Skip to main content
Version: Next

Extending the model

The Backstage catalog entity data model is based on the Kubernetes objects format, and borrows a lot of its semantics as well. This page describes those semantics at a higher level and how to extend them to fit your organization.

Backstage comes with a number of catalog concepts out of the box:

  • There are a number of builtin versioned kinds, such as Component, User etc. These encapsulate the high level concept of an entity, and define the schema for its entity definition data.
  • An entity has both a metadata object and a spec object at the root.
  • Each kind may or may not have a type. For example, there are several well known types of component, such as service and website. These clarify the more detailed nature of the entity, and may affect what features are exposed in the interface.
  • Entities may have a number of annotations on them. These can be added either by humans into the descriptor files, or added by automated processes when the entity is ingested into the catalog.
  • Entities may have a number of labels on them.
  • Entities may have a number of relations, expressing how they relate to each other in different ways.

We'll list different possibilities for extending this below.

Adding a New apiVersion of an Existing Kind​

Example intents:

"I want to evolve this core kind, tweaking the semantics a bit so I will bump the apiVersion a step"

"This core kind is a decent fit but we want to evolve it at will so we'll move it to our own company's apiVersion space and use that instead of backstage.io."

The backstage.io apiVersion space is reserved for use by the Backstage maintainers. Please do not change or add versions within that space.

If you add an apiVersion space of your own, you are effectively branching out from the underlying kind and making your own. An entity kind is identified by the apiVersion + kind pair, so even though the resulting entity may be similar to the core one, there will be no guarantees that plugins will be able to parse or understand its data. See below about adding a new kind.

Adding a New Kind​

Example intents:

"The kinds that come with the package are lacking. I want to model this other thing that is a poor fit for either of the builtins."

"This core kind is a decent fit but we want to evolve it at will so we'll move it to our own company's apiVersion space and use that instead of backstage.io."

A kind is an overarching family, or an idea if you will, of entities that also share a schema. Backstage comes with a number of builtin ones that we believe are useful for a large variety of needs that one may want to model in Backstage. The primary ambition is to map things to these kinds, but sometimes you may want or need to extend beyond them.

Introducing a new apiVersion is basically the same as adding a new kind. Bear in mind that most plugins will be compiled against the builtin @backstage/catalog-model package and have expectations that kinds align with that.

The catalog backend itself, from a storage and API standpoint, does not care about the kind of entities it stores. Extending with new kinds is mainly a matter of permitting them to pass validation when building the backend catalog using the CatalogBuilder, and then to make plugins be able to understand the new kind.

For the consuming side, it's a different story. Adding a kind has a very large impact. The very foundation of Backstage is to attach behavior and views and functionality to entities that we ascribe some meaning to. There will be many places where code checks if (kind === 'X') for some hard coded X, and casts it to a concrete type that it imported from a package such as @backstage/catalog-model.

If you want to model something that doesn't feel like a fit for either of the builtin kinds, feel free to reach out to the Backstage maintainers to discuss how to best proceed.

If you end up adding that new kind, you must namespace its apiVersion accordingly with a prefix that makes sense, typically based on your organization name - e.g. my-company.net/v1. Also do pick a new kind identifier that does not collide with the builtin kinds.

Adding a New Type of an Existing Kind​

Example intents:

"This is clearly a component, but it's of a type that doesn't quite fit with the ones I've seen before."

"We don't call our teams "team", can't we put "flock" as the group type?"

Some entity kinds have a type field in its spec. This is where an organization are free to express the variety of entities within a kind. This field is expected to follow some taxonomy that makes sense for yourself. The chosen value may affect what operations and views are enabled in Backstage for that entity. Inside Spotify our model has grown significantly over the years, and our component types now include ML models, apps, data pipelines and many more.

It might be tempting to put software that doesn't fit into any of the existing types into an Other catch-all type. There are a few reasons why we advise against this; firstly, we have found that it is preferred to match the conceptual model that your engineers have when describing your software. Secondly, Backstage helps your engineers manage their software by integrating the infrastructure tooling through plugins. Different plugins are used for managing different types of components.

For example, the Lighthouse plugin only makes sense for Websites. The more specific you can be in how you model your software, the easier it is to provide plugins that are contextual.

