= Advanced Associations

Sequel::Model's association support is powerful and flexible, but it can be difficult for
new users to understand what the support enables.  This guide shows off some of the more
advanced Sequel::Model association features.

You should probably review the {Model Associations Basics and Options guide}[rdoc-ref:doc/association_basics.rdoc]
before reviewing this guide.

== Sequel::Model Eager Loading

Sequel::Model offers two different ways to perform eager loading, +eager+ and
+eager_graph+. +eager+ uses an SQL query per association, +eager_graph+ uses a single
SQL query containing JOINs.

Assuming the following associations:

  Artist.one_to_many :albums
  Album.one_to_many :tracks
  Tracks.many_to_one :lyric

Let's say you wanted to load all artists and eagerly load the related albums, tracks, and lyrics.

  Artist.eager(albums: {tracks: :lyric})
  # 4 Queries:
  # SELECT * FROM artists;
  # SELECT * FROM albums WHERE (artist_id IN (...));
  # SELECT * FROM tracks WHERE (album_id IN (...));
  # SELECT * FROM lyrics WHERE (id IN (...));

  Artist.eager_graph(albums: {tracks: :lyric})
  # 1 Query:
  # SELECT artists.id, artists.name, ...
  #        albums.id AS albums_id, albums.name AS albums_name, ...
  #        tracks.id AS tracks_id, tracks.name AS tracks_name, ...
  #        lyric.id AS lyric_id, ...
  # FROM artists
  # LEFT OUTER JOIN albums ON (albums.artist_id = artists.id)
  # LEFT OUTER JOIN tracks ON (tracks.album_id = albums.id)
  # LEFT OUTER JOIN lyrics AS lyric ON (lyric.id = tracks.lyric_id);

In general, the recommendation is to use +eager+ unless you have a reason to use +eager_graph+.
+eager_graph+ is needed when you want to reference columns in an associated table.  For example,
if you want to order the loading of returned artists based on the names of the albums, you cannot
do:

  Artist.eager(albums: {tracks: :lyric}).order{albums[:name]}

because the initial query Sequel will use would be:

  # SELECT * FROM artists ORDER BY albums.name;

and +albums+ is not a valid qualifier in such a query.  In this situation, you must use +eager_graph+:

  Artist.eager_graph(albums: {tracks: :lyric}).order{albums[:name]}

Whether +eager+ or +eager_graph+ performs better is association and database dependent. If
you are concerned about performance, you should try benchmarking both cases with appropriate
data to see which performs better.

=== Mixing eager and eager_graph

Sequel offers the ability to mix +eager+ and +eager_graph+ when loading results.  This can
be done at the main level by calling both +eager+ and +eager_graph+ on the same dataset:

  Album.eager(:artist).eager_graph(:tracks)
  # 2 Queries:
  # SELECT albums.id, albums.name, ...
  #        artist.id AS artist_id, artist.name AS artist_name, ...
  # FROM albums
  # LEFT OUTER JOIN artists AS artist ON (artist.id = albums.artist_id);
  # SELECT * FROM artists WHERE (id IN (...));

You can also use +eager+ to load initial associations, and +eager_graph+ to load
remaining associations, by using +eager_graph+ in an eager load callback:

  Artist.eager(albums: {tracks: proc{|ds| ds.eager_graph(:lyric)}})
  # 3 Queries:
  # SELECT * FROM artists;
  # SELECT * FROM albums WHERE (artist_id IN (...));
  # SELECT tracks.id, tracks.name, ...
  #        lyric.id AS lyric_id, ...
  # FROM tracks
  # LEFT OUTER JOIN lyrics AS lyric ON (lyric.id = tracks.lyric_id)
  # WHERE (tracks.album_id IN (...));

Using the +eager_graph_eager+ plugin, you can use +eager_graph+ to load the
initial associations, and +eager+ to load the remaining associations.  When
you call +eager_graph_eager+, you must specify the dependency chain at
which to start the eager loading via +eager+:

  Artist.plugin :eager_graph_eager
  Artist.eager_graph(albums: :tracks).eager_graph_eager([:albums, :tracks], :lyric)
  # 2 Queries:
  # SELECT artists.id, artists.name, ...
  #        albums.id AS albums_id, albums.name AS albums_name, ...
  #        tracks.id AS tracks_id, tracks.name AS tracks_name, ...
  # FROM artists
  # LEFT OUTER JOIN albums ON (albums.artist_id = artists.id)
  # LEFT OUTER JOIN tracks ON (tracks.album_id= albums.id);
  # SELECT * FROM lyrics WHERE (id IN (...));

