UPWARD Specification

Unified Progressive Web App Response Definitions are simple files describing how a web server delivers and supports a [Progressive Web Application][pwa def]. They denote server behavior in a platform-independent way, so that a PWA client application expecting certain behavior from HTTP endpoints can be deployed on any type of tech stack that implements the UPWARD specification.

See RATIONALE.md for a detailed explanation of the context that led to the introduction of UPWARD, and the problems it intends to solve.

See EXECUTION_SCHEDULING_STRATEGIES.md for a deep dive into how the core logic of an UPWARD-compliant server might be implemented with a state machine.

See UPWARD_MAGENTO.md for context on how UPWARD fills a need in Magento PWA Studio and Magento 2 frontend development.

• UPWARD Specification
  • Quickstart
  • Summary
    • Echo Example
  • Configuration
  • Responding to requests
    • Execution scheduling and ordering
    • Cyclic dependencies
    • Response flush triggering
  • Context Reference
    • Initial context
      • Context Path Syntax
      • Using context
      • Context persistence and size
  • Resolver Reference
    • InlineResolver
      • InlineResolver Configuration Options
    • FileResolver
      • FileResolver Configuration Options
      • Parsing
      • FileResolver Error Handling
      • FileResolver shorthand
    • ServiceResolver
      • ServiceResolver Configuration Options
      • Example REST service call
      • Response Assignment
      • ServiceResolver Error Handling
    • TemplateResolver
      • TemplateResolver Configuration Options
      • Template Context
      • Template Engines
        • Example React DOM Server support
      • TemplateResolver Error Handling
    • ConditionalResolver
      • ConditionalResolver Configuration Options
      • Matchers
      • Match context
      • ConditionalResolver notes
    • ProxyResolver
      • ProxyResolver Example
      • ProxyResolver Configuration Options
      • ProxyResolver notes
    • DirectoryResolver
      • DirectoryResolver Example
      • DirectoryResolver Configuration Options
    • UrlResolver
      • UrlResolver Example
      • UrlResolver Configuration Options
      • UrlResolver Notes
    • ComputedResolver
      • ComputedResolver Example
      • ComputedResolver Configuration Options
      • ComputedResolver Notes
  • Reducing boilerplate
    • Default parameters
    • Builtin constants
    • Resolver type inference
    • YAML anchors

Quickstart

This repository is a test suite for UPWARD compliance, testing several scenarios and features on a live web server. It requires NodeJS v8 LTS or later. To test an UPWARD server:

1. Write or obtain a POSIX shell script which:

   • gets the path to an UPWARD definition file from the environment variable UPWARD_PATH
   • launches and/or binds the UPWARD server under test and runs it in the foreground (not as a daemon process).
   • prints the hostname and port of the now-running server instance to standard out
   • responds to SIGTERM by gracefully closing the server

   [Example here.][spec-shell-script]

2. Use npx to run upward-spec on your shell script

   npx @magento/upward-spec ./test_upward_server.sh
   

3. The shell script will run for each test suite with the environment variable UPWARD_YAML set to the path of a fixture YAML file for configuring a server instance. The script should launch a server (on a local port or a remote port, but resolvable to the local system) and print its host to standard out, staying in the foreground.

   The tests may run in parallel, so the server or launch script should seek open ports to bind to. When each test suite is over, the script will receive a SIGTERM or SIGKILL.

4. By default, the test runner will print human-readable results to stdout; the argument --xunit will make it print XUnit-compatible (and therefore JUnit-compatible) test result XML. The argument --tap will make it print Test Anything Protocol-compatible text. Under the hood, this uses tape and it can be piped to any number of open-source TAP reporters.

ℹ️ (The npx tool above is not required; it's a convenience script to avoid installing global NPM dependencies. You can also install upward-spec permanently using yarn global add upward-spec, and then simply invoke upward-spec ./test_upward_server.sh from that point forward.)

Summary

UPWARD definitions are YAML files which declare the behavior of an [application shell][application shell] server. An application shell server implements a strict subset of HTTP functionality: it proxies to API backends, it serves static files, and it handles an HTTP GET request for a resource by delivering a minimal app shell: the code and data to bootstrap a Progressive Web App which displays that resource.

An App Shell is purposefully minimal, and so is an UPWARD server. It is meant to initialize or refresh sessions, deliver small HTML documents with enough server-side rendering for initial display and SEO, and then hand off subsequent request handling to the PWA running in the client.

The declarative format of UPWARD means that an UPWARD-compliant server may be written in any programming language and run on any tech stack; therefore, a PWA can declare the URIs and behavior of the network endpoints it depends on by including an UPWARD file.

