Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
237 lines (186 sloc) 9.89 KB

Ember Data Architecture

Roles & Responsibilities


The store is the primary interface between the application developer and the data store. It is responsible for managing all available records, both materialized and immaterialized. At its core, it is a bookkeeping object that indexes loaded hashes, and serves as a coordinator between the other objects in the system.

  • Indexes data hashes by type and ID
  • Supplies a clientId for each requested record, and maps type/ID to clientIds and vice versa.
  • Serves as an identity map for records of a given type/ID
  • Creates new records and transactions
    • By default, Post.createRecord() asks the default store to create the record.
    • Optionally, coordinates with adapter to generate a client-generated ID for new records.
  • Coordinates with the adapter to request records (find, findMany, findAll, findQuery).
  • Sends lifecycle events to records. For example, the store notifies a record when the adapter has saved its pending changes (didCommit)
  • Serves as the callback target for the adapter (didCreateRecord, et al)
  • Responsible for managing indexes that power live record arrays
    • Filters: when a new data hash is loaded into the store, it updates any filters registered on that type. Records notify the store (via hashWasUpdated) when any properties change, causing the filters to update.
    • find(): find() is a special filter that matches all records for a given type.


A model defines the attributes and relationships for a given type. Instances of models, called records, are objects that provide an Ember interface to JSON hashes returned by the server. Internally, records keep track of their original JSON hash and any unsaved changes (see DataProxy below for more details).

Records move through states in a state manager throughout their life. For example, a newly created record begins its life in the loaded.created state. A record requested from the server starts in the loading state, and moves into the loaded.saved state once the server returns its JSON hash.

When a store materializes a record, it asks the adapter (see below) to extract the record's attributes and associations and normalize their names. This means that records will always have normalized data hashes.

  • Has a series of lifecycle flags (isLoaded, etc.)
  • Serializes the record into a persistable JSON hash, accepting adapter-provided options (such as includeForeignKeys).
  • Manages an underlying DataProxy
  • Manages a StateManager and sends any events to its state manager
  • Tracks its current transaction
  • Sends events to the transaction when the record becomes dirty
  • Updates materialized ManyArrays if the underlying data changes
  • Aliases store methods that require a type parameter to the DS.Model type. For example, instead of requiring you to call store.find(App.Person, 1), you can say App.Person.find(1).


A record's DataProxy wraps its server-returned JSON hash plus any unsaved changes in a single object.

It also supports commit, which collapses the unsaved changes into the saved changes, and rollback, which discards any unsaved changes.

Record State Manager

Manages the current state of a record. Every record has its own instance of the StateManager.

When events occur to the record (e.g. the data hash changes, the store acknowledges its commit), the record sends events to the state manager. This allows the record to have context-specific responses to these events, and initiate state transitions in response to events.

There is a lot of specific documentation in system/model/states.js.


A transaction represents a unit of work that can be atomically committed to the adapter. When a transaction is committed, it is responsible for providing all of the changes to the adapter to save. A transaction can also be rolled back, which reverts any changes that occurred but had not yet been saved to the adapter.

Every record must belong to a transaction. By default, records belong to the default transaction, which is a transaction that is implicitly created with the store.

Transactions are ephemeral objects. Once committed or rolled back, they should not be used again.

  • Stores references to records, grouped by the current state of the record.
    • For example, a newly created record is saved in the created bucket, while a record that has attributes changed is saved in the updated bucket.
  • Stores descriptions of changed relationships. When a relationship changes, information about its old parent, new parent, and new child is saved in the transaction.
  • Raises an exception if changes in relationships are made between records that are in different transactions.
  • Able to move records into itself from another transaction if it is legal.
  • When committed, provides changed records to the adapter and responsible for moving those records into an inFlight state.
  • After committing or rolling back, moves clean records into the store's default transaction.
  • When rolled back, the transaction notifies all changed records to discard changes.


Record arrays represent an ordered list of records. They are backed by an array of client IDs. When retrieving a record from the record array, it will be materialized lazily if necessary.

DS.RecordArray is an abstract base class that provides many of the features needed by its concrete implementations, described below.


Represents a one-to-many relationship. When the association is retrieved from a record, a ManyArray is created that contains an array of the client IDs that belong to that record.

  • Notifies the transaction if the relationship is modified
  • Tracks aggregate state of member records via isLoaded flag
  • Updates added records to point their inverse association to the new parent.


Represents an ordered list of records whose order and membership is determined by the adapter. For example, a query sent to the adapter may trigger a search on the server, whose results would be loaded into an instance of the AdapterPopulatedRecordArray.


Represents a list of records whose membership is determined by the store. As records are created, loaded, or modified, the store evaluates them to determine if they should be part of the record array.


The adapter is responsible for translating a store request into the appropriate action to take against a persistence layer. For example, a REST adapter may translate the request to find a record of type App.Photo with ID 1 into an HTTP GET request to /photos/1.

The responsibility of the adapter fall into two general categories: retrieving records and committing changes to records.

Finding Via an Adapter

  • Loading records into the store in response to find()
  • Loading multiple records into the store in response to findMany()
  • Loading the results of a query into an AdapterPopulatedRecordArray in response to a findQuery()
  • Loading records into the store in response to findAll()

Saving Changes

The adapter receives a list of all changes from a transaction in its commit() method. It is responsible for evaluating those changes, figuring out what to do in order to persist them, and letting the store know when the server acknowledged the save for a given record.

As part of this process, the adapter receives a list of all created, updated, and deleted records, as well as a list of all changes to relationships.

In order to make this easy for an adapter to implement this pattern, the DS.Adapter abstract class offers some conveniences:

  • If a record has no attribute changes, but is involved in a relationship change, the abstract DS.Adapter calls the shouldCommit method with the ambiguous record and the relationship changes.
    • In a relational model, for example, the adapter will return true if the record is the child of a relationship change and false if the record is the old or new parent.
    • If the shouldCommit method returns false, the abstract commit method will immediately call didUpdateRecord on the store.
  • If a record is involved in a relationship change, the abstract commit method will call the adapter's willCommit method with the record and the list of relationships.
    • This gives the adapter an opportunity to pend the record. For example, if a child record needs a foreign key, but the parent record's ID does not exist yet, the adapter can wait for the parent ID to become populated.
  • The abstract commit method will call createRecords, updateRecords, and deleteRecords to allow the adapter to break up the commits to the server in an appropriate way.

Client-Side ID Generation

Adapters can specify a mechanism for new records to generate client-side IDs. In general, this method should return a UUID or something with extremely low collission possibility.

When a store creates a new record, it first consults the adapter to determine whether the ID can be generated on the client. If so, it will apply the generated ID to the record immediately.

One major benefit of generating IDs on the client is that records do not need to wait for associated records to be saved in order to retrieve their foreign keys.

Naming Conventions

The adapter is also responsible for normalizing a server-provided data hash to the naming expected by Ember.

In general, this means converting underscored names to camelcased names.

It is also responsible for converting dirty records into a data hash expected by the server. For example, the adapter may need to add a foreign key to the data hash by adding _id to its association.

The abstract adapter class provides normalization functions that call into the namingConvention hash in the concrete classes. For very custom logic, the concrete classes may want to override the normalization directly, but that should be very rare.