GitLab Cells Development Guidelines

For background of GitLab Cells, refer to the design document.

Choose either the `gitlab_main_cell` or `gitlab_main_clusterwide` schema

Depending on the use case, your feature may be cell-local or clusterwide and hence the tables used for the feature should also use the appropriate schema.

When you choose the appropriate schema for tables, consider the following guidelines as part of the Cells architecture:

Default to gitlab_main_cell: We expect most tables to be assigned to the gitlab_main_cell schema by default. Choose this schema if the data in the table is related to projects or namespaces.
Consult with the Tenant Scale group: If you believe that the gitlab_main_clusterwide schema is more suitable for a table, seek approval from the Tenant Scale group. This is crucial because it has scaling implications and may require reconsideration of the schema choice.

To understand how existing tables are classified, you can use this dashboard.

After a schema has been assigned, the merge request pipeline might fail due to one or more of the following reasons, which can be rectified by following the linked guidelines:

Defining a sharding key for all cell-local tables

All tables with the following gitlab_schema are considered "cell-local":

gitlab_main_cell
gitlab_ci

All newly created cell-local tables are required to have a sharding_key defined in the corresponding db/docs/ file for that table.

The purpose of the sharding key is documented in the Organization isolation blueprint, but in short this column is used to provide a standard way of determining which Organization owns a particular row in the database. The column will be used in the future to enforce constraints on data not cross Organization boundaries. It will also be used in the future to provide a uniform way to migrate data between Cells.

The actual name of the foreign key can be anything but it must reference a row in projects or groups. The chosen sharding_key column must be non-nullable.

Setting multiple sharding_key, with nullable columns are also allowed, provided that the table has a check constraint that correctly ensures at least one of the keys must be non-nullable for a row in the table. See NOT NULL constraints for multiple columns for instructions on creating these constraints.

The following are examples of valid sharding keys:

The table entries belong to a project only:
```
sharding_key:
  project_id: projects
```
The table entries belong to a project and the foreign key is target_project_id:
```
sharding_key:
  target_project_id: projects
```
The table entries belong to a namespace/group only:
```
sharding_key:
  namespace_id: namespaces
```
The table entries belong to a namespace/group only and the foreign key is group_id:
```
sharding_key:
  group_id: namespaces
```

The table entries belong to a namespace or a project:

sharding_key:
  project_id: projects
  namespace_id: namespaces

The sharding key must be immutable

The choice of a sharding_key should always be immutable. Therefore, if your feature requires a user experience which allows data to be moved between projects or groups/namespaces, then you may need to redesign the move feature to create new rows. An example of this can be seen in the move an issue feature. This feature does not actually change the project_id column for an existing issues row but instead creates a new issues row and creates a link in the database from the original issues row. If there is a particularly challenging existing feature that needs to allow moving data you will need to reach out to the Tenant Scale team early on to discuss options for how to manage the sharding key.

Using the same sharding key for projects and namespaces

Developers may also choose to use namespace_id only for tables that can belong to a project where the feature used by the table is being developed following the Consolidating Groups and Projects blueprint. In that case the namespace_id would need to be the ID of the ProjectNamespace and not the group that the namespace belongs to.

Define a `desired_sharding_key` to automatically backfill a `sharding_key`

We need to backfill a sharding_key to hundreds of tables that do not have one. This process will involve creating a merge request like https://gitlab.com/gitlab-org/gitlab/-/merge_requests/136800 to add the new column, backfill the data from a related table in the database, and then create subsequent merge requests to add indexes, foreign keys and not-null constraints.

In order to minimize the amount of repetitive effort for developers we've introduced a concise declarative way to describe how to backfill the sharding_key for this specific table. This content will later be used in automation to create all the necessary merge requests.

An example of the desired_sharding_key was added in https://gitlab.com/gitlab-org/gitlab/-/merge_requests/139336 and it looks like:

--- # db/docs/security_findings.yml
table_name: security_findings
classes:
- Security::Finding

...

desired_sharding_key:
  project_id:
    references: projects
    backfill_via:
      parent:
        foreign_key: scanner_id
        table: vulnerability_scanners
        sharding_key: project_id
        belongs_to: scanner

To understand best how this YAML data will be used you can map it onto the merge request we created manually in GraphQL https://gitlab.com/gitlab-org/gitlab/-/merge_requests/136800. The idea will be to automatically create this. The content of the YAML specifies the parent table and its sharding_key to backfill from in the batched background migration. It also specifies a belongs_to relation which will be added to the model to automatically populate the sharding_key in the before_save.

Define a `desired_sharding_key` when the parent table also has one

By default, a desired_sharding_key configuration will validate that the chosen sharding_key exists on the parent table. However, if the parent table also has a desired_sharding_key configuration and is itself waiting to be backfilled, you need to include the awaiting_backfill_on_parent field. For example:

desired_sharding_key:
  project_id:
    references: projects
    backfill_via:
      parent:
        foreign_key: package_file_id
        table: packages_package_files
        sharding_key: project_id
        belongs_to: package_file
    awaiting_backfill_on_parent: true

There are likely edge cases where this desired_sharding_key structure is not suitable for backfilling a sharding_key. In such cases the team owning the table will need to create the necessary merge requests to add the sharding_key manually.

Exempting certain tables from having sharding keys

Certain tables can be exempted from having sharding keys by adding

exempt_from_sharding: true

to the table's database dictionary file. This can be used for:

JiHu specific tables, since they do not have any data on the .com database. !145905
tables that are marked to be dropped soon, like operations_feature_flag_scopes. !147541
tables that mandatorily need to be present per cell to support a cell's operations, have unique data per cell, but cannot have a sharding key defined. For example, zoekt_nodes.

When tables are exempted from sharding key requirements, they also do not show up in our progress dashboard.

Exempted tables must not have foreign key, or loose foreign key references, as this may cause the target cell's database to have foreign key violations when data is moved. See #471182 for examples and possible solutions.