Adding a new type takes relatively little effort and carries little risk. Any type value is accepted by the catalog backend, but plugins may have to be updated if you want particular behaviors attached to that new type.

Changing the Validation Rules for The Entity Envelope or Metadata Fields​

Example intents:

"We want to import our old catalog but the default set of allowed characters for a metadata.name are too strict."

"I want to change the rules for annotations so that I'm allowed to store any data in annotation values, not just strings."

After pieces of raw entity data have been read from a location, they are passed through a field format validation step. This ensures that the types and syntax of the base envelope and metadata make sense - in short, things that aren't entity-kind-specific. Some or all of these validators can be replaced when building the backend using the catalog's dedicated catalogModelExtensionPoint (or directly on the CatalogBuilder if you are still using the old backend system).

The risk and impact of this type of extension varies, based on what it is that you want to do. For example, extending the valid character set for kinds, namespaces and names can be fairly harmless, with a few notable exceptions - there is code that expects these to never ever contain a colon or slash, for example, and introducing URL-unsafe characters risks breaking plugins that aren't careful about encoding arguments. Supporting non-strings in annotations may be possible but has not yet been tried out in the real world - there is likely to be some level of plugin breakage that can be hard to predict.

You must also be careful about not making the rules more strict than they used to be after populating the catalog with data. This risks making previously valid entities start having processing errors and fail to update.

Before making this kind of extension, we recommend that you contact the Backstage maintainers or a support partner to discuss your use case.

This is an example of relaxing the format rules of the metadata.name field:

import { createBackend } from '@backstage/backend-defaults';
import { createBackendModule } from '@backstage/backend-plugin-api';
import { catalogModelExtensionPoint } from '@backstage/plugin-catalog-node/alpha';

const myCatalogCustomizations = createBackendModule({
  pluginId: 'catalog',
  moduleId: 'catalog-customization',
  register(reg) {
    reg.registerInit({
      deps: {
        catalogModel: catalogModelExtensionPoint,
      },
      async init({ catalogModel }) {
        catalogModel.setFieldValidators({
          // This is only one of many methods that you can pass into
          // setFieldValidators; your editor of choice should help you
          // find the others. The length checks and regexp inside are
          // just examples and can be adjusted as needed, but take care
          // to test your changes thoroughly to ensure that you get
          // them right.
          isValidEntityName(value) {
            return (
              typeof value === 'string' &&
              value.length >= 1 &&
              value.length <= 63 &&
              /^[A-Za-z0-9@+_.-]+$/.test(value)
            );
          },
        });
      },
    });
  },
});

const backend = createBackend();
// ... add other backend features and the catalog backend itself here ...
backend.add(myCatalogCustomizations);
backend.start();

Changing the Validation Rules for Core Entity Fields​

Example intent:

"I don't like that the owner is mandatory. I'd like it to be optional."

After reading and policy-checked entity data from a location, it is sent through the processor chain looking for processors that implement the validateEntityKind step, to see that the data is of a known kind and abides by its schema. There is a builtin processor that implements this for all known core kinds and matches the data against their fixed validation schema. This processor can be replaced when building the backend catalog using the CatalogBuilder, with a processor of your own that validates the data differently. This replacement processor must have a name that matches the builtin processor, BuiltinKindsEntityProcessor.

This type of extension is high risk, and may have high impact across the ecosystem depending on the type of change that is made. It is therefore not recommended in normal cases. There will be a large number of plugins and processors - and even the core itself - that make assumptions about the shape of the data and import the typescript data type from the @backstage/catalog-model package.

Adding New Fields to the Metadata Object​

Example intent:

"Our entities have this auxiliary property that I would like to express for several entity kinds and it doesn't really fit as a spec field."

The metadata object is currently left open for extension. Any unknown fields found in the metadata will just be stored verbatim in the catalog. However we want to caution against extending the metadata excessively. Firstly, you run the risk of colliding with future extensions to the model. Secondly, it is common that this type of extension lives more comfortably elsewhere - primarily in the metadata labels or annotations, but sometimes you even may want to make a new component type or similar instead.

There are some situations where metadata can be the right place. If you feel that you have run into such a case and that it would apply to others, do feel free to contact the Backstage maintainers or a support partner to discuss your use case. Maybe we can extend the core model to benefit both you and others.

