Background

I'm working on a website for a movie theather chain currently located in four different cities (might expand in the future). They use the same, single-database website for all cities, which means I have to have a column in certain tables which holds the ID of the city that each row belongs to.

Right now I have three different tables:

• 
  Cinemas - Contains each city's cinema (ID and name).


• 
  Movies - Contains all movies that has been/will be shown at the cinema.


• 
  Showtimes - Contains all showtimes for all movies in all cities.

The structure of the Showtimes table is the following:

Column Name   | Column Type  | Description

--------------+--------------+---------------

ID            | BIGINT       | (Primary) Unique ID for each showtime (perhaps unnecessary?)

CinemaID      | TINYINT      | Foreign key bound to Cinemas.ID

MovieID       | BIGINT       | Foreign key bound to Movies.ID

Showtime      | DATETIME     | At what date and time the movie will show 



(will contain multiple rows for each movie, i.e. one row for each showtime)

How this table will be used

A user of the website must be able to:

• 
  

  View all current/upcoming movies and showtimes (sorted by date) in the selected city.

  
  
  

  Example query (backend):

  
  
  

  SELECT MovieID, Showtime FROM Showtimes WHERE CinemaID = ? ORDER BY Showtime
  
  

  
  


• 
  

  Select a single movie and view all showtimes for that specific title only (in the selected city).

  
  
  

  Example query:

  
  
  

  SELECT Showtime FROM Showtimes WHERE CinemaID = ? AND MovieID = ? ORDER BY Showtime
  
  

  
  


• 
  

  Select a single day and view all movies and showtimes for that day only (in the selected city).

  
  
  

  Example query:

  
  
  

  SELECT MovieID, Showtime FROM Showtimes WHERE CinemaID = ? AND (Showtime BETWEEN [date 12:00 AM] AND [date 12:00 PM])
  
  

  
  

So naturally I decided that I needed to create indexes for the columns.

Problem

What I'm having trouble with is deciding/determining how to index the columns properly. One index for each column seems quite expensive^[1][2] so I started looking into composite indexes, which seems to be the right choice but also led to even more confusion.

From my understanding (based on what I've read) you should add the columns to the index by order of selectivity, making the most selective (I'm guessing that means the most unique/with the most cardinality?) column the first in the composite index^[3] (in my case that would be the Showtime column). The only problem with that is that the index can only be used by the database if the first column is included in the search query^[4][5], which it currently isn't in either of my queries.

Question

What kind of index(es) should I apply to my columns in order to cover all usage scenarios? (the last scenario may be omitted, but the first two are required)

Should I use a composite index on all columns, for some columns, or do I need a separate index for each column?

This table is updated at most a few times per week to add new showtimes.

Footnotes

¹MySQL indexes - what are the best practices?

²Indexing every column in a table

³How important is the order of columns in indexes? (question)

⁴How important is the order of columns in indexes? (#2 top-voted answer)

⁵When should I use a composite index?

Composite Primary Key

I would define the primary key as a composite key of (CinemaID, MovieID, Showtime).

These 3 columns uniquely identify each row, and so having a separate ID column is not necessary.

Composite (Secondary) Index

With this PK, the only additional index you will need for your queries is (CinemaID, Showtime).

Why these indexes?

A good way to think about indexes are used is to think of them as the order of columns in a spreadsheet.

Imagine a spreadsheed with (CinemaID, MovieID, Showtime) that is sorted by each column successively.

All your queries have CinemaID present, which means you can quickly find the "section" of the spreadsheet for that CinemaID. Then, for your query that searches by MovieID, you can easily find the "subsection" in the 2nd column, where MovieID matches the searched for value.

As the 3rd column of Showtime is also sorted, you can imagine how quickly and easily it is to find all those show times for that movie, in that cinema. The DBMS does things in a similar way and can retrieve those results extremely quickly.

As for your other queries, they all start with CinemaID and then use Showtime in some manner. They also need the MovieID in their results.

So, the (CinemaID, Showtime) index has you covered there. Again, the CinemaID easily finds the "section" of the spreadsheet (in the analogy), and all possible showtimes (and there are going to be duplicates, assuming there is more than one screen) will be listed in order and are easily searched and/or sorted by those values.

Even better, since your primary key includes MovieID, that column is included for all secondary indexes after the defined columns (at least for MySQL InnoDB - others engines too, but not necessarily all.)

Think of that is a 3rd column in the "spreadsheet" of our secondary index. The reason the column exists is to have all parts of the primary key available to do a lookup to the main table (aka the clustered index, in InnoDB), if necessary. In this simple case, no lookup is needed, so it is even more efficient as it doesn't require that double lookup.

Using just this primary key and a single secondary index, you should obtain excellent performance on any of the queries you listed.

Afterthought

My assumption that this uniquely identifies each row may NOT be correct if you have a movie showing on multiple screens at the same time. If you want to be able to separately identify those screens, then my solution is not best (I can provide another solution for that situation, please just let me know.)

