This page contains information related to upcoming products, features, and functionality. It is important to note that the information presented is for informational purposes only. Please do not rely on this information for purchasing or planning purposes. The development, release, and timing of any products, features, or functionality may be subject to change or delay and remain at the sole discretion of GitLab Inc.
Status Authors Coach DRIs Owning Stage Created
accepted -

Cells: Topology Service

This document describes design goals and architecture of Topology Service used by Cells.

Goals

The purpose of Topology Service is to provide essential features for Cells to operate. The Topology Service will implement a limited set of functions and serve as an authoritative entity within the Cluster. There’s only a single Topology Service, that can be deployed in many regions.

  1. Technology.

    The Topology Service will be written in Go and expose API over gRPC, and REST API.

  2. Cells aware.

    The Topology Service will contain a list of all Cells. The Topology Service will monitor Cells health, and could pass this information down to Cells itself or Routing Service. Whether the Cell is healthy will be determined by various factors:

    • Watchdog: last time Cell contacted,
    • Failure rate: information gathered from the Routing Service
    • Configuration: Cells explicitly marked as orphaned
  3. Cloud first.

    The Topology Service will be deployed in Cloud, and use Cloud managed services to operate. Those services at later point could be extended with on-premise equivalents if required.

    The Topology Service will be written using a dual dialect:

    • GoogleSQL to run at scale for GitLab.com with Cloud Spanner
    • PostgreSQL for use internally and later provide on-premise compatibility.
  4. Small.

    The Topology Service due to its criticality in architecture will be limited to provide only essential functions required for cluster to operate.

Requirements

Requirement Description Priority
Configurable contains information about all Cells high
Security only authorized cells can use it high
Cloud-managed can use cloud managed services to operate high
Latency Satisfactory Latency Threshold of 20ms, 99.95% Error SLO, 99.95% Apdex SLO high
Self-managed can be eventually used by self-managed low
Regional can route requests to different regions low

Non-Goals

Those Goals are outside of the Topology Service scope as they heavily inflate the complexity:

  • The Topology Service will not provide indexing of the user-facing information for Cells. Example: CI Catalog to show data available cluster-wide will have to use another means to merge the information from all Cells.
  • The Topology Service has no knowledge of the business logic of GitLab. In theory it can work with any other web application that has the same authentication/access tokens as GitLab. However, this is subject to change as part of implementation.

Architecture

The Topology Service implements the following design guidelines:

  • Topology Service implements only a few gRPC services.
  • Some services due to backward compatibility are additionally exposed with REST API.
  • Topology Service does not perform complex processing of information.
  • Topology Service does not aggregate information from Cells.
graph TD; user((User)); http_router[HTTP Routing Service]; ssh_router[SSH Routing Service]; topology[Topology Service]; cell_1{Cell 1}; cell_N{Cell N}; spanner[Google Cloud Spanner]; user--HTTP-->http_router; user--SSH-->ssh_router; http_router--REST-->topology; http_router--HTTP-->cell_1; http_router--HTTP-->cell_N; ssh_router--gRPC-->topology; ssh_router--HTTP-->cell_1; ssh_router--HTTP-->cell_N; cell_1--gRPC-->topology; cell_N--gRPC-->topology; topology-->spanner; subgraph Cloudflare http_router; end subgraph GitLab.com Cluster ssh_router; cell_1; cell_N; topology; end subgraph Google Cloud spanner; end

Sequence Service

message LeaseSequenceRequest {
  string uuid = 3;
  string table_name = 1;
  int64 block_size = 2;
}

service SequenceService {
  rpc ValidateSequence(ValidateSequenceRequest) returns (ValidateSequenceResponse) {}
  rpc LeaseSequence(LeaseSequenceRequest) returns (LeaseSequenceRequest) {}
  rpc ReleaseSequence(ReleaseSequenceRequest) returns (ReleaseSequenceRequest) {}
}

The purpose of this service is to global allocator of the Database Sequences.

