User-defined Order in SQL


Some applications, such as todo lists, need to maintain a user-defined order of items. The challenge is that the order is arbitrary and can change when the user rearranges items:

reordering a list in a gui

reordering a list in a gui

This article investigates the best way to model the situation in SQL. We’ll examine several ways to do it, and will assess three properties for each:

  • Efficiency in space and time. Storing row order should be compact on disk, and re-ordering items should use minimal CPU and I/O resources.
  • Robustness. There should be no practical limit to how many times items can be re-ordered.
  • Elegance. The solution shouldn’t require complicated PL/pgSQL functions or parsing lists of numbers in strings.

Approach 1: Integer Position Column

The most natural first attempt is to add an auto-incrementing integer column to track each item’s position:

create table todos ( task text, pos serial, -- <== add this unique (pos)
);

Populating the list is easy:

insert into todos (task) values ('experiment with sql'), ('write article'), ('relax'), ('repeat'); select * from todos order by pos asc;
/*
┌─────────────────────┬─────┐
│ task │ pos │
├─────────────────────┼─────┤
│ experiment with sql │ 1 │
│ write article │ 2 │
│ relax │ 3 │
│ repeat │ 4 │
└─────────────────────┴─────┘
*/

What’s more difficult is inserting items within the list, or reordering existing items. Suppose we want to insert a new “edit article” task between items 2 and 3. This requires shifting items 3 and greater one position ahead, and inserting the item at position 3. But even the first step runs into problems:

-- shift items 3 and greater one position ahead update todos set pos = pos+1 where pos >= 3;
/*
ERROR: 23505: duplicate key value violates unique constraint "todos_pos_key"
DETAIL: Key (pos)=(4) already exists.
*/

The uniqueness constraint makes the update sensitive to when each table row gets processed. In our case it attempted to move item 3 to position 4 without first moving 4 to 5.

We can enable more flexible behavior by deferring the uniqueness constraint inside a transaction.

create table todos ( task text, pos serial, unique (pos) deferrable initially deferred -- ^^^ add this
); -- now we can shift the list and insert an item
begin; update todos set pos = pos+1 where pos >= 3; insert into todos (pos, task) values (3, 'edit article'); -- don't forget to increment the sequence
select nextval('todos_pos_seq'); commit;

How does this technique stack up?

  • Efficient? No. It requires updating a bunch of rows to insert one between others.
  • Robust? Yes. It supports reliably inserting/moving an item between any others. Also, even at the “read committed” isolation level, concurrent item insertion doesn’t result in inconsistencies. (At least in my tests.)
  • Elegant? No. The technique requires a fragile sequence of steps, including deferring a constraint and adjusting a sequence.

What about leaving room?

Rather than using sequential integers, what about leaving space in between them? Just like skipping line numbers in the BASIC programming language, we can skip position values in the table leave room between them. Something like:

create sequence todos_gapped_seq increment by 65536; -- the todos table is declared same as before

To insert an item between two others, simply use the position that is the average of the surrounding items. However with our choice of 2^16 blanks between each item, we can support no more than sixteen consecutive insertions between the first and next item. After reaching this limit we would have revert to the previous approach of shifting items forward.

How does this compare with the sequential integer approach?

  • Efficient? More efficient on average, but must occasionally take the hit of shifting.
  • Robust? Yes, as effective as the earlier approach.
  • Elegant? No, it’s even worse than the earlier approach because of mixed logic.

