the Chromium logo

The Chromium Projects

Chrome Sync's Model API

Chrome Sync operates on discrete, explicitly defined model types (bookmarks, preferences, tabs, etc). These model types are individually responsible for implementing their own local storage and responding to remote changes. This guide is for developers interested in syncing data for their model type to the cloud using Chrome Sync. It describes the newest version of the API, known as Unified Sync and Storage (USS). There is also the deprecated SyncableService API (aka Directory), which as of early 2022 is still used by several legacy model types, but "wrapped into" USS (see SyncableServiceBasedBridge).

Overview

To correctly sync data, USS requires that sync metadata be stored alongside your model data in a way such that they are written together atomically. This is very important! Sync must be able to update the metadata for any local data changes as part of the same write to disk. If you attempt to write data to disk and only notify sync afterwards, a crash in between the two writes can result in changes being dropped and never synced to the server, or data being duplicated due to being committed more than once.

ModelTypeSyncBridge is the interface the model code must implement. The bridge is usually owned by a KeyedService. The correct place for the bridge generally lies as close to where your model data is stored as possible, as the bridge needs to be able to inject metadata updates into any local data changes that occur.

The bridge owns a ModelTypeChangeProcessor object, which it uses to communicate local changes to sync using the Put and Delete methods. The processor will communicate remote changes from sync to the bridge using the MergeSyncData and ApplySyncChanges methods, respectively for the initial merge of remote and local data, and for incremental changes coming from sync. MetadataChangeList is the way sync communicates metadata changes to the storage mechanism. Note that it is typically implemented on a per-storage basis, not a per-type basis.

Data

Specifics

Model types will define a proto that contains the necessary fields of the corresponding native type (e.g. TypedUrlSpecifics contains a URL and a list of visit timestamps) and include it as a field in the generic EntitySpecifics proto. This is the form that all communications with sync will use. This proto form of the model data is referred to as the specifics.

Identifiers

There are two primary identifiers for entities: storage key and client tag. The bridge will need to take an EntityData object (which contains the specifics) and be able generate both of these from it. For non-legacy types without significant performance concerns, these will generally be the same.

The storage key is meant to be the primary key in the local model/database. It’s what’s used to refer to entities most of the time and, as its name implies, the bridge needs to be able to look up local data and metadata entries in the store using it. Because it is a local identifier, it can change as part of database migrations, etc. This may be desirable for efficiency reasons.

The client tag is used to generate the client tag hash, which will identify entities across clients. This means that its implementation can never change once entities have begun to sync, without risking massive duplication of entities. This means it must be generated using only immutable data in the specifics. If your type does not have any immutable fields to use, you will need to add one (e.g. a GUID, though be wary as they have the potential to conflict). While the hash gets written to disk as part of the metadata, the tag itself is never persisted locally.

Storage

A crucial requirement of USS is that the model must add support for keeping sync’s metadata in the same storage as its normal data. The metadata consists of one EntityMetadata proto for each data entity, and one ModelTypeState proto containing metadata pertaining to the state of the entire type (the progress marker, for example). This typically requires two extra tables in a database to do (one for each type of proto).

Since the processor doesn’t know anything about the store, the bridge provides it with an implementation of the MetadataChangeList interface. The change processor writes metadata through this interface when changes occur, and the bridge simply has to ensure it gets passed along to the store and written along with the data changes.

ModelTypeStore

While the model type may store its data however it chooses, many types use ModelTypeStore, which was created specifically to provide a convenient persistence solution. It’s backed by a LevelDB to store serialized protos to disk. ModelTypeStore provides two MetadataChangeList implementations for convenience; both accessed via ModelTypeStore::WriteBatch. One passes metadata changes directly into an existing WriteBatch and another caches them in memory until a WriteBatch exists to consume them.

The store interface abstracts away the type and will handle setting up tables for the type’s data, so multiple ModelTypeStore objects for different types can share the same LevelDB backend just by specifying the same path and task runner. Sync provides a backend that can be shared by all types via the ModelTypeStoreService.

Implementing ModelTypeSyncBridge

The responsibility of the bridge is to accept incoming changes from Sync and apply them to the local model (via MergeSyncData and ApplySyncChanges), as well as watch for local changes and send them to Sync (via the passed-in ModelTypeChangeProcessor's Put and Delete methods).