WHERE CinemaID = ? ORDER BY Showtime  -- and

WHERE CinemaID = ? AND (Showtime BETWEEN [date 12:00 AM] AND [date 12:00 PM])  -- need:

INDEX(CinemaID, Showtime)



WHERE CinemaID = ? AND MovieID = ? ORDER BY Showtime  -- needs:

INDEX(CinemaID, MovieId, Showtime)  -- or

INDEX(MovieId, CinemaID, Showtime)

Assuming that the the triple (MovieId, CinemaID, Showtime) is unique, I suggest getting rid of id and having

PRIMARY KEY(CinemaID, MovieId, Showtime)

INDEX(CinemaID, Showtime)

Are there cases where the main part of the WHERE is MovieID=...?

Cinemas - A list of each city and its cinema (ID and name):

SELECT Cinema, CinemaID FROM Cimemas;  -- (no index needed)

Movies - A list of movies that has been/will be shown at the cinema.

SELECT DISTINCT MovieID FROM ShowTimes WHERE CinemaID=...

INDEX(CinemaID, MovieID)  -- already handled by my proposed PK

Showtimes - A list of all showtimes for all movies in all cities. -- This is a huge output; rethink the requirement for it. That is, think about what the Client will do with it.

Most of these indexes could be deduced from studying indexing cookbook and composite indexing

Note how I did not suggest indexes (single-column vs composite, plus specifics) until I saw the SELECTs.

"columns to the index by order of selectivity, making the most selective (I'm guessing that means the most unique/with the most cardinality?)" -- No. Selectivity is not the key for designing a composite index. Start with all the = columns, in any order. (My Cookbook dwells on this topic.)

"index can only be used by the table if the first column is included in the search query" -- Mostly true. Notice how I recommended 2 indexes (remember: the PK is an index). There are cases where an index can be used for GROUP BY or ORDER BY, ignoring the WHERE; but those are rare.

"first in the composite index ... Showtime" -- It is usually counterproductive to put a DATETIME first in a composite index. In particular, your 3rd query can use both columns of INDEX(CinemaID, Showtime), but not both columns of INDEX(Showtime, CinemaID). It is easy to see this: Think of writing out two lists of cinemas and showtimes. Have one list sorted first on cinema (a la INDEX(cinema, time); have the other sorted on time. Think about which one would have all the rows for a particular cinema over a time range clumped ("clustered") together.

If there is also a ScreenID due to two screens showing the same movie at the same time, tack it onto the PK. However, all the indexes will need rethinking.

(Sorry, Willem, I wrote my Answer before reading yours -- We say almost the same stuff.)

You are optimizing for size*, on an extremely small table. Since this basically a bridge table, it doesn’t really need a primary key, and even if it did, it wouldn’t need to be a bigint, the movies id doesn’t need to be a bigint either. If you think you need more than a couple of billion movies, then used unsigned but I think it’s unlikely that you’re going to run into that limitation for number of movies anytime soon.

Have an index on each of your primary terms, movie, cinema, showtime. See what your performance is, and if it’s unacceptable add additional index combinations.

AFTER you give it a try with no indexes at all (except possibly a pk on an auto_increment column used as a surrogate key) and a reasonable amount of data, say 200 cinemas X (5 new movies a week x 4 show times daily x 7 x (520 weeks aka 10 years)) = 14,560,000 records and see how it performs.

Frankly at this size, a full table scan shouldn’t be tooo bad.

Size wise a totally covering index is 13 bytes, rounds up to 16, so 16 x 14,560,000 / 1024 is 227,500k aka 227mb aka .277. Times 4 for the table and each of the possible covering index and we are up to almost 1.25 gig for the table and indexes. Note that any combination indexes would be the same size, just have a different weight to the data...

Note that the way to really speed this up, is to put your historical data in another table or use partitioning. In which case if you have the data for the upcoming month, it’s 5 x 4 x 7 x 30 x 200 = 120,000 records, and scanning 120k records should not be a problem. At that point you might just keep it all in memory and use a local dictionary to look things up.

Also note that I am assuming that you expand like crazy, as you currently have only one screen, so probably only 4-6 records a day per cinema, and a low number of cinemas, so a more realistic daily count is probably between 1k and 10k records per month. Totally possible you’d get decent enough performance out of this if you stored the values in an excel spreadsheet which was manually updated daily.

PS for index selectivity, you don’t want the most precise, you want the most coverage. You aren’t looking for a single record, you are looking for a group, which means you want the index to be on the column that returns your group. In particular, this means that if you want to look for movies by day, you should split the date and time into two columns so you can put an index on the date (and possibly one on the time).

*The reason I said you were optimizing for space is because you seem to be worried about the cost of the index. The cost of an index comes in two forms, a slower insert/delete/update and disk space. For a full covering indexes that basically means each index cost just as much as the table does. The time to add a single row to a table with 16 bytes (or 24) is small and non-accumulating (ie basically constant regardless of the number of rows). While disk space accumulates slowly, it does grow.