Adding New Fields to the Spec Object of an Existing Kind​

Example intent:

"The builtin Component kind is fine but we want to add an additional field to the spec for describing whether it's in prod or staging."

A kind's schema validation typically doesn't forbid "unknown" fields in an entity spec, and the catalog will happily store whatever is in it. So doing this will usually work from the catalog's point of view.

Adding fields like this is subject to the same risks as mentioned about metadata extensions above. Firstly, you run the risk of colliding with future extensions to the model. Secondly, it is common that this type of extension lives more comfortably elsewhere - primarily in the metadata labels or annotations, but sometimes you even may want to make a new component type or similar instead.

There are some situations where the spec can be the right place. If you feel that you have run into such a case and that it would apply to others, do feel free to contact the Backstage maintainers or a support partner to discuss your use case. Maybe we can extend the core model to benefit both you and others.

Adding a New Annotation​

Example intents:

"Our custom made build system has the concept of a named pipeline-set, and we want to associate individual components with their corresponding pipeline-sets so we can show their build status."

"We have an alerting system that automatically monitors service health, and there's this integration key that binds the service to an alerts pool. We want to be able to show the ongoing alerts for our services in Backstage so it'd be nice to attach that integration key to the entity somehow."

Annotations are mainly intended to be consumed by plugins, for feature detection or linking into external systems. Sometimes they are added by humans, but often they are automatically generated at ingestion time by processors. There is a set of well-known annotations, but you are free to add additional ones. This carries no risk or impact to other systems as long as you abide by the following naming rules.

  • The backstage.io annotation prefix is reserved for use by the Backstage maintainers. Reach out to us if you feel that you would like to make an addition to that prefix.
  • Annotations that pertain to a well known third party system should ideally be prefixed with a domain, in a way that makes sense to a reader and connects it clearly to the system (or the maker of the system). For example, you might use a pagerduty.com prefix for pagerduty related annotations, but maybe not ldap.com for LDAP annotations since it's not directly affiliated with or owned by an LDAP foundation/company/similar.
  • Annotations that have no prefix at all, are considered local to your Backstage instance and can be used freely as such, but you should not make use of them outside of your organization. For example, if you were to open source a plugin that generates or consumes annotations, then those annotations must be properly prefixed with your company domain or a domain that pertains to the annotation at hand.

Adding a New Label​

Example intents:

"Our process reaping system wants to periodically scrape for components that have a certain property."

"It'd be nice if our service owners could just tag their components somehow to let the CD system know to automatically generate SRV records or not for that service."

Labels are mainly intended to be used for filtering of entities, by external systems that want to find entities that have some certain property. This is sometimes used for feature detection / selection. An example could be to add a label deployments.my-company.net/register-srv: "true".

At the time of writing this, the use of labels is very limited and we are still settling together with the community on how to best use them. If you feel that your use case fits the labels best, we would appreciate if you let the Backstage maintainers know.

You are free to add labels. This carries no risk or impact to other systems as long as you abide by the following naming rules.

  • The backstage.io label prefix is reserved for use by the Backstage maintainers. Reach out to us if you feel that you would like to make an addition to that prefix.
  • Labels that pertain to a well known third party system should ideally be prefixed with a domain, in a way that makes sense to a reader and connects it clearly to the system (or the maker of the system). For example, you might use a pagerduty.com prefix for pagerduty related labels, but maybe not ldap.com for LDAP labels since it's not directly affiliated with or owned by an LDAP foundation/company/similar.
  • Labels that have no prefix at all, are considered local to your Backstage instance and can be used freely as such, but you should not make use of them outside of your organization. For example, if you were to open source a plugin that generates or consumes labels, then those labels must be properly prefixed with your company domain or a domain that pertains to the label at hand.

Adding a New Relation Type​

Example intents:

"We have this concept of service maintainership, separate from ownership, that we would like to make relations to individual users for."

"We feel that we want to explicitly model the team-to-global-department mapping as a relation, because it is core to our org setup and we frequently query for it."

