March | 2017 | Matt Mazur

While trying to wrangle a MySQL query to rank a set of results, I got sucked into a few Stack Overflow threads outlining various approaches.

In order to understand the pros and cons of each technique, I created some test data and reimplemented the solutions, all of which I’ve shared below.

If there are any other methods worth including here, please drop a note in the comments.

We’ll be working with this table:

DROP TABLE IF EXISTS users;

CREATE TABLE users (user_id INT PRIMARY KEY AUTO_INCREMENT, name VARCHAR(20), start_date DATE, team_id INT);
INSERT INTO users (name, start_date, team_id) VALUES ('Matt', '2017-01-01', 1);
INSERT INTO users (name, start_date, team_id) VALUES ('John', '2017-01-02', 2);
INSERT INTO users (name, start_date, team_id) VALUES ('Sara', '2017-01-02', 2);
INSERT INTO users (name, start_date, team_id) VALUES ('Tim', '2017-01-02', 3);
INSERT INTO users (name, start_date, team_id) VALUES ('Bob', '2017-01-03', 3);
INSERT INTO users (name, start_date, team_id) VALUES ('Bill', '2017-01-04', 3);
INSERT INTO users (name, start_date, team_id) VALUES ('Kathy', '2017-01-04', 3);
INSERT INTO users (name, start_date, team_id) VALUES ('Anne', '2017-01-05', 3);

+---------+-------+------------+---------+
| user_id | name  | start_date | team_id |
+---------+-------+------------+---------+
|       1 | Matt  | 2017-01-01 |       1 |
|       2 | John  | 2017-01-02 |       2 |
|       3 | Sara  | 2017-01-02 |       2 |
|       4 | Tim   | 2017-01-02 |       3 |
|       5 | Bob   | 2017-01-03 |       3 |
|       6 | Bill  | 2017-01-04 |       3 |
|       7 | Kathy | 2017-01-04 |       3 |
|       8 | Anne  | 2017-01-05 |       3 |
+---------+-------+------------+---------+

Ranked by start date

SELECT *
FROM users
ORDER by start_date ASC

Ranked by start date with ties broken by user id

SELECT *
FROM users
ORDER by start_date ASC, user_id ASC

First employee by start date with ties broken by user id

SELECT *
FROM users
ORDER by start_date ASC, user_id ASC
LIMIT 1

First employee by start date with ties

SELECT *
FROM users
WHERE start_date = (SELECT MIN(start_date) FROM users);

Second employee by start date with ties broken by user id

SELECT *
FROM users
ORDER by start_date ASC, user_id ASC
LIMIT 1
OFFSET 1

Second employee by start date with ties

SELECT *
FROM users
WHERE start_date = (
    SELECT DISTINCT start_date
    FROM users
    ORDER BY start_date ASC
    LIMIT 1
    OFFSET 1
)

Ranked by start date using variable

SET @rank := 0;
SELECT
    *,
    @rank := @rank + 1 AS rank
FROM users
ORDER BY start_date ASC

+---------+-------+------------+---------+------+
| user_id | name  | start_date | team_id | rank |
+---------+-------+------------+---------+------+
|       1 | Matt  | 2017-01-01 |       1 |    1 |
|       2 | John  | 2017-01-02 |       2 |    2 |
|       3 | Sara  | 2017-01-02 |       2 |    3 |
|       4 | Tim   | 2017-01-02 |       3 |    4 |
|       5 | Bob   | 2017-01-03 |       3 |    5 |
|       6 | Bill  | 2017-01-04 |       3 |    6 |
|       7 | Kathy | 2017-01-04 |       3 |    7 |
|       8 | Anne  | 2017-01-05 |       3 |    8 |
+---------+-------+------------+---------+------+

Ranked by start date using a variable

Based on this Stack Overflow comment:

SELECT
    *,
    @rank := @rank + 1 AS rank
FROM users, (SELECT @rank := 0) r
ORDER BY start_date ASC

+---------+-------+------------+---------+------------+------+
| user_id | name  | start_date | team_id | @rank := 0 | rank |
+---------+-------+------------+---------+------------+------+
|       1 | Matt  | 2017-01-01 |       1 |          0 |    1 |
|       2 | John  | 2017-01-02 |       2 |          0 |    2 |
|       3 | Sara  | 2017-01-02 |       2 |          0 |    3 |
|       4 | Tim   | 2017-01-02 |       3 |          0 |    4 |
|       5 | Bob   | 2017-01-03 |       3 |          0 |    5 |
|       6 | Bill  | 2017-01-04 |       3 |          0 |    6 |
|       7 | Kathy | 2017-01-04 |       3 |          0 |    7 |
|       8 | Anne  | 2017-01-05 |       3 |          0 |    8 |
+---------+-------+------------+---------+------------+------+