Sequence Allocation workflow

Sequences will be allocated once, at the Cell provisioning.

sequenceDiagram box participant Cell 1 participant Cell 1 DB end box participant Cell 2 participant Cell 2 DB end participant GS as GS / Sequence Service; critical Allocate sequence to projects Cell 1 ->>+ GS: LeaseSequence(projects, 1_000_000); GS -->>- Cell 1: SequenceInfo(projects, start: 10_000_000, size: 1_000_000) Cell 1 ->> Cell 1 DB: ALTER SEQUENCE projects_id_seq <br/>MINVALUE 10_000_000 <br/>MAXVALUE 10_999_999 <br/>START WITH 10_000_000 end critical Allocate sequence to projects Cell 2 ->> GS: LeaseSequence(projects, 1_000_000); GS ->> Cell 2: SequenceInfo(projects, start: 11_000_000, size: 1_000_000) Cell 2 ->> Cell 2 DB: ALTER SEQUENCE projects_id_seq <br/>MINVALUE 11_000_000 <br/>MAXVALUE 11_999_999 <br/>START WITH 11_000_000 end

Claim Service

enum ClaimType {
    Unknown = 0;
    Routes = 1;
};

message ClaimInfo {
    int64 id = 1;
    ClaimType claim_type = 2;
    string claim_value = 3;
    ...
}

service ClaimService {
    rpc CreateClaim(CreateClaimRequest) returns (CreateClaimResponse) {}
    rpc GetClaims(GetClaimsRequest) returns (GetClaimsResponse) {}
    rpc DestroyClaim(DestroyClaimRequest) returns (DestroyClaimResponse) {}
}

The purpose of this service is to provide a way to enforce uniqueness (ex. usernames, e-mails, tokens) within the cluster.

Example usage of Claim Service in Rails

class User < MainClusterwide::ApplicationRecord
  include CellsUniqueness

  cell_cluster_unique_attributes :username,
   sharding_key_object: -> { self },
   claim_type: Gitlab::Cells::ClaimType::Usernames,
 owner_type: Gitlab::Cells::OwnerType::User

  cell_cluster_unique_attributes :email,
   sharding_key_object: -> { self },
   claim_type: Gitlab::Cells::ClaimType::Emails,
 owner_type: Gitlab::Cells::OwnerType::User
end

The CellsUniqueness concern will implement cell_cluster_unique_attributes. The concern will register before and after hooks to call Topology Service gRPC endpoints for Claims within a transaction.

Classify Service

enum ClassifyType {
    Route = 1;
    Login = 2;
}

message ClassifyRequest {
    ClassifyType type = 2;
    string value = 3;
}

service ClassifyService {
    rpc Classify(ClassifyRequest) returns (ClassifyResponse) {
        option (google.api.http) = {
            get: "/v1/classify"
        };
    }
}

The purpose of this service is find owning cell of a given resource by string value. Allowing other Cells, HTTP Routing Service and SSH Routing Service to find on which Cell the project, group or organization is located.

Path Classification workflow with Classify Service

sequenceDiagram participant User1 participant HTTP Router participant TS / Classify Service participant Cell 1 participant Cell 2 User1->> HTTP Router :GET "/gitlab-org/gitlab/-/issues" Note over HTTP Router: Extract "gitlab-org/gitlab" from Path Rules HTTP Router->> TS / Classify Service: Classify(Route) "gitlab-org/gitlab" TS / Classify Service->>HTTP Router: gitlab-org/gitlab => Cell 2 HTTP Router->> Cell 2: GET "/gitlab-org/gitlab/-/issues" Cell 2->> HTTP Router: Issues Page Response HTTP Router->>User1: Issues Page Response

User login workflow with Classify Service

sequenceDiagram participant User participant HTTP Router participant Cell 1 participant Cell 2 participant TS / Classify Service User->>HTTP Router: Sign in with Username: john, password: test123 HTTP Router->>+Cell 1: Sign in with Username: john, password: test123 Note over Cell 1: User not found Cell 1->>+TS / Classify Service: Classify(Login) "john" TS / Classify Service-->>- Cell 1: "john": Cell 2 Cell 1 ->>- HTTP Router: "Cell 2". <br /> 307 Temporary Redirect HTTP Router ->> User: Set Header Cell "Cell 2". <br /> 307 Temporary Redirect User->>HTTP Router: Headers: Cell: Cell 2 <br /> Sign in with Username: john, password: test123. HTTP Router->>+Cell 2: Sign in with Username: john, password: test123. Cell 2-->>-HTTP Router: Success HTTP Router-->>User: Success

The sign-in request going to Cell 1 might at some point later be round-rubin routed to all Cells, as each Cell should be able to classify user and redirect it to correct Cell.