These two approaches can also be nested, with +eager+ -> +eager_graph+ -> +eager+:

  Album.plugin :eager_graph_eager
  Artist.eager(albums: proc{|ds| ds.eager_graph(:tracks).eager_graph_eager([:tracks], :lyric)})
  # 3 Queries:
  # SELECT * FROM artists;
  # SELECT albums.id, albums.name, ...
  #        tracks.id AS tracks_id, tracks.name AS tracks_name, ...
  # FROM albums 
  # LEFT OUTER JOIN tracks ON (tracks.album_id = albums.id)
  # WHERE (albums.artist_id IN (...));
  # SELECT * FROM lyrics WHERE (id IN (...));

Or with 2 separate +eager_graph+ queries:

  Artist.eager_graph(:albums).eager_graph_eager([:albums], tracks: proc{|ds| ds.eager_graph(:lyric)})
  # 2 Queries:
  # SELECT artists.id, artists.name, ...
  #        albums.id AS albums_id, albums.name AS albums_name, ...
  # FROM artists
  # LEFT OUTER JOIN albums ON (albums.artist_id = artists.id);
  # SELECT tracks.id, tracks.name, ...
  #        lyric.id AS lyric_id, ...
  # FROM tracks
  # LEFT OUTER JOIN lyrics AS lyric ON (lyric.id = tracks.lyric_id)
  # WHERE (tracks.album_id IN (...));

== Sequel::Model Association Loading Options

There are a bunch of advanced association options that are available to
handle more complex cases.  First we'll go over some of the simpler ones:

All associations take a block that can be used to further filter/modify the
default dataset:

  Artist.one_to_many :gold_albums, class: :Album do |ds|
    ds.where{copies_sold > 500000}
  end

There's also an :eager_block option if you want to use a different block when
eager loading via <tt>Dataset#eager</tt>.

There are many options for changing how the association is eagerly
loaded via <tt>Dataset#eager_graph</tt>:

:graph_join_type :: The type of join to do (<tt>:inner</tt>, <tt>:left</tt>, <tt>:right</tt>)
:graph_conditions :: Additional conditions to put on join (needs to be a
                     hash or array of all two pairs).  Automatically assumes unqualified symbols
                     or first element of the pair to be columns of the associated model, and
                     unqualified symbols of the second element of the pair to be columns of the
                     current model.
:graph_block :: A block passed to +join_table+, allowing you to specify
                conditions other than equality, or to use OR, or set up any arbitrary
                condition.  The block is passed the associated table alias, current table 
                alias, and an array of previous joins clause objects.
:graph_only_conditions :: Use these conditions instead of the standard
                          association conditions. This is necessary when you don't want to have an
                          equal condition between the foreign key and primary key of the tables.
                          You can also use this to have a JOIN USING (array of symbols), or a NATURAL
                          or CROSS JOIN (nil, with the appropriate <tt>:graph_join_type</tt>).
  
These can be used like this:

  # Makes Artist.eager_graph(:required_albums).all not return artists that
  # don't have any albums
  Artist.one_to_many :required_albums, class: :Album, graph_join_type: :inner
  
  # Makes sure all returned albums have the active flag set
  Artist.one_to_many :active_albums, class: :Album, graph_conditions: {active: true}
    
  # Only returns albums that have sold more than 500,000 copies
  Artist.one_to_many :gold_albums, class: :Album,
    graph_block: proc{|j,lj,js| Sequel[j][:copies_sold] > 500000}
    
  # Handles the case where the tables are associated by a case insensitive name string
  Artist.one_to_many :albums, key: :artist_name,
    graph_only_conditions: nil,
    graph_block: proc{|j,lj,js| {Sequel.function(:lower, Sequel[j][:artist_name])=>Sequel.function(:lower, Sequel[lj][:name])}}
    
  # Handles the case where both key columns have the name artist_name, and you want to use
  # a JOIN USING
  Artist.one_to_many :albums, key: :artist_name, graph_only_conditions: [:artist_name]

One advantage of using +eager_graph+ is that you can easily filter/order
on columns in an associated table on a per-query basis, using regular
Sequel dataset methods.  For example, if you only want to retrieve artists
who have albums that start with A, and eager load just those albums,
ordered by the albums name, you can do:

  albums = Artist.
    eager_graph(:albums).
    where{Sequel.like(albums[:name], 'A%')}.
    order{albums[:name]}.
    all

For lazy loading (e.g. Model[1].association), the <tt>:dataset</tt> option can be used
to specify an arbitrary dataset (one that uses different keys, multiple keys,
joins to other tables, etc.).

