ContextQMD
Back to Gitlabhq

Geo repository synchronization

doc/development/geo/repository_sync.md

18.11.28.0 KB
Original Source

Repository replication (also called synchronization or sync) is a core Geo feature that replicates Git repositories from primary to secondary sites. This ensures that secondary sites maintain up-to-date copies of all repositories, enabling disaster recovery and geographic distribution of Git data.

The Geo::ProjectRepositoryRegistry class defines the model used to track the state of repository replication on the secondary sites. For each project in the main (primary) database, one record in the tracking database is kept.

It records the following about repositories:

  • The last time they were synced.
  • The last time they were successfully synced.
  • If they need to be resynced.
  • When a retry should be attempted.
  • The number of retries.
  • If and when they were verified.

Sync triggers

Repository synchronization can be triggered through different mechanisms.

mermaid
flowchart TD
        A(["Start"])
        A --> B{"What triggers the repository sync?"}
        B --> C["Resync from the admin interface"]
        B --> D["An update event on the repository (for example, git push)"]
        B --> E["Cron worker periodic sync"]
        B --> F["Rails console sync: \n 'replicator.sync'"]
        C --> G["Geo::EventWorker"]
        D --> G
        E --> H["Geo::SyncWorker"]
        F --> I["Geo::FrameworkRepositorySyncService.execute"]
        H --> I
        G --> I

These triggers boil down to two categories: background and event-driven syncs.

Background syncs

To ensure all repositories are synchronized, Geo::RepositoryRegistrySyncWorker runs periodically in the background on secondary sites. Specifically, this worker:

  • Queries the tracking database for registries that need syncing. These can be:
    • Unsynced: Projects that have never been synced on the secondary site and so do not exist yet.
    • Updated recently: Projects that have a last_repository_updated_at timestamp that is more recent than the last_repository_successful_sync_at timestamp in the Geo::ProjectRepositoryRegistry model.
    • Manually resynced: The administrator can manually flag a repository to resync in the Geo Admin area.
  • Enqueues Geo::SyncWorker jobs for each repository that needs updating.
  • Provides baseline synchronization for all repositories.

Event-driven syncs

Event-driven syncs are triggered by repository updates on the primary site. When a repository is updated, a new record is created in the Geo events table for this particular repository update. The Geo Log Cursor daemon, which runs on the secondary site, picks the event up and schedules a Geo::EventWorker job, which runs the Geo::EventService. This service:

  • Finds the repository replicator, and runs its consume(event) method.
  • Changes registry state to pending before attempting sync to avoid race conditions
  • Ensures latest changes are replicated quickly to secondary sites

Sequence diagram

The following diagram illustrates how repository synchronization flows through the system. The two paths are represented on the same diagram as they both end in the same fashion:

mermaid
sequenceDiagram
    participant Worker as Geo::EventWorker
    participant Service as Geo::EventService
    participant Scheduler as Geo::RepositoryRegistrySyncWorker
    participant SyncWorker as Geo::SyncWorker
    participant SyncService as Geo::SyncService
    participant Replicator as Geo::ProjectRepositoryReplicator
    participant Registry as Geo::ProjectRepositoryRegistry
    participant FRSS as Geo::FrameworkRepositorySyncService
    participant Primary as Primary Site

    Note over Worker,Replicator: Event-driven sync path
    Worker->>Service: new(replicable_name, event_name, payload)
    Service->>Replicator: new(model_record_id: payload[:model_record_id])
    Replicator-->>Service: Geo:ProjectRepositoryReplicator
    Service->>Replicator: consume(event_name, **payload)
    Replicator->>Registry: find_or_initialize_by(model_foreign_key => model_record_id)
    Registry-->>Replicator: Geo::ProjectRepositoryRegistry
    Replicator->>Registry: pending!
    Registry-->>Replicator: true

    Note over Scheduler,Replicator: Background sync path
    Scheduler->>SyncWorker: enqueues
    SyncWorker->>SyncService: new(replicable_name, model_record_id)
    SyncService->>Replicator: new(model_record_id: payload[:model_record_id])
    Replicator-->>SyncService: Geo:ProjectRepositoryReplicator
    SyncService->>Replicator: sync

    Note over Replicator, Primary: Common path
    Replicator->>FRSS: execute
    FRSS->>Replicator: registry
    Replicator->>Registry: find_or_initialize_by(model_foreign_key => model_record_id)
    Registry-->>Replicator: Geo::ProjectRepositoryRegistry
    Replicator-->>FRSS: Geo::ProjectRepositoryRegistry
    FRSS->>Registry: started!
    FRSS->>Primary: Fetch repository
    Primary->>FRSS: Repository data
    FRSS->>Registry: synced!

Exclusive lease mechanism

The exclusive lease mechanism is a critical component that prevents concurrent synchronization operations on the same repository. The lease is taken when the Geo::FrameworkRepositorySyncService executes.

Its purpose is to:

  • Prevents race conditions when multiple sync jobs attempt to update the same repository simultaneously.
  • Ensures data consistency by serializing sync operations for each repository.
  • Protects against un-necessary load coming from duplicated event workers as well as sync stuck in pending state.

The lease has a timeout of 3 hours. The format for the lease key is geo_sync_ssf_service:{replicable_name}:{model_record_id}.

Behavior during sync operations

When a sync operation begins:

  1. The service attempts to obtain an exclusive lease for the repository
  2. If the lease is already held by another process:
    • The sync exits early with a log message: "Cannot obtain an exclusive lease".
    • No sync operation is performed.
    • The job will be retried later.
  3. If the lease is obtained:
    • The sync operation proceeds.
    • The lease is released in an ensure block after sync completes.
mermaid
sequenceDiagram
participant FRSS as FrameworkRepositorySyncService
participant Redis as Redis (Lease Store)
participant Primary as Primary Site

    FRSS->>Redis: Try obtain lease
(geo_sync_ssf_service:type:id, TTL=8h)

    alt Lease obtained
        Redis-->>FRSS: Lease granted
        FRSS->>FRSS: Change state to started
        FRSS->>Primary: Fetch repository
        Primary-->>FRSS: Repository data
        FRSS->>FRSS: Change state to synced
        FRSS->>Redis: Cancel lease
    else Lease taken
        Redis-->>FRSS: Lease denied
        Note over FRSS: Sync skipped
    end

Orphaned lease keys

Orphaned lease keys can occur when the sync process is killed uncleanly (for example, SIGKILL or process crash), which means the process terminates before the ensure block can release the lease. When a lease key is orphaned, subsequent sync attempts for that repository are blocked for up to 3 hours (until the lease expires).

To mitigate this problem, you have a few options:

  • Automatic expiration: Orphaned leases automatically expire after 3 hours.
  • Monitoring: Check for an increase in geo.log for "Cannot obtain an exclusive lease" messages and geo_sync_ssf_service:{replicable_name}:{model_record_id} value of lease_key to identify potentially affected repositories.
  • Manual release: As a last resort, you can manually release orphaned leases through the Rails console (see troubleshooting)