Echo Example

This example definition file echoes request data as text back to the client.

status:
  resolver: inline
  inline: 200
headers:
  resolver: inline
  inline:
    content-type: text/plain
body:
  resolver: template
  engine: mustache
  provide:
    - request
  template:
    resolver: inline
    inline: |
      {{#request}}
      Headers:
          {{#headerEntries}}
          {{name}}: {{value}}
          {{/headerEntries}}
      URL:
          {{#url}}{{#?protocol}}protocol: {{protocol}}
          {{/?protocol}}{{#?host}}host: {{host}}
          {{/?host}}{{#?hostname}}hostname: {{hostname}}
          {{/?hostname}}{{#?port}}port: {{port}}
          {{/?port}}pathname: {{pathname}}
          {{/url}}
      URL Query:
          {{#queryEntries}}
          {{name}}: {{value}}
          {{/queryEntries}}
      {{/request}}

This example demonstrates the initial properties of the context object, populated by the originating HTTP request. describes a server which always returns status 200 with a single header, content-type, and a text body which is a plaintext summary of the GET request properties. An example request to such a server results in:

$ curl 'https://localhost:54422/head/shoulders?and=knees&and=toes'

Headers:
    host: localhost:54422
    user-agent: curl/7.54.0
    accept: */*
URL:
    host: localhost:54422
    hostname: localhost
    port: 54422
    pathname: /head/shoulders
URL Query:
    and: knees,toes

Configuration

A spec-compliant UPWARD server should be configurable with a runtime parameter: the location of the UPWARD definition YML file. The file references external resources using Resolvers.

Responding to requests

An UPWARD definition file is an instruction manual for building an HTTP response. It links values together in a global namespace hereafter called the context. Each request handling cycle begins with a new context object, with the incoming Request assigned to the context value request, and current environment variables assigned to the context value env. Root properties of the definition file represent other named values in the context, which Resolvers populate. Resolvers can use other context properties as input, and they can also use other Resolvers directly; in this way, the definition itself can be considered an abstract decision tree, from which code could be statically generated.

Execution scheduling and ordering

1. Resolvers must execute only when needed. If a request handling cycle moves through ConditionalResolver branches in a way that never requires a particular context value, then that context value must never be resolved during that execution cycle.

2. Resolvers must execute as concurrently as possible. The maximum concurrency is left to the implementation. A compliant server detects when a Resolver uses a context value, and delays its execution until that context value becomes available, via [topological sorting][topological sorting] of resolver execution.

See EXECUTION_SCHEDULING_STRATEGIES.md for an example of how this could be implemented.

Cyclic dependencies

An UPWARD server must detect cyclic dependencies, and should detect them as soon as possible. Only bare string literals can be used to reference context objects in a definition file, so it should be possible to detect potential cyclic dependencies upon first processing the configuration file, before any requests are made. If a cyclic dependency is detected in this manner, the server should raise an error on startup. If a cyclic dependency occurs at runtime, the server should return a 500 error.

Response flush triggering

The root context must always eventually have non-null status, headers, and body properties. Once the resolution path has assigned these three values fo context, the UPWARD-compliant server should immediately use those three values to create an HTTP response and flush it to the client. No streaming or buffering interface should be provided; UPWARD servers should not deal in data large enough to require streaming. If all resolvers finish executing and the response is lacking any of the status, headers, or body properties, the server should emit a 500 error. If it is possible at startup time to trace a path through the decision tree where this occurs, the implementation may use static analysis to do so and raise an error on startup.

ℹ️ _In real-world scenarios, the generation of status, headers, and body will share a lot of logic. The recommended best practice is to use an InlineResolver to create a top-level object, called something like "response", with those properties, and then define the top-level status, headers, and body properties to refer to that object in context, e.g. status: response.status.

Context Reference

The context is a global namespace created for each request. It is a dictionary of values, like a JSON object. A typical response cycle may append intermediate values to the context, like a query result or a template string. Those values do not emit as part of the response. Only the status, headers, and body properties of the context will be flushed to the client. Since the context is the global namespace and the means of sharing values between resolvers, it may become a large object at some points in the cycle, but this should not affect performance on the client.

Context values cannot be overwritten. A resolver which attempts to overwrite an already-set context value must raise a context conflict error. UPWARD-compliant servers should be able to identify potential context conflict errors during static analysis of the definition file and raise an error on launch. UPWARD-compliant servers must be able to identify a context conflict during runtime execution and respond with a 500 error.