== Custom Eager Loaders

For eager loading via +eager+, the <tt>:eager_loader</tt> option can be used to specify
how to eagerly load a complex association.  This is an extremely powerful
option.  Though it can often be verbose (compared to other things in Sequel),
it allows you complete control over how to eagerly load associations for a
group of objects.

:eager_loader should be a proc that takes a single hash argument, which will
have at least the following keys:

:id_map :: A mapping of key values to arrays of current model instances,
           usage described below
:rows :: An array of model objects
:associations :: A hash of dependent associations to eagerly load
:self :: The dataset that is doing the eager loading
:eager_block :: A dynamic callback for this eager load.

Since you are given all of the records, you can do things like filter on
associations that are specified by multiple keys, or do multiple
queries depending on the content of the records (which would be
necessary for polymorphic associations).  Inside the <tt>:eager_loader</tt>
proc, you should get the related objects and populate the
associations cache for all objects in the array of records.  The hash
of dependent associations is available for you to cascade the eager
loading down multiple levels, but it is up to you to use it.

The id_map is a performance enhancement that is used by the default
association loaders and is also available to you.  It is a hash with keys
foreign/primary key values, and values being arrays of current model
objects having the foreign/primary key value associated with the key. 
This may be hard to visualize, so I'll give an example.  Let's say you
have the following associations

  Album.many_to_one :artist
  Album.one_to_many :tracks
    
and the following three albums in the database:

  album1 = Album.create(artist_id: 3) # id: 1
  album2 = Album.create(artist_id: 3) # id: 2
  album3 = Album.create(artist_id: 2) # id: 3

If you try to eager load this dataset:

  Album.eager(:artist, :tracks).all

Then the id_map provided to the artist :eager_loader proc would be:

  {3=>[album1, album2], 2=>[album3]}

The artist id_map contains a mapping of artist_id values to arrays of
album objects.  Since both album1 and album2 have the same artist_id,
the are both in the array related to that key.  album3 has a different
artist_id, so it is in a different array. Eager loading of artists is
done by looking for any artist having one of the keys in the hash:

  artists = Artist.where(id: id_map.keys).all

When the artists are retrieved, you can iterate over them, find entries
with matching keys, and manually associate them to the albums:

  artists.each do |artist|
    # Find related albums using the artist_id_map
    if albums = id_map[artist.id]
      # Iterate over the albums
      albums.each do |album|
        # Manually set the artist association for each album
        album.associations[:artist] = artist
      end
    end
  end

The id_map provided to the tracks :eager_loader proc would be:

  {1=>[album1], 2=>[album2], 3=>[album3]}

Now the id_map contains a mapping of id values to arrays of album objects (in this
case each array only has a single object, because id is the primary key).  So when
looking for tracks to eagerly load, you only need to look for ones that have an
album_id with one of the keys in the hash:

  tracks = Track.where(album_id: id_map.keys).all

When the tracks are retrieved, you can iterate over them, find entries with matching
keys, and manually associate them to the albums:

  tracks.each do |track|
    if albums = id_map[track.album_id]
      albums.each do |album|
        album.associations[:tracks] << track
      end
    end
  end

=== Two basic example eager loaders

Putting the code in the above examples together, you almost have enough for a basic
working eager loader.  The main important thing that is missing is you need to set
initial values for the eagerly loaded associations.  For the artist association, you
need to initial the values to nil:

  # rows here is the :rows entry in the hash passed to the eager loader
  rows.each{|album| album.associations[:artist] = nil}
  
For the tracks association, you set the initial value to an empty array:

  rows.each{|album| album.associations[:track] = []}

These are done so that if an album currently being loaded doesn't have an associated
artist or any associated tracks, the lack of them will be cached, so calling the
artist or tracks method on the album will not do another database lookup.