Any processor can emit relations for entities as they are being processed, and new processors can be added when building the backend catalog using the CatalogBuilder. They can emit relations based on the entity data itself, or based on information gathered from elsewhere. Relations are directed and go from a source entity to a target entity. They are also tied to the entity that originated them - the one that was subject to processing when the relation was emitted. Relations may be dangling (referencing something that does not actually exist by that name in the catalog), and callers need to be aware of that.

There is a set of well-known relations, but you are free to emit your own as well. You cannot change the fact that they are directed and have a source and target that have to be an entity reference, but you can invent your own types. You do not have to make any changes to the catalog backend in order to accept new relation types.

At the time of writing this, we do not have any namespacing/prefixing scheme for relation types. The type is also not validated to contain only some particular set of characters. Until rules for this are settled, you should stick to using only letters, dashes and digits, and to avoid collisions with future core relation types, you may want to prefix the type somehow. For example: myCompany-maintainerOf + myCompany-maintainedBy.

If you have a suggestion for a relation type to be elevated to the core offering, reach out to the Backstage maintainers or a support partner.

Using a Well-Known Relation Type for a New Purpose​

Example intents:

"The ownerOf/ownedBy relation types sound like a good fit for expressing how users are technical owners of our company specific ServiceAccount kind, and we want to reuse those relation types for that."

At the time of writing, this is uncharted territory. If the documented use of a relation states that one end of the relation commonly is a User or a Group, for example, then consumers are likely to have conditional statements on the form if (x.kind === 'User') {} else {}, which get confused when an unexpected kind appears.

If you want to extend the use of an established relation type in a way that has an effect outside of your organization, reach out to the Backstage maintainers or a support partner to discuss risk/impact. It may even be that one end of the relation could be considered for addition to the core.

Adding a New Status field​

Example intent:

"We would like to convey entity statuses through the catalog in a generic way, as an integration layer. Our monitoring and alerting system has a plugin with Backstage, and it would be useful if the entity's status field contained the current alert state close to the actual entity data for anyone to consume. We find the status.items semantics a poor fit, so we would prefer to make our own custom field under status for these purposes."

We have not yet ventured to define any generic semantics for the status object. We recommend sticking with the status.items mechanism where possible (see below), since third party consumers will not be able to consume your status information otherwise. Please reach out to the maintainers on Discord or by making a GitHub issue describing your use case if you are interested in this topic.

Adding a New Status Item Type​

Example intent:

"The semantics of the entity status.items field are fine for our needs, but we want to contribute our own type of status into that array instead of the catalog specific one."

This is a simple, low risk way of adding your own status information to entities. Consumers will be able to easily track and display the status together with other types / sources.

We recommend that any status type that are not strictly private within the organization be namespaced to avoid collisions. Statuses emitted by Backstage core processes will for example be prefixed with backstage.io/, your organization may prefix with my-org.net/, and pagerduty.com/active-alerts could be a sensible complete status item type for that particular external system.

The mechanics for how to emit custom statuses is not in place yet, so if this is of interest to you, you might consider contacting the maintainers on Discord or my making a GitHub issue describing your use case. This issue also contains more context.

Referencing different environments with the model​

Example intent:

"I have multiple versions of my API deployed in different environments so I want to have mytool-dev and mytool-prod as different entities."

While it's possible to have different versions of the same thing represented as separate entities, it's something we generally recommend against. We believe that a developer should be able to just find for example one Component representing a service, and to be able to see the different code versions that are deployed throughout your stack within its view. This reasoning works similarly for other kinds as well, such as API.

That being said - sometimes the differences between versions are so large, that they represent what is for all intents and purposes an entirely new entity as seen from the consumer's point of view. This can happen for example for different significant major versions of an API, and in particular if the two major versions coexist in the ecosystem for some time. In those cases, it can be motivated to have one my-api-v2 and one my-api-v3 named entity. This matches the end user's expectations when searching for the API, and matches the desire to maybe have separate documentation for the two and similar. But use this sparingly - only do it if the extra modelling burden is outweighed by any potential better clarity for users.

When writing your custom plugins, we encourage designing them such that they can show all the different variations through environments etc under one canonical reference to your software in the catalog. For example for a continuous deployment plugin, a user is likely to be greatly helped by being able to see the entity's versions deployed in all different environments next to each other in one view. That is also where they might be offered the ability to promote from one environment to the other, do rollbacks, see their relative performance metrics, and similar. This coherency and collection of tooling in one place is where something like Backstage can offer the most value and effectiveness of use. Splitting your entities apart into small islands makes this harder.