Approach 2: Decimal Position

What if we store the position of each row using float or numeric values rather than int or bigint? This would allow squeezing new elements between others, rather than shifting items forward to make room. Here’s the idea:

create sequence todos_seq; create table todos ( task text, pos float not null default nextval('todos_seq'), unique (pos)
); insert into todos (task) values ('experiment with sql'), ('write article'), ('relax'), ('repeat');

Now inserting an item between rows 2 and 3 is easy:

insert into todos (pos, task) values ((2.0+3)/2, 'edit article');

It seems like a perfect solution! However floating point numbers have limited precision. For example, if we repeatedly insert between 1000 and 1001, cutting by halves each time, then by the 38th iteration it will truncate to exactly 1000:

┌────┬──────────────────┐
│ i │ val │
├────┼──────────────────┤
│ 25 │ 1000.0000000298 │
│ 26 │ 1000.0000000149 │
│ 27 │ 1000.00000000745 │
│ 28 │ 1000.00000000373 │
│ 29 │ 1000.00000000186 │
│ 30 │ 1000.00000000093 │
│ 31 │ 1000.00000000047 │
│ 32 │ 1000.00000000023 │
│ 33 │ 1000.00000000012 │
│ 34 │ 1000.00000000006 │
│ 35 │ 1000.00000000003 │
│ 36 │ 1000.00000000001 │
│ 37 │ 1000.00000000001 │
│ 38 │ 1000 │
│ 39 │ 1000 │
│ 40 │ 1000 │
└────┴──────────────────┘
  • Efficient? Yes. A simple 64-bit column and some arithmetic makes inserting new elements into the list very fast.
  • Robust? No. It appears to work but eventually fails.
  • Elegant? Yes. Finding midpoints is easy.

Arbitrary Precision

OK, so the float type eventually runs out of precision – what about using the numeric (aka decimal) type in SQL? It has a variable size that grows as needed to store precise numbers. How big does it get? Let’s ask the database:

select (1 / power(2, i))::numeric as val, pg_column_size(1 / power(2, i)::numeric) as sz
from generate_series(1, 21) as i; /*
┌─────────────────────────┬────┐
│ val │ sz │
├─────────────────────────┼────┤
│ 0.5 │ 8 │
│ 0.25 │ 8 │
│ 0.125 │ 8 │
│ 0.0625 │ 8 │
│ 0.03125 │ 10 │
│ 0.015625 │ 10 │
│ 0.0078125 │ 10 │
│ 0.00390625 │ 10 │
│ 0.001953125 │ 12 │
│ 0.0009765625 │ 12 │
│ 0.00048828125 │ 12 │
│ 0.000244140625 │ 12 │
│ 0.0001220703125 │ 14 │
│ 0.00006103515625 │ 12 │
│ 0.000030517578125 │ 12 │
│ 0.0000152587890625 │ 12 │
│ 0.00000762939453125 │ 14 │
│ 0.000003814697265625 │ 14 │
│ 0.0000019073486328125 │ 14 │
│ 0.00000095367431640625 │ 14 │
│ 0.000000476837158203125 │ 16 │
└─────────────────────────┴────┘
*/

It reaches 12 bytes pretty quickly, when in fact cutting between 0 and 1 like this is the best case. As the integer part grows, numeric requires even more bytes. But, aside from disk usage, this is the best approach so far.

  • Efficient? Yes, mostly. It’s unusual for the numeric type to exceed 128 bits if there aren’t many list rearrangements, and most operations on the column will be indexed comparisons, not arithmetic.
  • Robust? Yes. Over a long time the numeric values may continue consuming more space as the list keeps being rearranged, but for all practical purposes you can rearrange the list forever.
  • Elegant? Yes, just like float.

Approach 3: True Fractions

Both the float and numeric types ultimately ran into limitations because of how we picked midpoints using averages. Those midpoints quickly consumed precision. The float couldn’t handle it at all, and the numeric got bloated. However there’s a mathematical trick we can use to get the robustness of a numeric type with the size of a float.

In the course of working on this article, I created a new base type for PostgreSQL called rational. It’s available as a database extension here. It performs exact fractional arithmetic and I designed it to always use exactly 64 bits per fraction, which is the same size as PostgreSQL’s float type. (I took inspiration from the postgres wiki.)