So putting everything together, the artist eager loader looks like:

  Album.many_to_one :artist, eager_loader: (proc do |eo_opts|
    eo_opts[:rows].each{|album| album.associations[:artist] = nil}
    id_map = eo_opts[:id_map]
    Artist.where(id: id_map.keys).all do |artist|
      if albums = id_map[artist.id]
        albums.each do |album|
          album.associations[:artist] = artist
        end
      end
    end
  end)

and the tracks eager loader looks like:

  Album.one_to_many :tracks, eager_loader: (proc do |eo_opts|
    eo_opts[:rows].each{|album| album.associations[:tracks] = []}
    id_map = eo_opts[:id_map]
    Track.where(album_id: id_map.keys).all do |track|
      if albums = id_map[track.album_id]
        albums.each do |album|
          album.associations[:tracks] << track
        end
      end
    end
  end)

Now, these are both overly simplistic eager loaders that don't respect cascaded
associations or any of the association options.  But hopefully they both
provide simple examples that you can more easily build and learn from, as 
the custom eager loaders described later in this page are more complex.

Basically, the eager loading steps can be broken down into:

1. Set default association values (nil/[]) for each of the current objects
2. Return just related associated objects by filtering the associated class
   to include only rows with keys present in the id_map.
3. Iterating over the returned associated objects, indexing into the id_map
   using the foreign/primary key value in the associated object to get
   current values associated to that specific object.
4. For each of those current values, updating the cached association value to
   include that specific object.

Using the :eager_loader proc, you should be able to eagerly load all associations
that can be eagerly loaded, even if Sequel doesn't natively support such eager
loading.

== Limited Associations

Sequel supports specifying limits and/or offsets for associations:

  Artist.one_to_many :first_10_albums, class: :Album, order: :release_date, limit: 10

For retrieving the associated objects for a single object, this just uses
a LIMIT:

  artist.first_10_albums
  # SELECT * FROM albums WHERE (artist_id = 1) LIMIT 10

=== Eager Loading via eager

However, if you want to eagerly load an association, you must use a different
approach.  Sequel has 4 separate strategies for dealing with such cases.

The default strategy used on all databases is a UNION-based approach, which
will submit multiple subqueries in a UNION query:

  Artist.where(id: [1,2]).eager(:first_10_albums).all
  # SELECT * FROM (SELECT * FROM albums WHERE (artist_id = 1) LIMIT 10) UNION ALL
  # SELECT * FROM (SELECT * FROM albums WHERE (artist_id = 2) LIMIT 10)

This is the fastest way to load the associated objects on most databases, as long as
there is an index on albums.artist_id.  Without an index it is probably the slowest
approach, so make sure you have an index on the key columns.  If you cannot add an
index, you'll want to manually specify the :eager_limit_strategy option as shown below.

On PostgreSQL, for *_one associations that don't use an offset, you can
choose to use a the distinct on strategy:

  Artist.one_to_one :first_album, class: :Album, order: :release_date,
    eager_limit_strategy: :distinct_on
  Artist.where(id: [1,2]).eager(:first_album).all
  # SELECT DISTINCT ON (albums.artist_id) *
  # FROM albums
  # WHERE (albums.artist_id IN (1, 2))
  # ORDER BY albums.artist_id, release_date
  
Otherwise, if the database supports window functions, you can choose to use
the window function strategy:

  Artist.one_to_many :first_10_albums, class: :Album, order: :release_date, limit: 10,
    eager_limit_strategy: :window_function
  Artist.where(id: [1,2]).eager(:first_10_albums).all
  # SELECT * FROM (
  #   SELECT *, row_number() OVER (PARTITION BY albums.artist_id ORDER BY release_date) AS x_sequel_row_number_x
  #   FROM albums
  #   WHERE (albums.artist_id IN (1, 2))
  # ) AS t1
  # WHERE (x_sequel_row_number_x <= 10)
  
Alternatively, you can use the :ruby strategy, which will fall back to
retrieving all records, and then will slice the resulting array to get
the first 10 after retrieval.

=== Dynamic Eager Loading Limits

If you need to eager load variable numbers of records (with limits that aren't
known at the time of the association definition), Sequel supports an
:eager_limit dataset option that can be defined in an eager loading callback:

  Artist.one_to_many :albums
  Artist.where(id: [1, 2]).eager(albums: lambda{|ds| ds.order(:release_date).clone(eager_limit: 3)}).all
  # SELECT * FROM (
  #   SELECT *, row_number() OVER (PARTITION BY albums.artist_id ORDER BY release_date) AS x_sequel_row_number_x
  #   FROM albums
  #   WHERE (albums.artist_id IN (1, 2))
  # ) AS t1
  # WHERE (x_sequel_row_number_x <= 3)

You can also customize the :eager_limit_strategy on a case-by-case basis by passing in that option in the same way:

  Artist.where(id: [1, 2]).eager(albums: lambda{|ds| ds.order(:release_date).clone(eager_limit: 3, eager_limit_strategy: :ruby)}).all
  # SELECT * FROM albums WHERE (albums.artist_id IN (1, 2)) ORDER BY release_date

The :eager_limit and :eager_limit_strategy options currently only work when
eager loading via #eager, not with #eager_graph.

