The data model

Andrew Gallant edited this page Jul 23, 2014 · 19 revisions

One of the goals of nfldb is for it to be simple to use. And for it to be simple to use, the data model should also be simple.

There are only 8 tables in the database. Here is a brief description of each:

  • meta stores information about the database or about the state of the world. For example, it keeps track of the version of the database and the current week of the current NFL season.
  • team stores a row for each team in the league. There is also a row that corresponds to an unknown team called UNK. This is used for players that are not on any current roster.
  • player stores ephemeral data about players. Namely, it is the most current information about each player known by nfldb. The data is nearly a total copy of the data in nflgame's JSON player database.
  • game stores a row for each NFL game in the preseason, regular season and postseason dating back to 2009. This includes games that are scheduled in the future but have not been played.
  • drive stores a row for each drive in a single game.
  • play stores a row for each play in a single drive.
  • agg_play stores a row for each play aggregating statistics from the corresponding rows in the play_player table.
  • play_player stores a row for each player statistic in a single play.

You can get an overview of the entire database and the relationships between each table with the Entity-Relationship (ER) diagrams section of this page.

The rest of this page describes the data stored in the database with SQL examples. An explicit effort was made not to talk about the nfldb Python module.

What kind of player meta data is stored?

The data in the player table corresponds to information scraped off of roster and player profile pages on (In fact, this is the only data in nfldb that is scraped.) pages are used so that players can be matched with their statistical data via unique identifiers, rather than having to rely on a fuzzy name matching algorithm. (If you're interested in how the scraping is done, please see the nflgame-update-players script in the nflgame repository.)

This data includes players who are no longer playing. In this case, their team is UNK. This leads to a nice property of the data in the player table: any player with a team not equal to UNK is currently on that team's roster. Therefore, the roster of a team (sorted by player status, position and name) as known by nfldb can easily be accessed with the following query:

SELECT full_name, position, status
FROM player
WHERE team = 'NE'
ORDER BY status ASC, position ASC, full_name ASC

Similarly, you could get every player currently on a roster by using team != 'UNK'.

Whether the data in the player table is current or not depends on how quickly updates their data and whether you're updating your database frequently. In my experience,'s data can be slow to update during the offseason, but is relatively quick during the season.

Finally, it is important to note that most of the columns in the player table can be NULL. This means that not all data is available for all players. (We are at the mercy of the consistency of's roster and player profile pages.) In my experience, the data is usually very complete for active players.

What is the play_player table?

The play_player table is arguably the most important table in the entire database. Namely, it is a single point of truth for player statistics. Each row in this table corresponds to statistics recorded by a single player in a single play. (Aggregated player statistics for each play are stored in the agg_play table. agg_play is a materialized view.)

Let's look at a fairly complex play that occurred in the Eagles/Redskins game during the first week of the 2013 regular season in the 4th quarter with 13:55 remaining on the clock:

M.Vick pass short right to J.Avant to PHI 33 for 6 yards (J.Wilson).
FUMBLES (J.Wilson), RECOVERED by WAS-P.Riley at PHI 35.
P.Riley to PHI 29 for 6 yards (J.Avant).

This particular play has a gsis_id of 2013090900, a drive_id of 21 and a play_id of 3717. We can then use that information to see the statistics recorded by each player in that play. (Note that I've restricted the SELECT fields in the query below to make the output readable here. You may want to try SELECT * ....)

    full_name, passing_yds, receiving_rec, receiving_yds,
    fumbles_forced, defense_ffum, defense_frec_yds
FROM play_player
LEFT JOIN player ON player.player_id = play_player.player_id
WHERE (gsis_id, drive_id, play_id) = ('2013090900', 21, 3717)

And the output of that query is:

  full_name   | passing_yds | receiving_rec | receiving_yds | fumbles_forced | defense_ffum | defense_frec_yds 
 Michael Vick |           6 |             0 |             0 |              0 |            0 |                0
 Jason Avant  |           0 |             1 |             6 |              1 |            0 |                0
 Josh Wilson  |           0 |             0 |             0 |              0 |            1 |                0
 Perry Riley  |           0 |             0 |             0 |              0 |            0 |                6

The statistics record that Michael Vick threw a pass for 6 yards, Jason Avant caught a pass for 6 yards and had the ball stripped by Josh Wilson, which was recovered by Perry Riley and returned for 6 yards.

Aggregate statistics

To keep things simple and to avoid duplication of data, nfldb does not store aggregate statistics beyond the play level. Instead, they must be computed on the fly. Thankfully, there is a fast way to do it with PostgreSQL's aggregate functions. For example, we could find the top quarterbacks in the 2012 regular season with more than 4500 passing yards:

SELECT player.full_name, SUM(play_player.passing_yds) AS passing_yds
FROM play_player
LEFT JOIN player ON player.player_id = play_player.player_id
LEFT JOIN game ON game.gsis_id = play_player.gsis_id
WHERE game.season_year = 2012 AND game.season_type = 'Regular'
GROUP BY player.full_name
HAVING SUM(play_player.passing_yds) >= 4500
ORDER BY passing_yds DESC

And the output is:

    full_name     | passing_yds 
 Drew Brees       |        5177
 Matthew Stafford |        4965
 Tony Romo        |        4903
 Tom Brady        |        4799
 Matt Ryan        |        4719
 Peyton Manning   |        4667

When writing aggregate queries, the WHERE clause specifies what to aggregate while the HAVING clause specifies which aggregate results to return. Namely, WHERE is used before aggregation while HAVING is used after aggregation.

The agg_play table

It is a materialized view. A materialized view is like a normal view (which is just a saved SELECT query), except it actually stores the data.

Why does nfldb aggregate player statistics for each play? For the most part, it is a decision guided by the performance of searching statistics. If we didn't aggregate any data, then filtering plays based on statistics---like passing yards---requires joining with the play_player table and summing all the joining rows with SUM. On large data sets, this is a very expensive operation. If we aggregate statistics, no joining or summing is necessary when filtering. The costs are pretty meager by comparison: the database is a little bigger and it takes a little longer to insert data.

Since agg_play is a materialized view, it requires no maintenance. It is automatically updated whenver new data is added, modified or deleted. It just works.

Will other materialized views with aggregated data be added? I'm not sure. The play table really benefits because filtering statistics by plays is so common.

Finally, the agg_play table is completely hidden from the public interface. Even though play data is actually stored across two tables, users of nfldb never need to know this.

ER diagrams

Entity-Relationship (ER) diagrams are used to graphically represent the schema of a database. They show each entity, its attributes and the relationships between each entity. In the ER diagrams for nfldb, entities correspond to tables and attributes correspond to columns in a table.

An example of a relationship would be one-to-many between games and drives. Namely, for each game, there can be zero or more drives associated with that game and for each drive, there must be exactly one game associated with that drive.

Note that the ER diagrams do not contain derived fields. You can see documentation for each statistical category (including derived fields) on the statistical categories page.

There are two ER diagrams. The first is a condensed version that omits many of the statistical categories for plays and players. (Click on the image to get a full PDF of the ER diagram.)

Shortened ER diagram for nfldb

The second is a full ER diagram, with all of the statistical categories:

Full ER diagram for nfldb

If you're curious, the above ER diagrams are automatically generated with the nfldb-write-erd script using a utility I wrote called erd.