Status | Authors | Coach | DRIs | Owning Stage | Created |
---|---|---|---|---|---|
proposed |
@pguinoiseau
@jcstephenson
@ayeung
|
@sxuereb
@andrewn
|
@product-manager
@sabrams
| devops data stores | 2024-03-22 |
Cell Architecture and Tooling
Summary
This blueprint lays out how cells will be architected, which components will be managed as cell-local or globally, and how a mix of existing and new tooling will be used to provision new cells.
The core of each cell is essentially a Dedicated instance that will be configured to form part of the GitLab.com cluster of cells. This lets us reuse the existing Dedicated tooling (specifically Instrumentor and Amp) for provisioning and configuration of the initial cells. On the other hand, as a consequence of being reliant on Dedicated, many decisions - specifically how a cell is internally architected - are already made for us.
The Foundations team will develop new tooling (provisionally named cellctl
) which will be used by callers wanting to interact with cells. This will abstract away the need for the callers to have any knowledge of how a cell is built. It does this by presenting a cell model schema - essentially a reduced version of the Dedicated tenant model schema, with additional cells-related fields added - to clients, which will be translated into a Dedicated tenant model schema that can be used by Dedicated tooling. cellctl
, using Terraform, will also be responsible for ensuring that GCP resources are appropriately located within the resource hierarchy. cells-tissue
will continue to be the source of truth for cell definitions.
For the initial stages of the cells rollout, we don’t believe it is necessary to have a separate provisioning process for cell-local infrastructure, but that option remains on the table in the event of the requirements of cells and Dedicated diverging too greatly to consolidate cleanly within Instrumentor.
Motivation
It has been apparent for some time that we have hit a scalability ceiling on GitLab.com. Noisy neighbor issues abound, and the underlying infrastructure demands an ever-increasing amount of toil from SREs due to the inevitable buildup of unaddressed technical debt under the hood.
To alleviate these problems, we aim to replace the current monolithic architecture with a resilient, horizontally scalable multi-tenanted architecture (“Cells”) that provides a high level of isolation between tenants while still providing a largely unchanged experience to the user. This change should also make implementation of customer requirements such as data residency possible.
In doing this, we’re also trying to achieve some degree of homogeneity in how GitLab is deployed. Currently, GitLab.com involves myriad different tools to provision: Helm, Chef, Tanka, and Terraform among them. Cells will be using the tooling used to provision GitLab Dedicated tenants, meaning that any improvements to the process for cells can be applied to Dedicated customers too, or vice versa.
Goals
This blueprint will:
- Describe the architecture of each cell
- Enumerate reasons for why certain architectural decisions were made
- Broadly address how existing and new tooling will be used for cell provisioning
Non-Goals
The following is out of scope:
- Details of cell provisioning or configuration management - these will be addressed in separate blueprints (for brevity)
- Process changes (e.g. which team is responsible for what task)
- Low-level implementation details of tooling
- Migration from monolith to cells
- How to run monolith and cells concurrently
Proposal
- Each cell is effectively its own GitLab Dedicated instance that can interface seamlessly with the rest of GitLab.com, yet is limited to using resources available to that cell only. Users on a cell will not perceive the isolation of the underlying resources
- Dedicated tooling (Instrumentor and Amp) will be used to provision new cells
- To reduce potential future disruption, common Terraform modules will be split out of the current Instrumentor repo and vendored instead. Foundations will share ownership of the modules with the Dedicated team
- A separate Amp environment will be setup for cells use
-
cellctl
is a tool that will be developed by Foundations to translate a cell model schema into a Dedicated tenant model schema, and calling Instrumentor to provision and configure the tenant. Users ofcellctl
need not be aware of the infrastructure underlying a cell-
cellctl
will also be responsible for ensuring the resource hierarchy in GCP matches the defined hierarchy incells-tissue
-
Design and implementation details
Architecture
Components
Each cell contains a Cloud Native Hybrid deployment of GitLab, sized appropriately for its expected workload. This is provided by the GitLab Environment Toolkit (GET) via the Dedicated tool Instrumentor which acts as an orchestration layer on top of GET.
Cell definitions (i.e. size of each cell, which components it includes, etc.) will be stored in the cells-tissue
repository, which already contains definitions for a number of cells and rings.
