Pattern matching with LIKE, SIMILAR TO or regular expressions in PostgreSQL

  • I had to write a simple query where I go looking for people's name that start with a B or a D :

    SELECT s.name 
    FROM spelers s 
    WHERE s.name LIKE 'B%' OR s.name LIKE 'D%'
    ORDER BY 1
    

    I was wondering if there is a way to rewrite this to become more performant. So I can avoid or and / or like?

    Why are you trying to rewrite? Performance? Neatness? Is `s.name` indexed?

    I want to write for performance, s.name is not indexed.

    Well as you are searching without leading wild cards and not selecting any additional columns an index on `name` could be useful here if you care about performance.

  • Your query is pretty much the optimum. Syntax won't get much shorter, query won't get much faster:

    SELECT name
    FROM   spelers
    WHERE  name LIKE 'B%' OR name LIKE 'D%'
    ORDER  BY 1;
    

    If you really want to shorten the syntax, use a regular expression with branches:

    ...
    WHERE  name ~ '^(B|D).*'
    

    Or slightly faster, with a character class:

    ...
    WHERE  name ~ '^[BD].*'
    

    A quick test without index yields faster results than for SIMILAR TO in either case for me.
    With an appropriate B-Tree index in place, LIKE wins this race by orders of magnitude.

    Read the basics about pattern matching in the manual.

    Index for superior performance

    If you are concerned with performance, create an index like this for bigger tables:

    CREATE INDEX spelers_name_special_idx ON spelers (name text_pattern_ops);
    

    Makes this kind of query faster by orders of magnitude. Special considerations apply for locale-specific sort order. Read more about operator classes in the manual. If you are using the standard "C" locale (most people don't), a plain index (with default operator class) will do.

    Such an index is only good for left-anchored patterns (matching from the start of the string).

    SIMILAR TO or regular expressions with basic left-anchored expressions can use this index, too. But not with branches (B|D) or character classes [BD] (at least in my tests on PostgreSQL 9.0).

    Trigram matches or text search use special GIN or GiST indexes.

    Overview of pattern matching operators

    • LIKE (~~) is simple and fast but limited in its capabilities.
      ILIKE (~~*) the case insensitive variant.
      pg_trgm extends index support for both.

    • ~ (regular expression match) is powerful but more complex and may be slow for anything more than basic expressions.

    • SIMILAR TO is just pointless. A peculiar halfbreed of LIKE and regular expressions. I never use it. See below.

    • % is the "similarity" operator, provided by the additional module pg_trgm. See below.

    • @@ is the text search operator. See below.

    pg_trgm - trigram matching

    Beginning with PostgreSQL 9.1 you can facilitate the extension pg_trgm to provide index support for any LIKE / ILIKE pattern (and simple regexp patterns with ~) using a GIN or GiST index.

    Details, example and links:

    pg_trgm also provides these operators:

    • % - the "similarity" operator
    • <% (commutator: %>) - the "word_similarity" operator in Postgres 9.6 or later
    • <<% (commutator: %>>) - the "strict_word_similarity" operator in Postgres 11 or later

    Text search

    Is a special type of pattern matching with separate infrastructure and index types. It uses dictionaries and stemming and is a great tool to find words in documents, especially for natural languages.

    Prefix matching is also supported:

    As well as phrase search since Postgres 9.6:

    Consider the introduction in the manual and the overview of operators and functions.

    Additional tools for fuzzy string matching

    The additional module fuzzystrmatch offers some more options, but performance is generally inferior to all of the above.

    In particular, various implementations of the levenshtein() function may be instrumental.

    Why are regular expressions (~) always faster than SIMILAR TO?

    The answer is simple. SIMILAR TO expressions are rewritten into regular expressions internally. So, for every SIMILAR TO expression, there is at least one faster regular expression (that saves the overhead of rewriting the expression). There is no performance gain in using SIMILAR TO ever.

    And simple expressions that can be done with LIKE (~~) are faster with LIKE anyway.

    SIMILAR TO is only supported in PostgreSQL because it ended up in early drafts of the SQL standard. They still haven't gotten rid of it. But there are plans to remove it and include regexp matches instead - or so I heard.

    EXPLAIN ANALYZE reveals it. Just try with any table yourself!

    EXPLAIN ANALYZE SELECT * FROM spelers WHERE name SIMILAR TO 'B%';
    

    Reveals:

    ...  
    Seq Scan on spelers  (cost= ...  
      Filter: (name ~ '^(?:B.*)$'::text)
    

    SIMILAR TO has been rewritten with a regular expression (~).

    Ultimate performance for this particular case

    But EXPLAIN ANALYZE reveals more. Try, with the afore-mentioned index in place:

    EXPLAIN ANALYZE SELECT * FROM spelers WHERE name ~ '^B.*;
    

    Reveals:

    ...
     ->  Bitmap Heap Scan on spelers  (cost= ...
           Filter: (name ~ '^B.*'::text)
            ->  Bitmap Index Scan on spelers_name_text_pattern_ops_idx (cost= ...
                  Index Cond: ((prod ~>=~ 'B'::text) AND (prod ~<~ 'C'::text))
    

    Internally, with an index that is not locale-aware (text_pattern_ops or using locale C) simple left-anchored expressions are rewritten with these text pattern operators: ~>=~, ~<=~, ~>~, ~<~. This is the case for ~, ~~ or SIMILAR TO alike.

    The same is true for indexes on varchar types with varchar_pattern_ops or char with bpchar_pattern_ops.