=== Eager Loading via eager_graph_with_options

When eager loading an association via eager_graph (which uses JOINs), the
situation is similar.  While the UNION-based strategy cannot be used as
you don't know the records being eagerly loaded in advance, Sequel can use
a variant of the other 3 strategies.  By default it retrieves all records
and then does the array slice in ruby.  As eager_graph does not support
options, to use an eager_graph limit strategy you have to use the
eager_graph_with_options method with the :limit_strategy option.

The :distinct_on strategy uses DISTINCT ON in a subquery and JOINs that
subquery:

  Artist.eager_graph_with_options(:first_album, limit_strategy: :distinct_on).all
  # SELECT artists.id, artists.name, first_album.id AS first_album_id,
  #        first_album.name AS first_album_name, first_album.artist_id,
  #        first_album.release_date
  # FROM artists 
  # LEFT OUTER JOIN (
  #   SELECT DISTINCT ON (albums.artist_id) *
  #   FROM albums
  #   ORDER BY albums.artist_id, release_date
  # ) AS first_album ON (first_album.artist_id = artists.id)

The :window_function approach JOINs to a nested subquery using a window
function:

  Artist.eager_graph_with_options(:first_10_albums, limit_strategy: :window_function).all
  # SELECT artists.id, artists.name, first_10_albums.id AS first_10_albums_id,
  #        first_10_albums.name AS first_10_albums_name, first_10_albums.artist_id,
  #        first_10_albums.release_date
  # FROM artists 
  # LEFT OUTER JOIN (
  #   SELECT id, name, artist_id, release_date
  #   FROM (
  #     SELECT *, row_number() OVER (PARTITION BY tracks.album_id ORDER BY release_date) AS x_sequel_row_number_x
  #     FROM albums
  #   ) AS t1 WHERE (x_sequel_row_number_x <= 10)
  # ) AS first_10_albums ON (first_10_albums.artist_id = artists.id)

The :correlated_subquery approach JOINs to a nested subquery using a correlated
subquery:

  Artist.eager_graph_with_options(:first_10_albums, limit_strategy: :correlated_subquery).all
  # SELECT artists.id, artists.name, first_10_albums.id AS first_10_albums_id,
  #        first_10_albums.name AS first_10_albums_name, first_10_albums.artist_id,
  #        first_10_albums.release_date
  # FROM artists 
  # LEFT OUTER JOIN (
  #   SELECT *
  #   FROM albums
  #   WHERE albums.id IN (
  #     SELECT t1.id
  #     FROM tracks AS t1
  #     WHERE (t1.album_id = tracks.album_id)
  #     ORDER BY release_date
  #     LIMIT 10
  #   )
  # ) AS first_10_albums ON (first_10_albums.artist_id = artists.id)

The reason that Sequel does not automatically use the :distinct_on, :window function
or :correlated_subquery strategy for eager_graph is that it can perform much worse than the
default of just doing the array slicing in ruby.  If you are only using eager_graph to
return a few records, it may be cheaper to get all of their associated records and filter
them in ruby as opposed to computing the set of limited associated records for all rows.

It's recommended to only use an eager_graph limit strategy if you have benchmarked
it against the default behavior and found it is faster for your use case.

=== Filtering By Associations

In order to return correct results, Sequel automatically uses a limit strategy when
using filtering by associations with limited associations, if the database supports
it.  As in the eager_graph case, the UNION-based strategy doesn't work.  Unlike
in the eager and eager_graph cases, the array slicing in ruby approach does not work,
you must use an SQL-based strategy.  Sequel will select an appropriate default
strategy based on the database you are using, and you can override it using the
:filter_limit_strategy option.