Initial context

When an UPWARD-compliant server receives an HTTP GET request, it must populate an initial context object with the following values:

• request: an object representing the incoming HTTP request. It must have these properties:

  • headers: An object representing HTTP headers. Header names are lower-cased, and multiple values are joined with commas.

  • headerEntries: An iterable array version of the headers object, suitable for use in a Mustache template (which cannot iterate over plain JSON objects). For this headers object:

    {
      "accept": "text/html",
      "host": "example.com"
    }
    

    the headerEntries array would be:

    [
      { "name": "accept", "value": "text/html" },
      { "name": "host", "value": "example.com" }
    ]
    

  • queryEntries: An iterable array version of the URL query object, suitable for Mustache like the headerEntries property explained above.

  • url: A subset of a [URL record][url spec] as specified by WHATWG. The following properties should at least be populated; the Host header can be used to infer the origin.

    Attribute	Example
    host	example.com:8080
    hostname	example.com
    port	8080
    pathname	/deep/blue/sea
    search	?baby=beluga
    query	{ "baby": "beluga" }

    Because HTTP servers are sometimes unable to ascertain their own domain names or origins, it is acceptable for one or more of the href, origin, protocol, username, password, host, hostname, and port properties to be undefined. However, a compliant server MUST provide pathname, search, and query.

    The query property is not part of the WHATWG specification, but it must be included in the url object nevertheless. Much like the headers and headerEntries properties, these objects exist for property lookup and iteration in logic-less templates, respectively.

• env: an object containing the environment variables set when the server was launched. For instance, if a Dockerfile launches the server through Apache, with an environment variable MAGENTO_GRAPHQL_ENDPOINT:

  ENV MAGENTO_GRAPHQL_ENDPOINT https://m2host.com/graphql
  CMD ["/usr/sbin/apache2ctl", "-D", "FOREGROUND"]
  

  then the context value env.MAGENTO_GRAPHQL_ENDPOINT would equal "https://m2host.com/graphql".

For convenience and concision, frequently used strings should be registered as self-referential top-level context objects. Examples include:

• GET: the string 'GET' must be included in the global context for convenience, to reduce boilerplate

• POST: the string 'POST' must be included in the global context for convenience, to reduce boilerplate

• mustache: the string 'mustache', included in the global context for convenience, to reduce boilerplate

• Other strings that must be preset as literals in the context include:

  • text/html
  • text/plain
  • application/json
  • utf-8
  • latin-1
  • base64
  • hex
  • All valid HTTP response codes, i.e. 200, 404, 500, and all others.

Context Path Syntax

In UPWARD configuration, a bare string is treated as a context lookup by default. Some resolver configuration properties must treat bare strings as literal strings, and the default contains many constants for common strings, as mentioned earlier, so a definition file may appear to use many literal strings, but almost all string values are context lookups.

A context lookup resembles "dot lookup" notation in JavaSript or Python, though it has simpler rules and cannot be dynamically generated.

Rules:

• Valid characters: A context lookup may contain any UTF-8 characters except control characters, whitespace, or newlines. All characters that are not lookup operators must be treated as part of a property name. For the sake of easier manipulation in common programming languages, it is a best to use context names which are legal identifiers in those languages.)
  • A context lookup cannot begin with a dot ., which is the lookup operator.
• Lookup operators: The dot . denotes a property lookup on an object. A positive integer property name in a context lookup, such as characters.0.name, will perform array access by index. Array indices must be positive integers; they must be specified literally, just like the rest of a context lookup string, and cannot be dynamically evaluated.
• A context lookup always starts with a basename. This may be any string that can be a valid YAML property name. The basename has special significance; when a top-level resolver completes resolution, it assigns a value to this name in the global context, which should trigger resolution of all paths beginning with that name.
• Behavior for undeclared properties: Undeclared property lookup should silently fail, resolving the empty string. If the basename foo is assigned, but it is not an object or it does not contain the property bar, then the context lookup foo.bar should wait until foo is assigned, and then yield the empty string.

Parts of a context lookup:

         string property names
basename    |               |
     |      |  array index  |
     |      |          |    |
  uxbridges.characters.0.name
           |             |
        named property lookups

The above context lookup behaves at runtime as an instruction to wait until the uxbridges basename has been assigned. There must be a definition elsewhere in the definition file which assigns this property, e.g.

uxbridges:
  resolver: service
  method: POST
    resolver: inline
      inline: false
  url:
    resolver: inline
    inline: https://stapi.co/api/v1/rest/character/search
  headers:
    resolver: inline
    inline:
      content-type: application/x-www-form-urlencoded
  query:
    resolver: inline
    inline: '{
      characters @rest(type: "Character", path: )
    }'

The above definition would assign an HttpResponse to the uxbridges basename once it has run. An HTTP response has no characters property, so the example context lookup would resolve to the empty string. However, an HTTP response does have a body property, so the lookup uxbridges.body.characters.0.name would resolve to Kevin Uxbridge.

