Postgres Full Text Search

22 April 2021 ⌚ 4 min read

Today, I want to talk about how easy it is to implement a full-text search in Postgres. Let's dive right into it!

But first, why you would want to implement text search when you can do queries with LIKE or ILIKE. There are many reasons to this:

Full-text search queries against a specialized stemmed version of your data. This means that it can search for complex English words like "justifiable" when searching for "justify".
It is fast, especially combined with indexes. LIKE and ILIKE will have to do a sequential scan of your data to select the results.
You can blacklist certain words from getting parsed.

And much more!

Setup

Postgres Instance

I am using docker to create a new postgres instance on my local machine. To create start a Postgres container, run:

docker run --name=postgresfts -p 5432:5432 -e POSTGRES_PASSWORD=postgres -d postgres

Feel free to use any frontend clients you like to connect to this postgres instance. For this job, psql will do fine.

Sample Data

We need some data to search. For this I am using the example of a books table that looks like this:

                              Table "public.books"
   Column    |  Type   | Collation | Nullable |              Default
-------------+---------+-----------+----------+-----------------------------------
 id          | bigint  |           | not null | nextval('books_id_seq'::regclass)
 title       | text    |           |          |
 description | text    |           |          |
 price       | numeric |           |          |
Indexes:
    "books_pkey" PRIMARY KEY, btree (id)

To create this books table, run:

CREATE TABLE books (id BIGSERIAL PRIMARY KEY, title text, description text, price numeric);

Now that we have our table, let's fill in some data to work with. Inserting few rows are fine. But we are doing a full-text search and it is good to have few thousand rows to spice things up 😉.

For this, I have put together this simple nodejs script:

const { Pool } = require("pg");
const format = require("pg-format");
const faker = require("faker");

const db = new Pool({
  host: "172.17.0.1", // postgres host
  database: "postgres",
  password: "passwd",
  user: "postgres",
});

async function main() {
  let books = [];

  for (let i = 1; i <= 100000; i++) {
    books[i - 1] = [
      faker.commerce.productName(),
      faker.commerce.productDescription(),
      faker.commerce.price(),
    ];
  }

  const query = format(
    "INSERT INTO books (title, description, price) VALUES %L",
    books
  );

  try {
    await db.query(query);
    console.log("Done!");
  } catch (error) {
    console.error(error);
  }
}

main();

The above script will create 100k rows using faker.js and insert them in the books table.

You can also clone this git repo from siddhantk232/node_postgres_scripts.

Let's Search!

Now that we have our data ready in the books table. Let's implement full-text search.

In Postgres, you need two things to start with, one is a tsvector to search and the other is the search query aka the tsquery.

tsvector?

The tsvector can be another column in your table that stores normalized keywords from the actual data.

In this case, I want the book's title and description column to be searchable. So I would create a tsvector of these columns.

Doing that is pretty easy, we can create a generated column of type tsvector.

To do this, run:

ALTER TABLE books ADD COLUMN fts tsvector 
  GENERATED ALWAYS AS (to_tsvector('english', title || ' ' || description)) 
  STORED;

This will add a column called fts to the books table. While inserting you have to ignore this column as this is a generated column.

The first argument to the to_tsvector is the language we are using. The default gets picked up from the instance's locale settings. You can change this to work with other languages.

Now that we have our data and the tsvector, let's do some search!

For searching, we use the @@ operator in Postgres. @@ returns true if the search succeeds and false otherwise.

Example:

SELECT title FROM books WHERE fts @@ plainto_tsquery('chip') LIMIT 2;

I was getting a few thousand results which were hard to show, so I put a limit of two here.

The output looks like this:

           title
---------------------------
 Refined Metal Chips
 Intelligent Granite Chips
(2 rows)

As you can see our search works. We searched for the chip and it returned the first two results it found.

But how fast?

Let's analyze the search, but this time with a new query so that we don't hit the cache and get raw results.

EXPLAIN ANALYZE 
  SELECT id, title, description, price 
  FROM books WHERE fts @@ plainto_tsquery('style');

This will output:

                                                        QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------
 Gather  (cost=1000.00..17860.20 rows=8297 width=150) (actual time=0.578..95.603 rows=8329 loops=1)
   Workers Planned: 2
   Workers Launched: 2
   ->  Parallel Seq Scan on books  (cost=0.00..16030.50 rows=3457 width=150) (actual time=0.461..86.162 rows=2776 loops=3)
         Filter: (fts @@ plainto_tsquery('style'::text))
         Rows Removed by Filter: 30557
 Planning Time: 0.191 ms
 Execution Time: 96.159 ms
(8 rows)

You can see we are getting about 8000 results in about 100ms! This is fast but we can do better.

Index to the rescue!

Postgres recommends using a GIN index. Actually, You can use either GIN or the GiST index type. You can learn the differences between them here.

Let's create a GIN index on our fts column:

CREATE INDEX books_search_idx ON books USING GIN (fts);

You can run \d books to see the created index in the psql client. The output will look something like this:

                                                                Table "public.books"
   Column    |   Type   | Collation | Nullable |                                               Default
-------------+----------+-----------+----------+-----------------------------------------------------------------------------------------------------
 id          | bigint   |           | not null | nextval('books_id_seq'::regclass)
 title       | text     |           |          |
 description | text     |           |          |
 price       | numeric  |           |          |
 fts         | tsvector |           |          | generated always as (to_tsvector('english'::regconfig, (title || ' '::text) || description)) stored
Indexes:
    "books_pkey" PRIMARY KEY, btree (id)
    "books_search_idx" gin (fts)

Let's try our search once again to see how much faster it is:

                                                           QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------
 Bitmap Heap Scan on books  (cost=84.55..7554.73 rows=8297 width=150) (actual time=4.786..17.013 rows=8329 loops=1)
   Recheck Cond: (fts @@ plainto_tsquery('style'::text))
   Heap Blocks: exact=4127
   ->  Bitmap Index Scan on books_search_idx  (cost=0.00..82.47 rows=8297 width=0) (actual time=3.125..3.126 rows=8329 loops=1)
         Index Cond: (fts @@ plainto_tsquery('style'::text))
 Planning Time: 0.425 ms
 Execution Time: 18.217 ms
(7 rows)

8k results in ~20ms in a 100k row table. This is crazy fast!

The End

There is a lot that you can do with text search in Postgres. It is impossible to cover all of them in a single blog.