The :distinct_on strategy:

  Artist.where(first_album: Album[1]).all
  # SELECT *
  # FROM artists
  # WHERE (artists.id IN (
  #   SELECT albums.artist_id
  #   FROM albums
  #   WHERE ((albums.artist_id IS NOT NULL) AND (albums.id IN (
  #     SELECT DISTINCT ON (albums.artist_id) albums.id
  #     FROM albums
  #     ORDER BY albums.artist_id, release_date
  #   )) AND (albums.id = 1))))

The :window_function strategy:

  Artist.where(first_10_albums: Album[1]).all
  # SELECT *
  # FROM artists
  # WHERE (artists.id IN (
  #   SELECT albums.artist_id
  #   FROM albums
  #   WHERE ((albums.artist_id IS NOT NULL) AND (albums.id IN (
  #     SELECT id FROM (
  #       SELECT albums.id, row_number() OVER (PARTITION BY albums.artist_id ORDER BY release_date) AS x_sequel_row_number_x
  #       FROM albums
  #     ) AS t1
  #     WHERE (x_sequel_row_number_x <= 10)
  #   )) AND (albums.id = 1))))

The :correlated_subquery strategy:

  Artist.where(first_10_albums: Album[1]).all
  # SELECT *
  # FROM artists
  # WHERE (artists.id IN (
  #   SELECT albums.artist_id
  #   FROM albums
  #   WHERE ((albums.artist_id IS NOT NULL) AND (albums.id IN (
  #     SELECT t1.id
  #     FROM albums AS t1
  #     WHERE (t1.artist_id = albums.artist_id)
  #     ORDER BY release_date
  #     LIMIT 1
  #   )) AND (albums.id = 1))))

Note that filtering by limited associations does not work on MySQL, as MySQL does not support
any of the strategies.  It's also not supported when using composite keys on databases
that don't support window functions and don't support multiple columns in IN.

=== Additional Association Types

While the above examples for limited associations showed one_to_many and one_to_one associations,
it's just because those are the simplest examples.  Sequel supports all of the same features for
many_to_many and one_through_one associations that are enabled by default, as well as the
many_through_many and one_through_many associations that are added by the many_through_many
plugin.

== More advanced association examples 

=== Association extensions

All associations come with an <tt><i>association</i>_dataset</tt> method that can be further filtered or
otherwise modified:

  class Author < Sequel::Model
    one_to_many :authorships
  end
  Author.first.authorships_dataset.where{number < 10}.first
 
You can extend a dataset with a module using the <tt>:extend</tt> association option.  You can reference
the model object that created the association dataset via the dataset's
+model_object+ method, and the related association reflection via the dataset's
+association_reflection+ method:

  module FindOrCreate
    def find_or_create(vals)
      first(vals) || model.create(vals.merge(association_reflection[:key]=>model_object.id))
    end
  end
  class Author < Sequel::Model
    one_to_many :authorships, extend: FindOrCreate
  end
  Author.first.authorships_dataset.find_or_create(name: 'Blah', number: 10)

=== many_to_many associations through model tables

The many_to_many association can be used even when the join table is a table used for a
model.  The only requirement is the join table has foreign keys to both the current
model and the associated model.  Anytime there is a one_to_many association from model A to
model B, and model B has a many_to_one association to model C, you can use a many_to_many
association from model A to model C.

  class Author < Sequel::Model
    one_to_many :authorships
    many_to_many :books, join_table: :authorships
  end

  class Authorship < Sequel::Model
    many_to_one :author
    many_to_one :book
  end

  @author = Author.first
  @author.books

=== many_to_many for three-level associations

You can even use a many_to_many association between model A and model C if model A has a
one_to_many association to model B, and model B has a one_to_many association to model C.
You just need to use the appropriate :right_key and :right_primary_key options.  And in
the reverse direction from model C to model A, you can use a one_through_one association
using the :left_key and :left_primary_key options.

  class Firm < Sequel::Model
    one_to_many :clients
    many_to_many :invoices, join_table: :clients, right_key: :id, right_primary_key: :client_id
  end 

  class Client < Sequel::Model
    many_to_one :firm
    one_to_many :invoices
  end

  class Invoice < Sequel::Model
    many_to_one :client
    one_through_one :firm, join_table: :clients, left_key: :id, left_primary_key: :client_id
  end

  Firm.first.invoices
  Invoice.first.firm

To handle cases where there are multiple join tables, you can use the many_through_many
plugin that ships with Sequel.

=== Polymorphic Associations

Sequel discourages the use of polymorphic associations, which is the reason they
are not supported by default.  All polymorphic associations can be made non-polymorphic
by using additional tables and/or columns instead of having a column
containing the associated class name as a string.