Implementing custom model extensions​

This section walks you through the steps involved extending the catalog model with a new Entity type.

Creating a custom entity definition​

The first step of introducing a custom entity is to define what shape and schema it has. We do this using a TypeScript type, as well as a JSONSchema schema.

Most of the time you will want to have at least the TypeScript type of your extension available in both frontend and backend code, which means you likely want to have an isomorphic package that houses these types. Within the Backstage main repo the package naming pattern of <plugin>-common is used for isomorphic packages, and you may choose to adopt this pattern as well.

You can generate an isomorphic plugin package by running:yarn new --select plugin-common or you can run yarn new and then select "plugin-common" from the list of options

There's at this point no existing templates for generating isomorphic plugins using the @backstage/cli. Perhaps the simplest way to get started right now is to copy the contents of one of the existing packages in the main repository, such as plugins/scaffolder-common, and rename the folder and file contents to the desired name. This example uses foobar as the plugin name so the plugin will be named foobar-common.

Once you have a common package in place you can start adding your own entity definitions. For the exact details on how to do that we defer to getting inspired by the existing scaffolder-common package. But in short you will need to declare a TypeScript type and a JSONSchema for the new entity kind.

Building a custom processor for the entity​

The next step is to create a custom processor for your new entity kind. This will be used within the catalog to make sure that it's able to ingest and validate entities of our new kind. Just like with the definition package, you can find inspiration in for example the existing ScaffolderEntitiesProcessor. We also provide a high-level example of what a catalog process for a custom entity might look like:

import { CatalogProcessor, CatalogProcessorEmit, processingResult } from '@backstage/plugin-catalog-node';
import { LocationSpec } from '@backstage/plugin-catalog-common'
import { Entity, entityKindSchemaValidator } from '@backstage/catalog-model';

// For an example of the JSONSchema format and how to use $ref markers to the
// base definitions, see:
// https://github.com/backstage/backstage/tree/master/packages/catalog-model/src/schema/kinds/Component.v1alpha1.schema.json
import { foobarEntityV1alpha1Schema } from '@internal/catalog-model';

export class FoobarEntitiesProcessor implements CatalogProcessor {
  // You often end up wanting to support multiple versions of your kind as you
  // iterate on the definition, so we keep each version inside this array as a
  // convenient pattern.
  private readonly validators = [
    // This is where we use the JSONSchema that we export from our isomorphic
    // package
    entityKindSchemaValidator(foobarEntityV1alpha1Schema),
  ];

  // Return processor name
  getProcessorName(): string {
    return 'FoobarEntitiesProcessor'
  }

  // validateEntityKind is responsible for signaling to the catalog processing
  // engine that this entity is valid and should therefore be submitted for
  // further processing.
  async validateEntityKind(entity: Entity): Promise<boolean> {
    for (const validator of this.validators) {
      // If the validator throws an exception, the entity will be marked as
      // invalid.
      if (validator(entity)) {
        return true;
      }
    }

    // Returning false signals that we don't know what this is, passing the
    // responsibility to other processors to try to validate it instead.
    return false;
  }

  async postProcessEntity(
    entity: Entity,
    _location: LocationSpec,
    emit: CatalogProcessorEmit,
  ): Promise<Entity> {
    if (
      entity.apiVersion === 'example.com/v1alpha1' &&
      entity.kind === 'Foobar'
    ) {
      const foobarEntity = entity as FoobarEntityV1alpha1;

      // Typically you will want to emit any relations associated with the
      // entity here.
      emit(processingResult.relation({ ... }))
    }

    return entity;
  }
}

Once the processor is created it can be wired up to the catalog via the CatalogBuilder in packages/backend/src/plugins/catalog.ts:

packages/backend/src/plugins/catalog.ts
import { FoobarEntitiesProcessor } from '@internal/plugin-foobar-backend';

export default async function createPlugin(
  env: PluginEnvironment,
): Promise<Router> {
  const builder = await CatalogBuilder.create(env);
  builder.addProcessor(new FoobarEntitiesProcessor());
  const { processingEngine, router } = await builder.build();
  // ..
}