Metadata Service (future, implemented for Cells 1.5)

The Metadata Service is a way for Cells to distribute information cluster-wide:

  • metadata is defined by the resource_id
  • metadata can be owned by all Cells (each Cell can modify it), or owned by a Cell (only Cell can modify the metadata)
  • get request returns all metadata for a given resource_id
  • the metadata structure is owned by the application, it is strongly preferred to use protobuf to encode information due to multi-version compatibility
  • metadata owned by Cell is to avoid having to handle race conditions of updating a shared resource

The purpose of the metadata is to allow Cells to own a piece of distributed information, and allow Cells to merge the distributed information.

Example usage for different owners:

  • owned by all Cells: a user profile metadata is published representing the latest snapshot of a user publicly displayable information.
  • owner by Cell: a list of organizations to which user belongs is owned by the Cell (a distributed information), each Cell can get all metadata shared by other Cells and aggregate it.
enum MetadataOwner {
    Global = 1; // metadata is shared and any Cell can overwrite it
    Cell = 2; // metadata is scoped to Cell, and only Cell owning metadata can overwrite it
}

enum MetadataType {
    UserProfile = 1; // a single global user profile
    UserOrganizations = 2; // a metadata provided by each Cell individually
    OrganizationProfile = 3; // a single global organization information profile
}

message ResourceID {
    ResourceType type = 1;
    int64 id = 2;
};

message MetadataInfo {
    bytes data = 1;
    MetadataOwner owner = 2;
    optional CellInfo owning_cell = 3;
};

message CreateMetadataRequest {
    string uuid = 1;
    ResourceID resource_id = 2;
    MetadataOwner owner = 3;
    bytes data = 4;
};

message GetMetadataRequest {
    ResourceID resource_id = 1;
};

message GetMetadataResponse {
    repeated MetadataInfo metadata = 1;
};

service MetadataService {
    rpc CreateMetadata(CreateMetadataRequest) returns (CreateaMetadataResponse) {}
    rpc GetMetadata(GetMetadataRequest) returns (GetMetadataResponse) {}
    rpc DestroyMetadata(DestroyMetadataRequest) returns (DestroyMetadataResponse) {}
}

Example: User profile published by a Cell

sequenceDiagram participant Cell 1 participant Cell 2 participant TS as TS / Metadata Service Service; participant CS as Cloud Spanner; Cell 1 ->>+ TS: CreateMetadata(UserProfile, 100,<br/>"{username:'joerubin',displayName:'Joe Rubin'})") TS ->>- CS: INSERT INTO metadata SET (resource_id, data, cell_id)<br/>VALUES("user_profile/100",<br/>"{username:'joerubin',displayName:'Joe Rubin'})", NULL) Cell 2 ->>+ TS: CreateMetadata(UserProfile, 100,<br/>"{username:'joerubin',displayName:'Rubin is on PTO'})") TS ->>- CS: INSERT INTO metadata SET (resource_id, data, cell_id)<br/>VALUES("user_profile/100",<br/>"{username:'joerubin',displayName:'Rubin is on PTO'})", NULL) Cell 1 ->>+ TS: GetMetadata(UserProfile, 100) TS ->>- Cell 1: global => "{username:'joerubin',displayName:'Rubin is on PTO'}"

Example: Globally accessible list of Organizations to which user belongs

sequenceDiagram participant Cell 1 participant Cell 2 participant TS as TS / Metadata Service Service; participant CS as Cloud Spanner; Cell 1 ->>+ TS: CreateMetadata(UserOrganizations, 100,<br/>"[{id:200,access:'developer'}]") TS ->>- CS: INSERT INTO metadata SET (resource_id, data, cell_id)<br/>VALUES("user_organizations/100", "[{id:200,access:'developer'}]", "cell_1") Cell 2 ->>+ TS: CreateMetadata(UserOrganizations, 100,<br/>"[{id:300,access:'developer'},{id:400,access:'owner'}]") TS ->>- CS: INSERT INTO metadata SET (resource_id, data, cell_id)<br/>VALUES("user_organizations/100", "[{id:300,access:'developer'},{id:400,access:'owner'}]", "cell_2") Cell 1 ->>+ TS: GetMetadata(UserOrganizations, 100) TS ->>- Cell 1: cell_1 => "[{id:200,access:'developer'}]", "cell_1"<br/>cell_2 => "[{id:300,access:'developer'},{id:400,access:'owner'}]"

Reasons

The original Cells 1.0 described Primary Cell API, this changes this decision to implement Topology Service for the following reasons:

  1. Provide stable and well described set of cluster-wide services that can be used by various services (HTTP Routing Service, SSH Routing Service, each Cell).
  2. As part of Cells 1.0 PoC we discovered that we need to provide robust classification API to support more workflows than anticipated. We need to classify various resources (username for login, projects for SSH routing, etc.) to route to correct Cell. This would put a lot of dependency on resilience of the First Cell.
  3. It is our desire long-term to have Topology Service for passing information across Cells. This does a first step towards long-term direction, allowing us to much easier perform additional functions.