Polymorphic associations break referential integrity and are significantly more
complex than non-polymorphic associations, so their use is not recommended unless
you are stuck with an existing design that uses them.

If you must use them, look for the sequel_polymorphic external plugin, as it makes using
polymorphic associations in Sequel about as easy as it is in ActiveRecord.  However,
here's how they can be done using Sequel's custom associations (the sequel_polymorphic
external plugin is just a generic version of this code):

  class Asset < Sequel::Model
    many_to_one :attachable, reciprocal: :assets,
      setter: (lambda do |attachable|
        self[:attachable_id] = (attachable.pk if attachable)
        self[:attachable_type] = (attachable.class.name if attachable)
      end),
      dataset: (proc do
        klass = attachable_type.constantize
        klass.where(klass.primary_key=>attachable_id)
      end),
      eager_loader: (lambda do |eo|
        id_map = {}
        eo[:rows].each do |asset|
          asset.associations[:attachable] = nil 
          ((id_map[asset.attachable_type] ||= {})[asset.attachable_id] ||= []) << asset
        end
        id_map.each do |klass_name, id_map|
          klass = klass_name.constantize
          klass.where(klass.primary_key=>id_map.keys).all do |attach|
            id_map[attach.pk].each do |asset|
              asset.associations[:attachable] = attach
            end
          end
        end
      end)
  end 

  class Post < Sequel::Model
    one_to_many :assets, key: :attachable_id, reciprocal: :attachable, conditions: {attachable_type: 'Post'},
      adder: lambda{|asset| asset.update(attachable_id: pk, attachable_type: 'Post')},
      remover: lambda{|asset| asset.update(attachable_id: nil, attachable_type: nil)},
      clearer: lambda{assets_dataset.update(attachable_id: nil, attachable_type: nil)}
  end

  class Note < Sequel::Model
    one_to_many :assets, key: :attachable_id, reciprocal: :attachable, conditions: {attachable_type: 'Note'},
      adder: lambda{|asset| asset.update(attachable_id: pk, attachable_type: 'Note')},
      remover: lambda{|asset| asset.update(attachable_id: nil, attachable_type: nil)},
      clearer: lambda{assets_dataset.update(attachable_id: nil, attachable_type: nil)}
  end

  @asset.attachable = @post
  @asset.attachable = @note

=== Joining on multiple keys

Let's say you have two tables that are associated with each other with multiple
keys.  This can be handled using Sequel's built in composite key support for
associations:

  # Both of these models have an album_id, number, and disc_number fields.
  # All FavoriteTracks have an associated track, but not all tracks have an
  # associated favorite track
  
  class Track < Sequel::Model
    many_to_one :favorite_track, key: [:disc_number, :number, :album_id], primary_key: [:disc_number, :number, :album_id]
  end
  class FavoriteTrack < Sequel::Model
    one_to_one :tracks, key: [:disc_number, :number, :album_id], primary_key: [:disc_number, :number, :album_id]
  end

=== Tree - All Ancestors and Descendants

Let's say you want to store a tree relationship in your database, it's pretty
simple:

  class Node < Sequel::Model
    many_to_one :parent, class: self
    one_to_many :children, key: :parent_id, class: self
  end
  
You can easily get a node's parent with node.parent, and a node's children with
node.children.  You can even eager load the relationship up to a certain depth:

  # Eager load three generations of generations of children for a given node 
  Node.where(id: 1).eager(children: {children: :children}).all.first
  # Load parents and grandparents for a group of nodes
  Node.where{id < 10}.eager(parent: :parent).all

What if you want to get all ancestors up to the root node, or all descendants,
without knowing the depth of the tree?