    So, applied to the original question, this is the fastest possible way:

    SELECT name
    FROM   spelers  
    WHERE  name ~>=~ 'B' AND name ~<~ 'C'
        OR name ~>=~ 'D' AND name ~<~ 'E'
    ORDER  BY 1;
    

    Of course, if you should happen to search for adjacent initials, you can simplify further:

    WHERE  name ~>=~ 'B' AND name ~<~ 'D'   -- strings starting with B or C
    

    The gain over plain use of ~ or ~~ is tiny. If performance isn't your paramount requirement, you should just stick with the standard operators - arriving at what you already have in the question.

    The OP doesn't have an index on name but do you happen to know, if they did, would their original query involve 2 range seeks and `similar` a scan?

    @MartinSmith: A quick test with `EXPLAIN ANALYZE` shows 2 bitmap index scans. Multiple bitmap index scans can be combined rather quickly.

    Thanks. So would there be any milage with replacing the `OR` with `UNION ALL` or replacing `name LIKE 'B%'` with `name >= 'B' AND name <'C'` in Postgres?

    The weird thing is similar was faster :/

    @LucasKauffman - But if you create the index and try again sounds like you may get a different result :-)

    I don't have enough rights on the database to create one so unfortunately I can't test it :(

    @MartinSmith: `UNION` won't but, yes, combining the ranges into one `WHERE` clause will speed up the query. I have added more to my answer. Of course, you have to take your locale into account. Locale-aware search is always slower.

    @LucasKauffman: Try `WHERE name ~ '^[BD].*'`. That's faster than `SIMILAR TO` under all circumstances - as I have explained in my amended answer.

    @a_horse_with_no_name: I expect not. The new capabilities of pg_tgrm with GIN indexes are a treat for generic text search. A search anchored at the start is already faster than that.

    Is there a way to combine regular expression constraints such as ^ or $ with a field name, when you compare two fields? Example: (...) WHERE field1 ~ ^field2;

    @SébastienClément: Yes, concatenation. The regexp pattern is a plain string. So: `WHERE field1 ~ ('^' || field2)`. You may or may not want to escape characters with special meaning in `field2`. See: https://stackoverflow.com/a/45741630/939860

    @Erwin Brandstetter: fantastic, I had just forgotten the parentheses, thanks!

    Thanks very much for the note on `text_pattern_ops` as the operator class for the index. I have a table where the primary key is text but with with `text_ops` and Postgres prefers a scan over the whole table rather than using the index.

  • How about adding a column to the table. Depending on your actual requirements:

    person_name_start_with_B_or_D (Boolean)
    
    person_name_start_with_char CHAR(1)
    
    person_name_start_with VARCHAR(30)
    

    PostgreSQL doesn't support computed columns in base tables a la SQL Server but the new column can be maintained via trigger. Obviously, this new column would be indexed.

    Alternatively, an index on an expression would give you the same, cheaper. E.g.:

    CREATE INDEX spelers_name_initial_idx ON spelers (left(name, 1)); 
    

    Queries that match the expression in their conditions can utilize this index.

    This way, the performance hit is taken when the data is created or amended, so may only be appropriate for a low activity environment (i.e. much fewer writes than reads).

  • You could try

    SELECT s.name
    FROM   spelers s
    WHERE  s.name SIMILAR TO '(B|D)%' 
    ORDER  BY s.name
    

    I've no idea whether or not either the above or your original expression are sargable in Postgres though.

    If you create the suggested index would also be interested to hear how this compares with the other options.

    SELECT name
    FROM   spelers
    WHERE  name >= 'B' AND name < 'C'
    UNION ALL
    SELECT name
    FROM   spelers
    WHERE  name >= 'D' AND name < 'E'
    ORDER  BY name
    

    It worked and I got a cost of 1.19 where I had 1.25. Thanks !

  • What I have done in the past, faced with a similar performance issue, is to increment the ASCII character of the last letter, and do a BETWEEN. You then get the best performance, for a subset of the LIKE functionality. Of course, it only works in certain situations, but for ultra-large datasets where you're searching on a name for instance, it makes performance go from abysmal to acceptable.

  • Very old question, but I found another fast solution to this problem:

    SELECT s.name 
    FROM spelers s 
    WHERE ascii(s.name) in (ascii('B'),ascii('D'))
    ORDER BY 1
    

    Since function ascii() looks only at first character of the string.

    Does this use an index on `(name)`?

  • For checking of initials, I often use casting to "char" (with the double quotes). It's not portable, but very fast. Internally, it simply detoasts the text and returns the first character, and "char" comparison operations are very fast because the type is 1-byte fixed length:

    SELECT s.name 
    FROM spelers s 
    WHERE s.name::"char" =ANY( ARRAY[ "char" 'B', 'D' ] )
    ORDER BY 1
    

    Note that casting to "char" is faster than the ascii() slution by @Sole021, but it is not UTF8 compatible (or any other encoding for that matter), returning simply the first byte, so should only be used in cases where the comparison is against plain old 7-bit ASCII characters.

  • There are two methods not mentioned yet for dealing with such cases:

    1. partial (or partitioned - if created for full range manually) index - most useful when only a subset of data is required (for example during some maintenance or temporary for some reporting):

      CREATE INDEX ON spelers WHERE name LIKE 'B%'
      
    2. partitioning the table itself (using the first character as partitioning key) - this technique is especially worth considering in PostgreSQL 10+ (less painful partitioning) and 11+ (partition pruning during query execution).

    Moreover, if the data in a table is sorted, one can benefit from using BRIN index (over the first character).

  • Probably faster to do a single character comparison:

    SUBSTR(s.name,1,1)='B' OR SUBSTR(s.name,1,1)='D'
    

    Not really. `column LIKE 'B%'` will be more efficient than using substring function on the column.

License under CC-BY-SA with attribution


Content dated before 6/26/2020 9:53 AM