Spanner

Spanner will be a new data store introduced into the GitLab Stack, the reasons we are going with Spanner are:

  1. It supports Multi-Regional read-write access with a lot less operations when compared to PostgreSQL helping with out regional DR
  2. The data is read heavy not write heavy.
  3. Spanner provides 99.999% SLA when using Multi-Regional deployments.
  4. Provides consistency whilst still being globally distributed.
  5. Shards/Splits are handled for us.

The cons of using Spanners are:

  1. Vendor lock-in, our data will be hosted in a proprietary data.
    • How to prevent this: Topology Service will use generic SQL.
  2. Not self-managed friendly, when we want to have Topology Service available for self-managed customers.
    • How to prevent this: Spanner supports PostgreSQL dialect.
  3. Brand new data store we need to learn to operate/develop with.

GoogleSQL vs PostgreSQL dialects

Spanner supports two dialects one called GoogleSQL and PostgreSQL. The dialect doesn’t change the performance characteristics of Spanner, it’s mostly how the Database schemas and queries are written. Choosing a dialect is a one-way door decision, to change the dialect we’ll have to go through a data migration process.

We will use the GoogleSQL dialect for the Topology Service, and go-sql-spanner to connect to it, because:

  1. Using Go’s standard library database/sql will allow us to swap implementations which is needed to support self-managed.
  2. GoogleSQL data types are narrower and don’t allow to make mistakes for example choosing int32 because it only supports int64.
  3. New features seem to be released on GoogleSQL first, for example, https://cloud.google.com/spanner/docs/ml. We don’t need this feature specifically, but it shows that new features support GoogleSQL first.
  4. A more clear split in the code when we are using Google Spanner or native PostgreSQL, and won’t hit edge cases.

Citations:

  1. Google (n.d.). PostgreSQL interface for Spanner. Google Cloud. Retrieved April 1, 2024, from https://cloud.google.com/spanner/docs/postgresql-interface
  2. Google (n.d.). Dialect parity between GoogleSQL and PostgreSQL. Google Cloud. Retrieved April 1, 2024, from https://cloud.google.com/spanner/docs/reference/dialect-differences

Multi-Regional

Running Multi-Regional read-write is one of the biggest selling points of Spanner. When provisioning an instance you can choose single Region or Multi-region. After provisioning you can move an instance whilst is running but this is a manual process that requires assistance from GCP.

We will provision a Multi-Regional Cloud Spanner instance because:

  1. Won’t require migration to Multi-Regional in the future.
  2. Have Multi Regional on day 0 which cuts the scope of multi region deployments at GitLab.

This will however increase the cost considerably, using public facing numbers from GCP:

  1. Regional: $1,716
  2. Multi Regional: $9,085

Citations:

  1. Google (n.d.). Regional and multi-region configurations. Google Cloud. Retrieved April 1, 2024, from https://cloud.google.com/spanner/docs/instance-configurations
  2. Google (n.d.). FeedbackReplication. Google Cloud. Retrieved April 1, 2024, from https://cloud.google.com/spanner/docs/replication

Architecture of multi-regional deployment of Topology Service

graph TD; user_eu((User in EU)); user_us((User in US)); gitlab_com_gcp_load_balancer[GitLab.com GCP Load Balancer]; topology_service_gcp_load_balancer[Topology Service GCP Load Balancer]; http_router[HTTP Routing Service]; topology_service_eu[Topology Service in EU]; topology_service_us[Topology Service in US]; cell_us{Cell US}; cell_eu{Cell EU}; spanner[Google Cloud Spanner]; subgraph Cloudflare http_router; end subgraph Google Cloud subgraph Multi-regional Load Balancers / AnyCast DNS gitlab_com_gcp_load_balancer; topology_service_gcp_load_balancer; end subgraph Europe topology_service_eu; cell_eu; end subgraph US topology_service_us; cell_us; end subgraph Multi-regional Cloud Spanner spanner; end end user_eu--HTTPS-->http_router; user_us--HTTPS-->http_router; http_router--REST/mTLS-->topology_service_gcp_load_balancer; http_router--HTTPS-->gitlab_com_gcp_load_balancer; gitlab_com_gcp_load_balancer--HTTPS-->cell_eu; gitlab_com_gcp_load_balancer--HTTPS-->cell_us; topology_service_gcp_load_balancer--HTTPS-->topology_service_eu; topology_service_gcp_load_balancer--HTTPS-->topology_service_us; cell_eu--gRPC/mTLS-->topology_service_eu; cell_us--gRPC/mTLS-->topology_service_us; topology_service_eu--gRPC-->spanner; topology_service_us--gRPC-->spanner;

