Skip to content
Advertisement

How do you do date math that ignores the year?

I am trying to select dates that have an anniversary in the next 14 days. How can I select based on dates excluding the year? I have tried something like the following.

SELECT * FROM events
WHERE EXTRACT(month FROM "date") = 3
AND EXTRACT(day FROM "date") < EXTRACT(day FROM "date") + 14

The problem with this is that months wrap.
I would prefer to do something like this, but I don’t know how to ignore the year.

SELECT * FROM events
WHERE (date > '2013-03-01' AND date < '2013-04-01')

How can I accomplish this kind of date math in Postgres?

Advertisement

Answer

If you don’t care for explanation and details, use the “Black magic version” below.

All queries presented in other answers so far operate with conditions that are not sargable – they cannot use an index and have to compute an expression for every single row in the base table to find matching rows. Doesn’t matter much with small tables. Matters (a lot) with big tables.

Given the following simple table:

CREATE TABLE event (
  event_id   serial PRIMARY KEY
, event_date date
);

Query

Version 1. and 2. below can use a simple index of the form:

CREATE INDEX event_event_date_idx ON event(event_date);

But all of the following solutions are even faster without index.

1. Simple version

SELECT *
FROM  (
   SELECT ((current_date + d) - interval '1 year' * y)::date AS event_date
   FROM       generate_series( 0,  14) d
   CROSS JOIN generate_series(13, 113) y
   ) x
JOIN  event USING (event_date);

Subquery x computes all possible dates over a given range of years from a CROSS JOIN of two generate_series() calls. The selection is done with the final simple join.

2. Advanced version

WITH val AS (
   SELECT extract(year FROM age(current_date + 14, min(event_date)))::int AS max_y
        , extract(year FROM age(current_date,      max(event_date)))::int AS min_y
   FROM   event
   )
SELECT e.*
FROM  (
   SELECT ((current_date + d.d) - interval '1 year' * y.y)::date AS event_date
   FROM   generate_series(0, 14) d
        ,(SELECT generate_series(min_y, max_y) AS y FROM val) y
   ) x
JOIN  event e USING (event_date);

Range of years is deduced from the table automatically – thereby minimizing generated years.
You could go one step further and distill a list of existing years if there are gaps.

Effectiveness co-depends on the distribution of dates. Few years with many rows each make this solution more useful. Many years with few rows each make it less useful.

Simple SQL Fiddle to play with.

3. Black magic version

Updated 2016 to remove a “generated column”, which would block H.O.T. updates; simpler and faster function.
Updated 2018 to calculate MMDD with IMMUTABLE expressions to allow function inlining.

Create a simple SQL function to calculate an integer from the pattern 'MMDD':

CREATE FUNCTION f_mmdd(date) RETURNS int LANGUAGE sql IMMUTABLE AS
'SELECT (EXTRACT(month FROM $1) * 100 + EXTRACT(day FROM $1))::int';

I had to_char(time, 'MMDD') at first, but switched to the above expression which proved fastest in new tests on Postgres 9.6 and 10:

db<>fiddle here

It allows function inlining because EXTRACT (xyz FROM date) is implemented with the IMMUTABLE function date_part(text, date) internally. And it has to be IMMUTABLE to allow its use in the following essential multicolumn expression index:

CREATE INDEX event_mmdd_event_date_idx ON event(f_mmdd(event_date), event_date);

Multicolumn for a number of reasons:
Can help with ORDER BY or with selecting from given years. Read here. At almost no additional cost for the index. A date fits into the 4 bytes that would otherwise be lost to padding due to data alignment. Read here.
Also, since both index columns reference the same table column, no drawback with regard to H.O.T. updates. Read here.

One PL/pgSQL table function to rule them all

Fork to one of two queries to cover the turn of the year:

CREATE OR REPLACE FUNCTION f_anniversary(date = current_date, int = 14)
  RETURNS SETOF event AS
$func$
DECLARE
   d  int := f_mmdd($1);
   d1 int := f_mmdd($1 + $2 - 1);  -- fix off-by-1 from upper bound
BEGIN
   IF d1 > d THEN
      RETURN QUERY
      SELECT *
      FROM   event e
      WHERE  f_mmdd(e.event_date) BETWEEN d AND d1
      ORDER  BY f_mmdd(e.event_date), e.event_date;

   ELSE  -- wrap around end of year
      RETURN QUERY
      SELECT *
      FROM   event e
      WHERE  f_mmdd(e.event_date) >= d OR
             f_mmdd(e.event_date) <= d1
      ORDER  BY (f_mmdd(e.event_date) >= d) DESC, f_mmdd(e.event_date), event_date;
      -- chronological across turn of the year
   END IF;
END
$func$  LANGUAGE plpgsql;

Call using defaults: 14 days beginning “today”:

SELECT * FROM f_anniversary();

Call for 7 days beginning ‘2014-08-23’:

SELECT * FROM f_anniversary(date '2014-08-23', 7);

SQL Fiddle comparing EXPLAIN ANALYZE.

February 29

When dealing with anniversaries or “birthdays”, you need to define how to deal with the special case “February 29” in leap years.

When testing for ranges of dates, Feb 29 is usually included automatically, even if the current year is not a leap year. The range of days is extended by 1 retroactively when it covers this day.
On the other hand, if the current year is a leap year, and you want to look for 15 days, you may end up getting results for 14 days in leap years if your data is from non-leap years.

Say, Bob is born on the 29th of February:
My query 1. and 2. include February 29 only in leap years. Bob has birthday only every ~ 4 years.
My query 3. includes February 29 in the range. Bob has birthday every year.

There is no magical solution. You have to define what you want for every case.