Service | Runs on |
---|---|
Webservice | Kubernetes |
Sidekiq | Kubernetes |
Supporting services | Kubernetes |
Gitaly | Google Compute Engine VM |
PostgreSQL | CloudSQL |
Redis | Memorystore |
Object storage | Google Cloud Storage |
Per-cell observability stack | Kubernetes |
Inter-cell networking | Private Service Connect |
ClickHouse | ClickHouse Cloud |
Auxiliary services with a global context - that is, services handling data not specific to any individual cell - will have their lifecycle managed completely independently of cells. Such services include:
- Hashicorp Vault (for storage and management of secrets used in CI, and shared infrastructure secrets such as those used by the observability stack)
- We will be using the existing production Vault instance which is configured using
config-mgmt
and deployed ingitlab-helmfiles
.
- We will be using the existing production Vault instance which is configured using
- Camoproxy (for proxying assets)
- We will be using the existing Camoproxy instance deployed in
gitlab-helmfiles
.
- We will be using the existing Camoproxy instance deployed in
Naming conventions for cells
The ID field can only be up to 30 characters long due to restrictions imposed by the Dedicated tenant model and GCP project naming. This limits us in what kind of metadata we can include in the cell name: for example, we can’t include the region name.
With this in mind, cells will use a ULID, truncated to 18 characters, as its tenant ID. For example, a single cell’s ID might be 01HW6TQSAMAVPGR3N8
. Per the ULID spec, the ID consists of:
ttttttttttrrrrrrrr
where
t is Timestamp (10 characters)
r is Randomness (8 characters)
A prettier name prefixed with cell-
will be used for situations where human readability is preferred, e.g. in the GCP project name. For example, if a cell has the ID of 01HW6TQSAMAVPGR3N8
, its prettified name will be cell-01HW6TQSAMAVPGR3N8
.
The length of the ID allows for an additional 7 characters which can be used to name additional GCP projects (should they be needed) related to a single cell. For example, if the cell needs additional projects for Runner or Gitaly resources, they can be named like so, while staying within the 30 character limit:
-
cell-01HW6TQSAMAVPGR3N8-rnr123
orcell-01HW6TQSAMAVPGR3N8-runner
-
cell-01HW6TQSAMAVPGR3N8-git594
orcell-01HW6TQSAMAVPGR3N8-gitaly
See this issue for discussion.
Relationship to GCP projects
Each cell will have its own project in GCP to provide greater isolation between cells and simplify management. This is also in accordance with the existing Dedicated tooling, which assumes one GCP project per tenant.
See this issue for discussion, and this ADR.
Networking
Each cell will have its own VPC, with one subnet per region.
Having a single VPC per cell will make provisioning and management of a cell easier as there will be no need to worry about IP address space overlap issues, as well as make cells secure by default as they will be fully isolated.
Cells will use Private Service Connect to communicate with other cells if required. The definitions specifying how cells are connected will be kept alongside cell definitions that will be used by cellctl
to provision the needed resources.
See this issue for discussion, and this ADR.
GCP organization
A new GCP organization (provisionally named GitLab Cells
) for the sole purpose of hosting cells will be created. While it will use the same centralized billing account as the existing gitlab.com
organization, all resources and permissions under the new org will be controlled authoritatively using Terraform and access will be locked down to only team members who need it. Admin access would be granted in break-glass situations only, and fully auditable.
This was deemed preferable to thoroughly auditing and consolidating the existing organization, which is relied upon by many different teams (so we run the risk of breaking something business-critical) and is only partially managed by Terraform. See this issue for discussion.
Cloud resource hierarchy
The cells tooling (cellctl
) will be responsible for creating and maintaining a cloud resource hierarchy under the Gitlab Cells
organization. This hierarchy should reflect the filesystem hierarchy of the definitions in the cells-tissue
repository.
In this hierarchy, every environment will have its own folder, which in turn contains an Ops
folder, and a Cells
folder.
Each cell will have its own folder, named for its (prettified) ID, in the Cells
folder. This gives us the option of adding new projects for cell-specific components (such as runners or additional Gitaly nodes) to the folder, thus grouping them with the cell concerned, in the future.
The Ops
folder will contain the Amp environment GCP project.
Number of GKE clusters serving a cell
Each cell will run on one GKE cluster. Routing to and between cells will be simpler; the Helm charts do not support deploying one instance of the application across multiple clusters, while GET does not support managing multiple instances of the application in one cluster.