  class Node < Sequel::Model
    many_to_one :ancestors, class: self,
     eager_loader: (lambda do |eo|
      # Handle cases where the root node has the same parent_id as primary_key
      # and also when it is NULL
      non_root_nodes = eo[:rows].reject do |n| 
        if [nil, n.pk].include?(n.parent_id)
          # Make sure root nodes have their parent association set to nil
          n.associations[:parent] = nil 
          true
        else
          false
        end 
      end 
      unless non_root_nodes.empty?
        id_map = {}
        # Create an map of parent_ids to nodes that have that parent id
        non_root_nodes.each{|n| (id_map[n.parent_id] ||= []) << n}
        # Doesn't cause an infinite loop, because when only the root node
        # is left, this is not called.
        Node.where(id: id_map.keys).eager(:ancestors).all do |node|
          # Populate the parent association for each node
          id_map[node.pk].each{|n| n.associations[:parent] = node}
        end 
      end 
    end)
    many_to_one :descendants, eager_loader: (lambda do |eo|
      id_map = {}
      eo[:rows].each do |n| 
        # Initialize an empty array of child associations for each parent node
        n.associations[:children] = []
        # Populate identity map of nodes
        id_map[n.pk] = n 
      end 
      # Doesn't cause an infinite loop, because the :eager_loader is not called
      # if no records are returned.  Exclude id = parent_id to avoid infinite loop
      # if the root note is one of the returned records and it has parent_id = id
      # instead of parent_id = NULL.
      Node.where(parent_id: id_map.keys).exclude(id: :parent_id).eager(:descendants).all do |node|
        # Get the parent from the identity map
        parent = id_map[node.parent_id]
        # Set the child's parent association to the parent 
        node.associations[:parent] = parent
        # Add the child association to the array of children in the parent
        parent.associations[:children] << node
      end 
    end)
  end

Note that Sequel ships with an rcte_tree plugin that does all of the above and more:

  class Node < Sequel::Model
    plugin :rcte_tree
  end

=== Joining multiple keys to a single key, through a third table

Let's say you have a database of songs, lyrics, and artists.  Each song
may or may not have a lyric (most songs are instrumental).  The lyric can be
associated to an artist in each of four ways: composer, arranger, vocalist,
or lyricist.  These may all be the same, or they could all be different, and
none of them are required.  The songs table has a lyric_id field to associate
it to the lyric, and the lyric table has four fields to associate it to the
artist (composer_id, arranger_id, vocalist_id, and lyricist_id).

What you want to do is get all songs for a given artist, ordered by the song's
name, with no duplicates?

  class Artist < Sequel::Model
    one_to_many :songs, order: Sequel[:songs][:name],
      dataset: proc{Song.select_all(:songs).join(:lyrics, id: :lyric_id, id=>[:composer_id, :arranger_id, :vocalist_id, :lyricist_id])},
      eager_loader: (lambda do |eo|
        h = eo[:id_map]
        ids = h.keys
        eo[:rows].each{|r| r.associations[:songs] = []}
        Song.select_all(:songs).
          select_append{[lyrics[:composer_id], lyrics[:arranger_id], lyrics[:vocalist_id], lyrics[:lyricist_id]]}.
          join(:lyrics, id: :lyric_id){Sequel.or(composer_id: ids, arranger_id: ids, vocalist_id: ids, lyricist_id: ids)}.
          order{songs[:name]}.all do |song|
            [:composer_id, :arranger_id, :vocalist_id, :lyricist_id].each do |x|
              recs = h[song.values.delete(x)]
              recs.each{|r| r.associations[:songs] << song} if recs
            end
          end
        eo[:rows].each{|r| r.associations[:songs].uniq!}
      end)
  end

=== Statistics Associations (Sum of Associated Table Column)

In addition to getting associated records, you can use Sequel's association support
to get aggregate information for columns in associated tables (sums, averages, etc.).

Let's say you have a database with projects and tickets.  A project can have many
tickets, and each ticket has a number of hours associated with it.  You can use the
association support to create a Project association that gives the sum of hours for all
associated tickets.

  class Project < Sequel::Model
    one_to_many :tickets
    many_to_one :ticket_hours, read_only: true, key: :id,
     dataset: proc{Ticket.where(project_id: id).select{sum(hours).as(hours)}},
     eager_loader: (lambda do |eo|
      eo[:rows].each{|p| p.associations[:ticket_hours] = nil}
      Ticket.where(project_id: eo[:id_map].keys).
       select_group(:project_id).
       select_append{sum(hours).as(hours)}.
       all do |t|
        p = eo[:id_map][t.values.delete(:project_id)].first
        p.associations[:ticket_hours] = t
       end
     end)
    # The association method returns a Ticket object with a single aggregate
    # sum-of-hours value, but you want it to return an Integer/Float of just the
    # sum of hours, so you call super and return just the sum-of-hours value.
    # This works for both lazy loading and eager loading.
    def ticket_hours
      if s = super
        s[:hours]
      end
    end
  end
  class Ticket < Sequel::Model
    many_to_one :project
  end

Note that it is often better to use a sum cache instead of this approach.  You can implement
a sum cache using +after_create+, +after_update+, and +after_delete+ hooks, or preferably using a database trigger.