Test

To substantiate my point I ran an extensive test with all the presented solutions. I adapted each of the queries to the given table and to yield identical results without ORDER BY.

The good news: all of them are correct and yield the same result – except for Gordon’s query that had syntax errors, and @wildplasser’s query that fails when the year wraps around (easy to fix).

Insert 108000 rows with random dates from the 20th century, which is similar to a table of living people (13 or older). 

INSERT INTO  event (event_date)
SELECT '2000-1-1'::date - (random() * 36525)::int
FROM   generate_series (1, 108000);

Delete ~ 8 % to create some dead tuples and make the table more “real life”. 

DELETE FROM event WHERE random() < 0.08;
ANALYZE event;

My test case had 99289 rows, 4012 hits.

C – Catcall

WITH anniversaries as (
   SELECT event_id, event_date
         ,(event_date + (n || ' years')::interval)::date anniversary
   FROM   event, generate_series(13, 113) n
   )
SELECT event_id, event_date -- count(*)   --
FROM   anniversaries
WHERE  anniversary BETWEEN current_date AND current_date + interval '14' day;

C1 – Catcall’s idea rewritten

Aside from minor optimizations, the major difference is to add only the exact amount of years date_trunc('year', age(current_date + 14, event_date)) to get this year’s anniversary, which avoids the need for a CTE altogether:

SELECT event_id, event_date
FROM   event
WHERE (event_date + date_trunc('year', age(current_date + 14, event_date)))::date
       BETWEEN current_date AND current_date + 14;

D – Daniel

SELECT *   -- count(*)   -- 
FROM   event
WHERE  extract(month FROM age(current_date + 14, event_date))  = 0
AND    extract(day   FROM age(current_date + 14, event_date)) <= 14;

E1 – Erwin 1

See “1. Simple version” above.

E2 – Erwin 2

See “2. Advanced version” above.

E3 – Erwin 3

See “3. Black magic version” above.

G – Gordon

SELECT * -- count(*)   
FROM  (SELECT *, to_char(event_date, 'MM-DD') AS mmdd FROM event) e
WHERE  to_date(to_char(now(), 'YYYY') || '-'
                 || (CASE WHEN mmdd = '02-29' THEN '02-28' ELSE mmdd END)
              ,'YYYY-MM-DD') BETWEEN date(now()) and date(now()) + 14;

H – a_horse_with_no_name

WITH upcoming as (
   SELECT event_id, event_date
         ,CASE 
            WHEN date_trunc('year', age(event_date)) = age(event_date)
                 THEN current_date
            ELSE cast(event_date + ((extract(year FROM age(event_date)) + 1)
                      * interval '1' year) AS date) 
          END AS next_event
   FROM event
   )
SELECT event_id, event_date
FROM   upcoming
WHERE  next_event - current_date  <= 14;

W – wildplasser

CREATE OR REPLACE FUNCTION this_years_birthday(_dut date) RETURNS date AS
$func$
DECLARE
    ret date;
BEGIN
    ret :=
    date_trunc( 'year' , current_timestamp)
        + (date_trunc( 'day' , _dut)
         - date_trunc( 'year' , _dut));
    RETURN ret;
END
$func$ LANGUAGE plpgsql;

Simplified to return the same as all the others:

SELECT *
FROM   event e
WHERE  this_years_birthday( e.event_date::date )
        BETWEEN current_date
        AND     current_date + '2weeks'::interval;

W1 – wildplasser’s query rewritten

The above suffers from a number of inefficient details (beyond the scope of this already sizable post). The rewritten version is much faster:

CREATE OR REPLACE FUNCTION this_years_birthday(_dut INOUT date) AS
$func$
SELECT (date_trunc('year', now()) + ($1 - date_trunc('year', $1)))::date
$func$ LANGUAGE sql;

SELECT *
FROM   event e
WHERE  this_years_birthday(e.event_date)
        BETWEEN current_date
        AND    (current_date + 14);

Test results

I ran this test with a temporary table on PostgreSQL 9.1.7. Results were gathered with EXPLAIN ANALYZE, best of 5.

Results

Without index
C:  Total runtime: 76714.723 ms
C1: Total runtime:   307.987 ms  -- !
D:  Total runtime:   325.549 ms
E1: Total runtime:   253.671 ms  -- !
E2: Total runtime:   484.698 ms  -- min() & max() expensive without index
E3: Total runtime:   213.805 ms  -- !
G:  Total runtime:   984.788 ms
H:  Total runtime:   977.297 ms
W:  Total runtime:  2668.092 ms
W1: Total runtime:   596.849 ms  -- !

With index
E1: Total runtime:    37.939 ms  --!!
E2: Total runtime:    38.097 ms  --!!

With index on expression
E3: Total runtime:    11.837 ms  --!!

All other queries perform the same with or without index because they use non-sargable expressions.

Conclusio

  • So far, @Daniel’s query was the fastest.

  • @wildplassers (rewritten) approach performs acceptably, too.

  • @Catcall’s version is something like the reverse approach of mine. Performance gets out of hand quickly with bigger tables.
    The rewritten version performs pretty well, though. The expression I use is something like a simpler version of @wildplassser’s this_years_birthday() function.

  • My “simple version” is faster even without index, because it needs fewer computations.

  • With index, the “advanced version” is about as fast as the “simple version”, because min() and max() become very cheap with an index. Both are substantially faster than the rest which cannot use the index.

  • My “black magic version” is fastest with or without index. And it is very simple to call.
    The updated version (after the benchmark) is a bit faster, yet.

  • With a real life table an index will make even greater difference. More columns make the table bigger, and sequential scan more expensive, while the index size stays the same.

User contributions licensed under: CC BY-SA
5 People found this is helpful
Advertisement