See this issue for discussion, and this ADR.
Secrets management for cells
We will be using both Google Secret Manager and Hashicorp Vault for secrets management.
Secret Manager will store secrets that are specific to a single cell and thus scoped to a specific GCP project.
Hashicorp Vault will store secrets that are shared across multiple (or all) cells. Such secrets may include credentials for the observability stack.
All Kubernetes secrets will be injected using external-secrets
and not provisioned directly.
See this issue for discussion.
Access to cells
This is still being discussed. As much of the underlying infrastructure of a cell is derived from a Dedicated on GCP instance, we expect that any procedure for accessing a cell would be shared (or broadly similar) between the two products.
Observability
In general, the lifecycle of observability components for cells will be owned by the Scalability:Observability
team.
By default, each Dedicated tenant is provisioned with a fully functional Prometheus/Grafana stack. Cells will reuse this stack, with the intention of aggregating metrics so that queries can be run over multiple cells. More information can be found here.
See this issue for discussion.
Tooling
Cells 1.0
In order to achieve the Cells 1.0 milestone in the minimum time possible, Foundations will be doing the following:
- Extracting Terraform modules required for a Dedicated GCP instance out of the Instrumentor monorepo
- Each module will be split into its own repo and vendored in as an external Terraform module in Instrumentor
- Thereafter, both the Dedicated team and Foundations will be responsible for maintaining these modules
- Adding reference architectures for cells and their related Jsonnet files to Instrumentor so that appropriate variables can be templated in
- Developing a new tool, provisionally called
cellctl
, to orchestrate cells- For Cells 1.0,
cellctl
will simply be a thin translation wrapper, transforming its own cell tenant model into an opinionated instance of the tenant model schema used by Dedicated - The cells model will be an opinionated derivative of the tenant model schema used by Dedicated and include additional cell-specific fields for use by orchestrators of
cellctl
.
- For Cells 1.0,
- Set up a new Amp environment in GCP specifically for provisioning cells
Other tooling, such as ringctl
, should be able to interface with cellctl
to perform operations on one or many cells without needing to be aware of any specifics of how cells are implemented or orchestrated.
Cell tenant model
The tenant model of a cell will follow this structure, split into 2 core fields:
- a metadata struct which contains information used to identify the cell instance (name and labels), in addition to tooling specific information such as which template to use to render the cell definition. Labels can be used to store information relevant to other tools, such as the ring that the cell belongs to.
- a spec struct that contains information about the definition of a cell instance, including the location in which to create the instance.
// Model entry point
type Cell struct {
metadata CellMetadata
spec CellSpec
}
// Contains fields specific to Cells
type CellMetadata struct {
// The identifier of the cell
name string
// Optional template specifier
//
// If this is not set then the template must be provided
// externally.
template *string `json:",omitempty"`
// Additional annotations
//
// To be used for defining features and passing additional
// context through to the cell template.
annotations map[string]string `json:",omitempty"`
// Additional labels
// Used to add human understandable context to cell instances.
// Can be used to perform searches across cell instances.
labels map[string]string `json:",omitempty"`
}
// Contains fields related to the definition of the cell through the dedicated tooling
type CellSpec struct {
// Primary GCP project ID
project_id string
// The region we want this cell to be provisioned within
region string
// Reference architecture
reference_architecture string
// Instrumentor version
instrumentor_version *string `json:",omitempty"`
// GitLab version
gitlab_version *string `json:",omitempty"`
// Prerelease version
prerelease_version *string `json:",omitempty"`
}
cellctl
processes these cell instance definitions from YAML files, and translates them through a mapping layer of jsonnet templates to output a tenant model for provisioning through Amp and Instrumentor.
These cell definition files can also be processed by other tooling at the same plane as cellctl
to create additional infrastructure with a “global” view of all cell instances.
Beyond Cells 1.0
As cells are gradually built out, we may find that the requirements for Dedicated tenants and cells will diverge, and attempting to consolidate this into Instrumentor may result in unacceptable complexity. This may necessitate an additional layer of resources provisioned and managed by cells-specific tooling, which would be developed and maintained by Foundations.
Where should something be implemented for cells?
There are four places containing resource definitions for components making up a single cell. Each builds on top of the last:
As we build out cells, there will inevitably be questions about the best place to add something. The following chart should help guide the decision: