Navigation

Threading

Overview

To make your Xamarin and C#/.NET apps fast and responsive, you must balance the computing time needed to lay out the visuals and handle user interactions with the time needed to process your data and run your business logic. Typically, app developers spread this work across multiple threads: the main or UI thread for all of the user interface-related work, and one or more background threads to compute heavier workloads before sending it to the UI thread for presentation. By offloading heavy work to background threads, the UI thread can remain highly responsive regardless of the size of the workload. But it can be notoriously difficult to write thread-safe, performant, and maintainable multithreaded code that avoids issues like deadlocking and race conditions. MongoDB Realm aims to simplify this for you.

Three Rules to Keep in Mind

MongoDB Realm enables simple and safe multithreaded code when you follow these three rules:

Don’t lock to read:
Realm Database’s Multiversion Concurrency Control (MVCC) architecture eliminates the need to lock for read operations. The values you read will never be corrupted or in a partially-modified state. You can freely read from realms on any thread without the need for locks or mutexes. Unnecessarily locking would be a performance bottleneck since each thread might need to wait its turn before reading.
Avoid writes on the UI thread if you write on a background thread:
You can write to a realm from any thread, but there can be only one writer at a time. Consequently, write transactions block each other. A write on the UI thread may result in your app appearing unresponsive while it waits for a write on a background thread to complete. If you are using Realm Sync, avoid writing on the UI thread as Sync writes on a background thread.
Don’t pass live objects, collections, or realms to other threads:
Live objects, collections, and realm instances are thread-confined: that is, they are only valid on the thread on which they were created. Practically speaking, this means you cannot pass live instances to other threads. However, Realm Database offers several mechanisms for sharing objects across threads.

Communication Across Threads

You can have the same realm open on multiple threads as separate realm instances. You are free to read and write with realm instances on the thread where you first opened them. One of the key rules when working with Realm Database in a multithreaded environment is that objects are thread-confined: you may not access the instances of a realm, collection, or object that originated on other threads. Realm Database’s Multiversion Concurrency Control (MVCC) architecture means that there could be many active versions of an object at any time. Thread-confinement ensures that all instances in that thread are of the same internal version.

A realm instance is designed to work with one version at a time, not several different versions. Consider what Realm Database would have to do to support this: it would need to store a potentially enormous graph to allow the realm instance to reconcile the different object versions internally. Faced with this, it seems more reasonable to impose the limitation that you cannot pass live instances across threads. This design choice keeps the implementation relatively simple, more space-efficient, and more performant as a result.

When you need to communicate across threads, you have several options depending on your use case:

  • To work modify the data on two threads, query for the object on both threads.
  • To react to changes made on any thread, use Realm Database’s notifications.
  • To see changes from other threads in the realm on the current thread, refresh your realm instance.
  • To send a fast, read-only view of the object to other threads, “freeze” the object.
  • To keep and share many read-only views of the object in your app, copy the object from the realm.

Refreshing Realms

When you open a realm, it reflects the most recent successful write commit and remains on that version until it is refreshed. This means that the realm will not see changes that happened on another thread until the next refresh. Realms on the UI thread – more precisely, on any event loop thread – automatically refresh themselves at the beginning of that thread’s loop. However, you must manually refresh realm instances that do not exist on loop threads or that have auto-refresh disabled.

Frozen Objects

Live, thread-confined objects work fine in most cases. However, some apps – those based on reactive, event stream-based architectures, for example – need to send immutable copies around to many threads for processing before ultimately ending up on the UI thread. Making a deep copy every time would be expensive, and Realm Database does not allow live instances to be shared across threads. In this case, you can freeze objects, collections, and realms.

Freezing creates an immutable view of a specific object, collection, or realm that still exists on disk and does not need to be deeply copied when passed around to other threads. You can freely share a frozen object across threads without concern for thread issues.

When working with frozen objects, an attempt to do any of the following throws an exception:

  • Opening a write transaction on a frozen realm.
  • Modifying a frozen object.
  • Adding a change listener to a frozen realm, collection, or object.

Once frozen, it is not possible to unfreeze an object. You can use isFrozen() to check if the object is frozen. This method is always thread-safe.

To modify a frozen object, query for it on an unfrozen realm, then modify it.

Frozen objects are not live and do not automatically update. They are effectively snapshots of the object state at the time of freezing.

When you freeze a realm, its child objects also become frozen.

Frozen objects remain valid as long as the live realm that spawned them stays open. Therefore, avoid closing the live realm until all threads are done with the frozen objects. You can close frozen realm before the live realm is closed.

On caching frozen objects

Caching too many frozen objects can have a negative impact on the realm file size. “Too many” depends on your specific target device and the size of your Realm objects. If you need to cache a large number of versions, consider copying what you need out of the realm instead.

Realm’s Threading Model in Depth

Realm Database provides safe, fast, lock-free, and concurrent access across threads with its Multiversion Concurrency Control (MVCC) architecture.

Compared and Contrasted with Git

If you are familiar with a distributed version control system like Git, you may already have an intuitive understanding of MVCC. Two fundamental elements of Git are:

  • Commits, which are atomic writes.
  • Branches, which are different versions of the commit history.

Similarly, Realm Database has atomically-committed writes in the form of transactions. Realm Database also has many different versions of the history at any given time, like branches.

Unlike Git, which actively supports distribution and divergence through forking, a realm only has one true latest version at any given time and always writes to the head of that latest version. Realm Database cannot write to a previous version. This makes sense: your data should converge on one latest version of the truth.

Internal Structure

A realm is implemented using a B+ tree data structure. The top-level node represents a version of the realm; child nodes are objects in that version of the realm. The realm has a pointer to its latest version, much like how Git has a pointer to its HEAD commit.

Realm Database uses a copy-on-write technique to ensure isolation and durability. When you make changes, Realm Database copies the relevant part of the tree for writing. Realm Database then commits the changes in two phases:

  • Realm Database writes changes to disk and verifies success.
  • Realm Database then sets its latest version pointer to point to the newly-written version.

This two-step commit process guarantees that even if the write failed partway, the original version is not corrupted in any way because the changes were made to a copy of the relevant part of the tree. Likewise, the realm’s root pointer will point to the original version until the new version is guaranteed to be valid.

Diagram

The following diagram illustrates the commit process:

  1. The realm is structured as a tree. The realm has a pointer to its latest version, V1.
  2. When writing, Realm Database creates a new version V2 based on V1. Realm Database makes copies of objects for modification (A1, C1), while links to unmodified objects continue to point to the original versions (B, D).
  3. After validating the commit, Realm Database updates the realm’s pointer to the new latest version, V2. Realm Database discards old nodes no longer connected to the tree.

Realm Database uses zero-copy techniques like memory mapping to handle data. When you read a value from the realm, you are virtually looking at the value on the actual disk, not a copy of it. This is the basis for live objects. This is also why a realm head pointer can be set to point to the new version after the write to disk has been validated.

Summary

  • MongoDB Realm enables simple and safe multithreaded code when you follow three rules: - don’t lock to read
    • avoid writes on the UI thread if you write on background threads or use Realm Sync
    • don’t pass live objects to other threads.
  • There is a proper way to share objects across threads for each use case.
  • In order to see changes made on other threads in your realm instance, you must manually refresh realm instances that do not exist on “loop” threads or that have auto-refresh disabled.
  • For apps based on reactive, event-stream-based architectures, you can freeze objects, collections, and realms in order to pass shallow copies around efficiently to different threads for processing.
  • Realm Database’s multiversion concurrency control (MVCC) architecture is similar to Git’s. Unlike Git, Realm Database has only one true latest version for each realm.
  • Realm Database commits in two stages to guarantee isolation and durability.