Fractions are much more interesting than they’re presented in school. Non-negative fractions actually form a binary tree, with every fraction (in lowest terms) appearing at a unique node.

stern-brocot tree

stern-brocot tree

This “Stern-Brocot tree” is named after its independent discoverers from the 1860s: a mathematician and a clockmaker. If you want to learn more of the theory, check out this youtube video.

The tree helps us like this: if you want to find a fraction within certain bounds (a < x < b), then traverse the Stern-Brocot tree – constructing it as you go – in a binary search. As soon as you hit a node within bounds, stop. That may sound difficult, but pg_rational has the logic built-in.

-- the "rational" type comes from an extension
-- https://github.com/begriffs/pg_rational create extension pg_rational; create sequence todos_seq as integer; create table todos ( task text, pos rational not null default nextval('todos_seq')::integer, unique (pos)
); insert into todos (task) values ('experiment with sql'), ('write article'), ('relax'), ('repeat'); select * from todos order by pos asc;
/*
┌─────────────────────┬─────┐
│ task │ pos │
├─────────────────────┼─────┤
│ experiment with sql │ 1/1 │
│ write article │ 2/1 │
│ relax │ 3/1 │
│ repeat │ 4/1 │
└─────────────────────┴─────┘
*/

That was as easy as using a float or integer! Inserting a new value is easy too:

insert into todos (pos, task) values (rational_intermediate(2,3), 'edit article'); select * from todos order by pos asc;
/*
┌─────────────────────┬─────┐
│ task │ pos │
├─────────────────────┼─────┤
│ experiment with sql │ 1/1 │
│ write article │ 2/1 │
│ edit article │ 5/2 │
│ relax │ 3/1 │
│ repeat │ 4/1 │
└─────────────────────┴─────┘
*/

The terms of these fractions are expressed in lowest terms and grow slowly at each insertion. For instance you can see from the tree diagram earlier that inserting between 1 and 0 toward 0 generates 1/2, 1/3, 1/4 … which can go a very long time because numerators and denominators in pg_rational each get 32 bits of storage.

  • Efficient? Yes. Each rational uses 64 bits, and the extension is written in C.
  • Robust? Yes. It would take virtually forever to run out of space through repeated list reordering.
  • Elegant? Yes.

A number of observant readers including Michael Wolfe, David Marcin and “arnioxux” on Hacker News, pointed out that there are paths through the Stern-Brocot tree where the fraction terms rapidly increase. For instance continually reversing directions and walking down the middle of the tree (L, R, L, R, L…) increases the denominators in a Fibonacci sequence: 1/1, 1/2, 2/3, 3/5, 5/8, 8/13 etc. The 46th Fibonacci number – 1,836,311,903 – is the largest that can fit in a signed 32-bit integer.

So although finding midpoints with this method still usually beats taking averages, it’s not a silver bullet for any pattern of insertions. It works especially well for repeated inserts in one direction, and worse for a sort of weaving insert around new elements. The pathological pattern does not feel like a common thing for a user to do (since simply swapping the positions of two items can happen by swapping their numbers, not inserting new numbers). Nonetheless I should investigate adding a function in pg_rational to re-normalize a list into simpler fractions.

Approach 4: True Fractions … as floats

This final approach for storing list order is to do calculations with the rational type, but store the result as a float. Rationals in the library can convert back and forth:

-- convert float to rational
select 0.263157894737::float::rational;
-- => 5/19 -- convert rational to float
select '-1/2'::rational::float;
-- => -0.5

So we can define the todos table as in the float method:

create sequence todos_seq; create table todos ( task text, pos float not null default nextval('todos_seq'), unique (pos)
); insert into todos (task) values ('experiment with sql'), ('write article'), ('relax'), ('repeat');

But now insert using rational numbers, which will get coerced to float when saving into the column.

