Compression

Bit Packing / Mostly Encoding

Realm optimizes the memory usages for some data structures. For example, Realm converts a giant list of strings into enums, which are similar to the tagged pointers in Objective-C, in order for fast lookup. Moreover, the database utilizes integer packing to store the integer with optimized space so that it is not necessary for users to specify the integer type when declaration.

Concurrency Control

Multi-version Concurrency Control (MVCC)

In order to support concurrent read/write operations under the constraint of data consistency, Realm implements Multi-version Concurrency Control based on copy-on-write mechanism. Every time a new transaction begins, the database creates a snapshot of the whole database and all the following read/write operations are performed on the snapshot without any modification on the existing data. Once the transaction commits, the database verifies everything and atomically persists the changes into the disk safely based on two-phase commit protocol. Under this architecture, concurrent threads are not affected by each other and the data in the database is intact even if the write operations are interrupted because of the software failure.

Data Model

Object-Oriented

Realm is an object-oriented database which provides developers with an object interface in multiple programming languages in order to solve the mismatch between the object layer and database layer in object-oriented development.

Foreign Keys

Not Supported

Not Supported

Indexes

B+Tree

Realm's core storage architecture is based on B+Trees.

Joins

Not Supported

Logging

Not Supported

Log component in Realm is used for synchronization, which records the changes made by each transaction and coordinates the operations between processes/threads. It cannot be used for recovery in Realm.

Query Execution

Tuple-at-a-Time Model

Realm implements the lazy-loading feature during the query execution. When the query is executed, not all of the results are read from local disk into the memory. Instead, the system only retrieves the data that is related to the object or property accessed by users. The properties of all other objects are not fetched.

Query Interface

Custom API

Realm provides users a custom API to get the objects from the database based on properties. Moreover, users can filter and sort the query results as needed.

Storage Architecture

Disk-oriented

Realm supports both the disk-oriented and in-memory mode, which is configured in the run-time. When working as a disk-oriented DBMS, the database uses the memory-mapped files to place the contents of an object into the process's address space so that the system can get direct access to the raw data in the local disk without serialization/deserialization as if the file is in the memory.

Storage Model

Decomposition Storage Model (Columnar)

In Realm, the properties are stored contiguously in a columnar way. When a single property of the object in the database needs to be read/updated, the system doesn't need to load the entire row of data that belongs to the object but instead iterates over only that single property of a certain column. In this way, the database avoids unnecessary I/O operations for reading unused properties .

Storage Organization

Copy-on-Write / Shadow Paging

Under the MVCC architecture, Realm implements copy-on-write to avoid the data inconsistency caused by concurrent read/write operations and system crash. The system creates a snapshot of whole database at the beginning of transaction, performs the read/write behaviors on the snapshot and atomically updates the top-level pointer to migrate writes into the disk after everything verified based on the two-phase commit protocol.

System Architecture

Embedded

Realm is an embedded database designed for the mobile device.

Views

Virtual Views Materialized Views

Realm supports reflective views and imperative views. Essentially they are similar to materialized views and virtual views. For the imperative view, it never reruns the query so that the results of query it holds are fresh only in the beginning of the transaction. For the reflective view, the query results it holds always match the data in the database. Whenever the underlying tables changes, the corresponding changes are immediately reflected in the views.