First employee by start date using by setting a variable

SET @rank := 0;
SELECT *
FROM (
    SELECT
        *,
        @rank := @rank + 1 AS rank
    FROM users
    ORDER BY start_date ASC
) ranked
WHERE rank = 1

Ranked by start date with ties

Based on this Stack Overflow comment:

SET @prev_start_date = NULL;
SET @rank := 0;
SELECT
    *,
    CASE
        WHEN @prev_start_date = start_date THEN @rank
        -- Note that the assignment here will always be true
        WHEN @prev_start_date := start_date THEN @rank := @rank + 1
    END AS rank
FROM users
ORDER BY start_date ASC

+---------+-------+------------+---------+------+
| user_id | name  | start_date | team_id | rank |
+---------+-------+------------+---------+------+
|       1 | Matt  | 2017-01-01 |       1 |    1 |
|       2 | John  | 2017-01-02 |       2 |    2 |
|       3 | Sara  | 2017-01-02 |       2 |    2 |
|       4 | Tim   | 2017-01-02 |       3 |    2 |
|       5 | Bob   | 2017-01-03 |       3 |    3 |
|       6 | Bill  | 2017-01-04 |       3 |    4 |
|       7 | Kathy | 2017-01-04 |       3 |    4 |
|       8 | Anne  | 2017-01-05 |       3 |    5 |
+---------+-------+------------+---------+------+

Ranked by user id within each team

SELECT
    a.*,
    COUNT(*) AS rank
FROM users a
INNER JOIN users b
ON a.team_id = b.team_id AND a.user_id >= b.user_id
GROUP BY a.team_id, a.user_id

or, based on this Stack Overflow comment:

SELECT
    a.*,
    (
        SELECT COUNT(*)
        FROM users b
        WHERE a.team_id = b.team_id AND a.user_id >= b.user_id
    ) AS ranked
FROM users a

+---------+-------+------------+---------+--------+
| user_id | name  | start_date | team_id | ranked |
+---------+-------+------------+---------+--------+
|       1 | Matt  | 2017-01-01 |       1 |      1 |
|       2 | John  | 2017-01-02 |       2 |      1 |
|       3 | Sara  | 2017-01-02 |       2 |      2 |
|       4 | Tim   | 2017-01-02 |       3 |      1 |
|       5 | Bob   | 2017-01-03 |       3 |      2 |
|       6 | Bill  | 2017-01-04 |       3 |      3 |
|       7 | Kathy | 2017-01-04 |       3 |      4 |
|       8 | Anne  | 2017-01-05 |       3 |      5 |
+---------+-------+------------+---------+--------+

Note that both of these techniques require that there be a column without duplicates that we can rank on within the partition. For example, we can’t use start_date due to the duplicates within team 2 (2017-01-02) and team 3 (2017-01-14):

SELECT
    a.*,
    COUNT(*) AS rank
FROM users a
INNER JOIN users b
ON a.team_id = b.team_id AND a.start_date >= b.start_date
GROUP BY a.team_id, a.user_id

+---------+-------+------------+---------+------+
| user_id | name  | start_date | team_id | rank |
+---------+-------+------------+---------+------+
|       1 | Matt  | 2017-01-01 |       1 |    1 |
|       2 | John  | 2017-01-02 |       2 |    2 |
|       3 | Sara  | 2017-01-02 |       2 |    2 |
|       4 | Tim   | 2017-01-02 |       3 |    1 |
|       5 | Bob   | 2017-01-03 |       3 |    2 |
|       6 | Bill  | 2017-01-04 |       3 |    4 |
|       7 | Kathy | 2017-01-04 |       3 |    4 |
|       8 | Anne  | 2017-01-05 |       3 |    5 |
+---------+-------+------------+---------+------+

Return the last person to join within each team based on user id

Based on this Stack Overflow comment:

SELECT a.*
FROM users a
LEFT JOIN users b
ON a.team_id = b.team_id AND a.user_id < b.user_id
WHERE b.team_id IS NULL

SELECT a.*
FROM users a
WHERE user_id IN (
    SELECT MAX(user_id)
    FROM users
    GROUP BY team_id
)

+---------+------+------------+---------+
| user_id | name | start_date | team_id |
+---------+------+------------+---------+
|       1 | Matt | 2017-01-01 |       1 |
|       3 | Sara | 2017-01-02 |       2 |
|       8 | Anne | 2017-01-05 |       3 |
+---------+------+------------+---------+