Performance

We haven’t run any benchmarks ourselves because we don’t have a full schema designed. However looking at the performance documentation, both the read and write throughput of a Spanner instance scale linearly as you add more compute capacity.

Alternatives

  1. PostgreSQL: Having a multi-regional deployment requires a lot of operations.
  2. ClickHouse: It’s an OLAP database not an OLTP.
  3. Elasticsearch: Search and analytics document store.

Disaster Recovery

We must stay in our Disaster Recovery targets for the Topology Service. Ideally, we need smaller windows for recovery because this service is in the critical path.

The service is stateless, which should be much easier to deploy to multiple regions using runway. The state is stored in Cloud Spanner, the state consists of database sequences, projects, username, and anything we need to keep global uniqueness in the application. This data is critical, and if we loose this data we won’t be able to route requests accordingly or keep global uniqueness to have the ability to move data between cells in the future. For this reason we are going to set up Multi-Regional read-write deployment for Cloud Spanner so even if a region goes down, we can still read-write to the state.

Cloud Spanner provides 3 ways of recovery:

  1. Backups: A backup of a database inside of the instance. You can copy the backup to another instance but this requires an instance of the same size of storage which can 2x the costs. One concern with using backups is if the instance gets deleted by mistake (even with deletion protection)
  2. Import/Export: Export the database as a medium priority task inside of Google Cloud Storage.
  3. Point-in-time recovery: Version retention period up to 7 days, this can help with recovery of a portion of the database or create a backup/restore from a specific time to recover the full database. Increasing the retention period does have performance implications

As you can see all these options only handle the data side, not the storage/compute side, this is because storage/compute is managed for us. This means our Disaster Recovery plan should only account for potential logical application errors where it deletes/logically corrupts the data.

These require testing, and validation but to have all the protection we can have:

  1. Import/Export: Daily
  2. Backups: Hourly
  3. Point-in-time recovery: Retention period of 2 days.

On top of those backups we’ll also make sure:

  1. We have database deletion protection on.
  2. Make sure the application user doesn’t have spanner.database.drop IAM.
  3. The Import/Export bucket will have bucket lock configured to prevent deletion.

Citations:

  1. Google (n.d.). Choose between backup and restore or import and export. Google Cloud. Retrieved April 2, 2024, from https://cloud.google.com/spanner/docs/backup/choose-backup-import

FAQ

  1. Does Topology Service implement all services for Cells 1.0?

    No, for Cells 1.0 Topology Service will implement ClaimService and ClassifyService only. Due to complexity the SequenceService will be implemented by the existing Cell of the cluster. The reason is to reduce complexity of deployment: as we would only add a function to the first cell. We would add new feature, but we would not change “First Cell” behavior. At later point the Topology Service will take over that function from First Cell.

  2. How we will push all existing claims from “First Cell” into Topology Service?

    We would add rake gitlab:cells:claims:create task. Then we would configure First Cell to use Topology Service, and execute the Rake task. That way First Cell would claim all new records via Topology Service, and concurrently we would copy data over.

  3. How and where the Topology Service will be deployed?

    We will use Runway, and configure Topology Service to use Spanner for data storage.

  4. How Topology Service handle regions?

    We anticipate that Spanner will provide regional database support, with high-performance read access. In such case the Topology Service will be run in each region connected to the same multi-write database. We anticipate one Topology Service deployment per-region that might scale up to desired number of replicas / pods based on the load.

  5. Will Topology Service information be encrypted at runtime?

    This is yet to be defined. However, Topology Service could encrypt customer sensitive information allowing for the information to be decrypted by the Cell that did create that entry. Cells could transfer encrypted/hashed information to Topology Service making the Topology Service to only store metadata without the knowledge of information.

  6. Will Topology Service data to be encrypted at rest?

    This is yet to be defined. Data is encrypted during transport (TLS/gRPC and HTTPS) and at rest by Spanner.

Topology Service discussions