english'::regconfig, (COALESCE("CONTENT", ''::character varying))::text), 'C'::"char") @@ '''cyberspace'''::tsquery)"
"              Rows Removed by Filter: 6"
"Total runtime: 519667.429 ms"
4
  • 1
    From the top of my head: GIN indexes have received major improvements in Postgres 9.4 (and some more in the upcoming 9.5). It will certainly pay to upgrade to the current 9.4. And I would also investigate the performance of GiST instead of the GIN index. The culprit in your query is ORDER BY "RANK" DESC. I would investigate pg_trgm with GiST index and the similarity / distance operators as alternative. Consider: dba.stackexchange.com/questions/56224/…. Might even produce "better" results (besides being faster). Commented Sep 11, 2015 at 0:47
  • What OS are you running your PostgreSQL instance on?
    – Kassandry
    Commented Sep 13, 2015 at 20:01
  • Can you repeat these with explain (analyze, buffers), preferably with track_io_timing set to ON? There is no way it should take 520 seconds to seq scan that table, unless you have it stored on a RAID of floppy disks. Something is definitely pathological there. Also, what is your setting for random_page_cost, and the other cost parameters?
    – jjanes
    Commented Sep 14, 2015 at 18:11
  • @danjo I am facing the same issues even when I do not use ordering . Can you tell me how you fixed it ?
    – Sahil Bahl
    Commented Sep 20, 2017 at 2:35

2 Answers 2

11
+150

Questionable use case

...each CONTENT entry consists of one random word and a text string that is the same for all rows.

A text string that is the same for all rows is just dead freight. Remove it and concatenate it in a view if you need to show it.

Obviously, you are aware of that:

Granted, it is not realistic ... But since I can't control the text ...

Upgrade your Postgres version

Running PostgreSQL 9.3.4

While still on Postgres 9.3, you should at least upgrade to the latest point release (currently 9.3.9). The official recommendation of the project:

We always recommend that all users run the latest available minor release for whatever major version is in use.

Better yet, upgrade to 9.4 which has received major improvements for GIN indexes.

Major problem 1: Cost estimates

The cost of some textsearch functions has been seriously underestimated up to and including version 9.4. That cost is raised by factor 100 in the upcoming version 9.5 like @jjanes describes in his recent answer:

Here are the respective thread where this was discussed and the commit message by Tom Lane.

As you can see in the commit message, to_tsvector() is among those functions. You can apply the change immediately (as superuser):

ALTER FUNCTION to_tsvector (regconfig, text) COST 100;

which should make it much more likely that your functional index is used.

Major problem 2: KNN

The core problem is that Postgres has to calculate a rank with ts_rank() for 260k rows (rows=261011) before it can order by and pick the top 5. This is going to be expensive, even after you have fixed other problems as discussed. It's a K-nearest-neighbour (KNN) problem by nature and there are solutions for related cases. But I cannot think of a general solution for your case, since the rank calculation itself depends on user input. I would try to eliminate the bulk of low ranking matches early so that the full calculation only has to be done for few good candidates.

One way I can think of is to combine your fulltext search with trigram similarity search - which offers a working implementation for the KNN problem. This way you can pre-select the "best" matches with LIKE predicate as candidates (in a subquery with LIMIT 50 for example) and then pick the 5 top-ranking rows according to your rank-calculation in the main query.

Or apply both predicates in the same query and pick the closest matches according to trigram similarity (which would produce different results) like in this related answer:

I did some more research and you are not the first to run into this problem. Related posts on pgsql-general:

Work is ongoing to eventually implement a tsvector <-> tsquery operator.

Oleg Bartunov and Alexander Korotkov even presented a working prototype (using >< as operator instead of <-> back then) but it's very complex to integrate in Postgres, the whole infrastructure for GIN indexes has to be reworked (most of which is done by now).

Major problem 3: weights and index

And I identified one more factor adding to the slowness of the query. Per documentation:

GIN indexes are not lossy for standard queries, but their performance depends logarithmically on the number of unique words. (However, GIN indexes store only the words (lexemes) of tsvector values, and not their weight labels. Thus a table row recheck is needed when using a query that involves weights.)

Bold emphasis mine. As soon as weight are involved, each row has to be fetched from the heap (not just a cheap visibility check) and long values have to be de-toasted, which adds to the cost. But there seems to be a solution for that:

Index definition

Looking at your index again, it doesn't seem to make sense to begin with. You assign a weight to a single column, which is meaningless as long as you don't concatenate other columns with a different weight.

COALESCE() also makes no sense as long as you don't actually concatenate more columns.

Simplify your index:

CREATE INDEX "File_contentIndex" ON "File" USING gin
(to_tsvector('english', "CONTENT");

And your query:

SELECT "ITEMID", ts_rank(to_tsvector('english', "CONTENT")
                       , plainto_tsquery('english', 'searchTerm')) AS rank
FROM   "File"
WHERE  to_tsvector('english', "CONTENT")
       @@ plainto_tsquery('english', 'searchTerm')
ORDER  BY rank DESC
LIMIT  5;

Still expensive for a search term that matches every row, but probably much less.

Asides

All of these issues combined, the insane cost of 520 seconds for your second query is beginning to make sense. But there still may be more problems. Did you configure your server?
All the usual advice for performance optimization applies.

It makes your life easier if you don't work with double-quotes CaMeL-case identifiers:

2
  • I am running into this as well. With Postgresql 9.6, we use query rewriting for synonyms so I don't think using trigram similarity search to limit the number of documents will work well. Commented Aug 9, 2017 at 20:07
  • Amazing! USING gin (to_tsvector('english', "CONTENT")
    – Kerem
    Commented Feb 23, 2019 at 10:13
1

I had similar problem. I took care of it by precomputing the ts_rank of every popular text query term against a field:table tuple and storing it in a lookup table. This saved me a lot of time(factor of 40X) during search of popular words in the text heavy corpus.

  1. Get popular words in the corpus by scanning the doc and counting its occurrences.
  2. Sort by most popular word.
  3. precompute ts_rank of the popular words and store it in a table.

Query: lookup this table and get the sorted document ids by their respective rank. if not there, do it the old way.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.