Return the last people to join within each team based on start date

SELECT a.*
FROM users a
LEFT JOIN users b
ON a.team_id = b.team_id AND a.start_date < b.start_date
WHERE b.team_id IS NULL

or, based on this groupwise max post:

SELECT a.*
FROM users a
INNER JOIN (
    SELECT team_id, MAX(start_date) AS max_start_date
    FROM users b
    GROUP BY team_id
) max_start_dates
ON a.team_id = max_start_dates.team_id AND a.start_date = max_start_dates.max_start_date

+---------+------+------------+---------+
| user_id | name | start_date | team_id |
+---------+------+------------+---------+
|       1 | Matt | 2017-01-01 |       1 |
|       2 | John | 2017-01-02 |       2 |
|       3 | Sara | 2017-01-02 |       2 |
|       8 | Anne | 2017-01-05 |       3 |
+---------+------+------------+---------+

Ranked with gaps

Based on this Stack Overflow comment:

SELECT
    user_id,
    name,
    start_date,
    team_id,
    rank
FROM (
    SELECT
        *,
        IF(start_date = @_last_start_date, @cur_rank := @cur_rank, @cur_rank := @_sequence) AS rank,
        @_sequence := @_sequence + 1,
        @_last_start_date := start_date
    FROM users, (SELECT @cur_rank := 1, @_sequence := 1, @_last_start_date := NULL) r
    ORDER BY start_date
) ranked

Notice that after the three tied for second earliest start date, the next one jumps to 5 (not 3):

+---------+-------+------------+---------+------+
| user_id | name  | start_date | team_id | rank |
+---------+-------+------------+---------+------+
|       1 | Matt  | 2017-01-01 |       1 |    1 |
|       2 | John  | 2017-01-02 |       2 |    2 |
|       3 | Sara  | 2017-01-02 |       2 |    2 |
|       4 | Tim   | 2017-01-02 |       3 |    2 |
|       5 | Bob   | 2017-01-03 |       3 |    5 |
|       6 | Bill  | 2017-01-04 |       3 |    6 |
|       7 | Kathy | 2017-01-04 |       3 |    6 |
|       8 | Anne  | 2017-01-05 |       3 |    8 |
+---------+-------+------------+---------+------+

Technically this is known as the rank whereas the other examples, which didn’t include gaps, is the dense rank.

Happy ranking!

In order to level up my data visualization skills, I’m going to be attempting to implement well-designed charts that others have created on my own using R and eventually other tools. I’ll be open sourcing the code and sharing any lessons learned along the way.

For the first chart, I grabbed a seemingly simple chart from Scott Berinato’s excellent book, Good Charts: The HBR Guide to Making Smarter, More Persuasive Data Visualizations.

Here’s the chart that I set out to reproduce in R (he created it in Illustrator):

Here’s my attempt using R and ggplot2 whose code you can find in my new data viz repo on GitHub:

There were several tricky aspects of the original chart that presented ~~challenges~~ learning opportunities while implementing it in R:

In order to get the chart title, y-axis title, and source aligned with the left side of the chart, I wound up having to use R’s Graphical User Objects aka “grobs”; hat-tip to this SO thread for the idea. It was possible to get pretty close by adjusting the hjust value for the chart title and subtitle, but the grob approach seemed less hacky.
Speaking of hacky, I wasn’t able to figure out an elegant way to center the x-axis values between the ticks. I resorted to tinkering with the hjust value, but that just shifts them all to the right by a fixed amount. The problem is that some text like “2010” is wider than other text like “2011” which makes it impossible to get them all centered by shifting them by the same fixed amount. If anyone has a solution to this, I’d love to learn a better way to do it.
Removing the tick for the bottom ($0) y-axis value was “fun”: I had to make it transparent because I couldn’t figure out how to remove it completely. Again, Stack Overflow is awesome.
From a coding best practices perspective, there were several appearance-related values that had to be used in multiple locations in the code. To avoid duplication, I stored them in variables at the top of the script. It seems strange though having some appearance values stored in variables but others directly in the ggplot2 code. Storing all of the appearance values in variables – even the ones only used once – seems like bad practice too though.
I couldn’t figure out how to get the ends of the x-axis line to stop at the ticks so they wind up extending beyond them slightly in my version.
I’m not sure what fonts Scott used in Illustrator to create this chart, but even if I was 100% sure I don’t want to shell out money to buy custom fonts for this little project, so I used similar ones that were already available on my Macbook (PT Sans Narrow for the title, Proxima Nova for everything else).
Towards the end there were diminishing returns trying to get everything pixel perfect so I settled on a 90% good enough solution.

As I recreate more of charts there will undoubedly be “a-ha!” moments that cause me to go back and clean up the code for my old charts. I’ll try to share as many of those lessons learned here as well.