Best SQL indexes for join table

Question

With performance improvements in mind, I was wondering if and which indexes are helpful on a join table (specifically used in a Rails 3 has_and_belongs_to_many context).

Model and Table Setup

My models are Foo and Bar and per rails convention, I have a join table called bars_foos. There is no primary key or timestamps making the old fields in this table bar_id:integer and foo_id:integer. I'm interested in knowing which of the following indexes is best and is without duplication:

1. A compound index: add_index :bars_foos, [:bar_id, :foo_id]
2. • Two indexes
   • A. add_index :bars_foos, :bar_id
   • B. add_index :bars_foos, :foo_id
3. A combination of both 1 and 2-B

Basically, I'm not sure if the compound index is enough assuming it is helpful to begin with. I believe that a compound index can be used as a single index for the first item which is why I am pretty sure that using all three lines would certainly result in unnecessary duplication.

Likely Usage

The most common usage will be given an instance of model Foo, I will be asking for its associated bars using the RoR syntax of foo.bars and vice versa with bar.foos for an instance of the model Bar.

These will generate queries of the type SELECT * FROM bars_foos WHERE foo_id = ? and SELECT * FROM bars_foos WHERE bar_id = ? respectively and then using those resultant IDs to SELECT * FROM bars WHERE ID in (?) and SELECT * FROM foos WHERE ID in (?).

Please correct me in the comments if I am incorrect, but I do not believe that, in the context of the Rails application, it is ever going to try to do a query where it specifies both IDs like SELECT * FROM bars_foos where bar_id = ? AND foo_id = ?.

Databases

In the event there are database specific optimization techniques, I will most likely be using PostgreSQL. However, others using this code may want to use it in MySQL or SQLite depending on their Rails configuration so all answers are appreciated.

Answer 1

The Answer

The oft repeated answer, which tends to always be the case more often than not is, "it depends." More specifically, it depends on what your data is and how it will be used.

tl;dr Explanation

The short tl;dr answer for my specific case (and to cover all future bases) is choice #2 which is what I suspected. However, choice #3 would work just fine as, depending on my usage of the data, the extra time and space used creating the compound index could reduce future query lookups.

The Full Explanation

The reason for this is that databases try to be smart and try to do things as fast as possible regardless of programmer input. The most basic item to consider when adding an index is will this object be looked up by this key. If yes, an index can potentially help speed that up. However, whether this index is even used all comes down to selectivity and the cardinality of the field.

Since foreign keys are typically the IDs of another AR class, cardinality usually will be high. But again, this depends on your data. In my example if there are many Foos but few Bars, many of the entries in my join table will have simliar bar_ids. With bar_ids having a low cardinality, an index on bar_id may never be used and may be getting in the way by having the database devote time and resources* to adding to this index every time a new bars_foos entry is created. The same goes with many Bars and few Foos and few of both.

The general lesson is that when considering an index on a table, decide if the entries will be both looked up by this field and if this field has a high cardinality. That is, does this field have many distinct values? In the case of most join tables "it depends" and we must think more carefully about what the data represents and the relationships themselves. In my case, I will have both many Foos and Bars and will be looking up Foos by their associated bars and vice versa.

Another good answer I got at the office was, "why are you worrying about your indexes? Build your app!"

Footnotes

* In a similar question on indexes on STI it was pointed out that the cost of an index is very low so when in doubt, just add it.