insert into todos (pos, task) values (rational_intermediate(2,3), 'edit article'); select * from todos order by pos asc;
/*
┌─────────────────────┬─────┐
│ task │ pos │
├─────────────────────┼─────┤
│ experiment with sql │ 1 │
│ write article │ 2 │
│ edit article │ 2.5 │
│ relax │ 3 │
│ repeat │ 4 │
└─────────────────────┴─────┘
*/

Floats created by rational_intermediate don’t seem to run up against precision problems like those generated by taking averages. For example taking repeated intermediates from zero going toward one makes this pattern:

select (i, i+1)::ratt::rational AS fract, i::float / (i + 1)::float AS float from generate_series(1, 15) AS i;
/*
┌───────┬───────────────────┐
│ fract │ float │
├───────┼───────────────────┤
│ 1/2 │ 0.5 │
│ 2/3 │ 0.666666666666667 │
│ 3/4 │ 0.75 │
│ 4/5 │ 0.8 │
│ 5/6 │ 0.833333333333333 │
│ 6/7 │ 0.857142857142857 │
│ 7/8 │ 0.875 │
│ 8/9 │ 0.888888888888889 │
│ 9/10 │ 0.9 │
│ 10/11 │ 0.909090909090909 │
│ 11/12 │ 0.916666666666667 │
│ 12/13 │ 0.923076923076923 │
│ 13/14 │ 0.928571428571429 │
│ 14/15 │ 0.933333333333333 │
│ 15/16 │ 0.9375 │
└───────┴───────────────────┘
*/

Floats can handle this sequence for a very long time.

-- how long until floats cannot distinguish successive fractions?
select min(i) from generate_series(1, 100000000) as i where (i::float/(i+1.0)) = ((i::float+1.0)/(i+2.0)); -- => 94911150

Other fractions may pose more difficulties for the float representation, but from my limited experiments it looks like this way of choosing midpoints works a lot better than taking averages.

Ultimately this approach has its pros and cons compared with just using pg_rational. The space on disk is the same. The speed of comparing floats is faster than comparing fractions (a single CPU instruction vs an extra cross-multiplication of integer terms). But finding intermediates requires first turning floats to fractions using an iterative algorithm. This is adds overhead compared to simply storing the fractions.

  • Efficient? For comparisons, yes. But extra overhead for fractional conversion.
  • Robust? Yes? It appears so…
  • Elegant? Yes, but not as straightforward as dealing with fractions alone.

Conclusion

Based on the arguments above, I’d recommend using true fractions with pg_rational.

Tips

Here are a few extra notes to help you get the most out of the extension.

  • Insert the first item as 1/1 rather than 0/1. The algorithm for finding intermediate values can never get below zero, so starting at one will allow you to insert future items before the first item. You could even add a constraint to the position column enforcing that all values are strictly greater than zero.
  • Add a btree index to the position column. The rational type supports an index. You’ll get an index automatically simply by adding a uniqueness constraint to the column.
  • Here’s the easiest way to insert a new item immediately after a given one, cutting earlier than any existing subsequent items:

    -- add item immediately after item 2
    INSERT INTO todos (pos, task)
    SELECT rational_intermediate(2, min(pos)), 'edit article' FROM todos WHERE pos > 2;

    This works even if the item is the very biggest one, since rational_intermediate treats NULL in the second argument as infinity.

  • Inserting an item immediately before another uses similar logic except with max.
  • You probably don’t even want to use a sequence to keep track of the largest item inserted. You can insert at the end of the list by doing this:

    INSERT INTO todos (pos, task)
    SELECT rational_intermediate (max(pos), NULL), 'last one' FROM todos;

    This would work even on an empty table. I used a sequence earlier just to cut down on the code for inserting a batch of items.

  • To relabel the positions and hide the fractional numbering:

    select row_number() over () as pos, task from todos order by pos asc;

    Although sending this to a client application would prevent the client from knowing the true positions in order to ask the database to rearrange them, so the relabeling should probably happen at display time.