Threading

The bridge should live on the "model thread", i.e. the thread on which the local model itself lives. This can either be a background/database thread (typically for types which have a pre-existing persistence layer, e.g. TypedURLSyncBridge or PasswordSyncBridge), or it can be the UI thread (e.g. SendTabToSelfBridge or ReadingListStore). Either way, the bridge must be able to synchronously handle updates from Sync. If the bridge lives on the UI thread, then the actual persistence (e.g. to ModelTypeStore) will be asynchronous.

Initialization

The bridge is required to load all of the metadata for its type from storage and provide it to the processor via the ModelReadyToSync method before any local changes occur. This can be tricky if the thread the bridge runs on is different from the storage mechanism. No data will be synced with the server if the processor is never informed that the model is ready.

Since the tracking of changes and updating of metadata is completely independent, there is no need to wait for the sync engine to start before changes can be made.

MergeSyncData

This method is called only once, when a type is first enabled. Sync downloads all the data it has for the type from the server and provides it to the bridge using this method. Sync filters out any tombstones for this call, so EntityChange::type() will never be ACTION_DELETE for the provided entities. The bridge must then examine the sync data and the local data and merge them together:

The MetadataChangeList passed into the function is already populated with metadata for all the data passed in (note that neither the data nor the metadata have been committed to storage yet at this point). It must be given to the processor for any Put or Delete calls so the relevant metadata can be added/updated/deleted, and then passed to the store for persisting along with the data.

Note that if sync gets disabled and the metadata cleared, entities that originated from other clients will exist as “local” entities the next time sync starts and merge is called. Since tombstones are not provided for merge, this can result in reviving the entity if it had been deleted on another client in the meantime.

ApplySyncChanges

This method is called whenever new changes have been downloaded from the server. These changes must be applied to the local model.

Here’s an example implementation of a type using ModelTypeStore:

absl::optional<ModelError> DeviceInfoSyncBridge::ApplySyncChanges(
    std::unique_ptr<MetadataChangeList> metadata_change_list,
    EntityChangeList entity_changes) {
  std::unique_ptr<WriteBatch> batch = store_->CreateWriteBatch();
  for (const EntityChange& change : entity_changes) {
    if (change.type() == EntityChange::ACTION_DELETE) {
      batch->DeleteData(change.storage_key());
    } else {
      batch->WriteData(change.storage_key(),
                       change.data().specifics.your_type().SerializeAsString());
    }
  }

  batch->TakeMetadataChangesFrom(std::move(metadata_change_list));
  store_->CommitWriteBatch(std::move(batch), base::BindOnce(...));
  NotifyModelOfChanges();
  return {};
}

A conflict can occur when an entity has a pending local commit when an update for the same entity comes from another client. In this case, the bridge’s ResolveConflict method will have been called prior to the ApplySyncChanges call in order to determine what should happen. This method defaults to having the remote version overwrite the local version unless the remote version is a tombstone, in which case the local version wins.

Local changes

The ModelTypeChangeProcessor must be informed of any local changes via its Put and Delete methods. Since the processor cannot do any useful metadata tracking until MergeSyncData is called, the IsTrackingMetadata method is provided. It can be checked as an optimization to prevent unnecessary processing preparing the parameters to a Put or Delete call.

Here’s an example of handling a local write using ModelTypeStore:

void WriteLocalChange(std::string key, ModelData data) {
  std::unique_ptr<WriteBatch> batch = store_->CreateWriteBatch();
  if (change_processor()->IsTrackingMetadata()) {
    change_processor()->Put(key, ModelToEntityData(data),
                            batch->GetMetadataChangeList());
  }
  batch->WriteData(key, data.specifics->SerializeAsString());
  store_->CommitWriteBatch(std::move(batch), base::BindOnce(...));
}

Error handling

If any errors occur during store operations that could compromise the consistency of the data and metadata, the processor’s ReportError method should be called. The only exception to this is errors during MergeSyncData or ApplySyncChanges, which should just return a ModelError.

This will inform sync of the error, which will stop all communications with the server so bad data doesn’t get synced. Since the metadata might no longer be valid, the bridge will asynchronously receive an ApplyStopSyncChanges call with a non-null MetadataChangeList parameter. All the metadata will be cleared from the store (if possible), and the type will be started again from scratch on the next client restart.

Sync Integration Checklist

Testing

The TwoClientTypedUrlsSyncTest suite is probably a good place to start for integration testing. Especially note the use of a StatusChangeChecker to wait for events to happen.