ℹ️ (Array and list handling is intentionally rudimentary in UPWARD, because of the potential for scope confusion, performance and security issues in iteration. The only recommended use case for list lookup is when a web service returns a list of items expected to have only one result, so that the result may be lifted out into a scalar value.)

Since arbitrary property lookups do not through exceptions and instead return a default empty string, use pattern matching to test the success or failure of requests and responses:

matches: uxbridges.body.characters.0..name
pattern: '.'

The regex dot character will pass if the string is non-empty, and fail if it is empty. This pattern should be common in UPWARD definitions.

Using context

Anywhere a in the definition file where a resolver is allowed, you may substitute a context lookup instead. A context path indicates a dependency on another root context value, which causes the current branch to wait until that value is resolved. In this example:

body:
  resolver: conditional
  when:
    - matches: request.url.query.shipmentId
      pattern: '\d+'
      use:
        resolver: template
        engine: mustache
        template:
          resolver: file
          file:
            resolver: inline
            inline: './renderShipment.mst'
        provide:
          shipment:
            resolver: service
            query:
              resolver: file
              file:
                resolver: inline
                inline: './getShipment.graphql'
            variables:
              id: request.url.query.shipmentId
    - matches: request.url.query.shipmentName
      pattern: '\w+'
      use:
        resolver: template
        engine: mustache
        template:
          resolver: file
          file:
            resolver: inline
            inline: './renderShipment.mst'
        provide:
          shipment:
            resolver: service
            url: env.SHIPMENTS_SVC
            query:
              resolver: file
              file:
                resolver: inline
                inline: './getShipment.graphql'
            variables:
              name: request.url.query.shipmentName
    - matches: request.url.query.shipmentTrackingNumber
      pattern: '[\w\d\.]+'
      use:
        resolver: template
        engine: mustache
        template:
          resolver: file
          file:
            resolver: inline
            inline: './renderShipment.mst'
        provide:
          shipment:
            resolver: service
            url: env.SHIPMENTS_SVC
            query:
              resolver: file
              file:
                resolver: inline
                inline: './getShipment.graphql'
            variables:
              trackingNumber: request.url.query.shipmentTrackingNumber
    default:
      inline: 'Please provide a query'

The matchers use the context lookups request.url.query.shipmentId, request.url.query.shipmentName, and request.url.query.shipmentTrackingNumber to asynchronously obtain values for comparison, as soon as they have been derived. The request object is part of the initial context, but you can also declare dependencies on other context values you have defined. The above definition could be more concisely expressed using intermediate values and more context lookups:

body:
  resolver: conditional
  when:
    - matches: gqlVariables
      pattern: null
      use:
        inline: 'Please provide a query'
  default:
    resolver: template
    engine: mustache
    template:
      resolver: file
      file:
        resolver: inline
          inline: './renderShipment.mst'
    provide:
      shipment:
        resolver: service
        url: env.SHIPMENTS_SVC
        query:
          resolver: file
          file:
            resolver: inline
              inline: './getShipment.graphql'
        variables: gqlVariables
gqlVariables:
  resolver: conditional
  when:
    - matches: request.url.query.shipmentId
      pattern: '\d+'
      use:
        resolver: inline
        inline:
          id: request.url.query.shipmentId
    - matches: request.url.query.shipmentName
      pattern: '\w+'
      use:
        resolver: inline
        inline:
          name: request.url.query.shipmentName
    - matches: request.url.query.shipmentTrackingNumber
      pattern: '[\w\d\.]+'
      use:
        resolver: inline
        inline:
          trackingNumber: request.url.query.trackingNumber
  default: null

The above definition produces a server with identical behavior to the previous query, in 30% fewer lines. This is not the only way to use context values to reduce the first, repetitive definition, but it demonstrates the idea. If the GraphQL service set an order of precedence for the shipment query which used the same priority logic as the ConditionalResolver, the definition could be further reduced to use the same variable set in all cases:

body:
  resolver: conditional
  when:
    - matches: request.url.query.shipmentId
      pattern: '\d+'
      use: getShipment
    - matches: request.url.query.shipmentName
      pattern: '\w+'
      use: getShipment
    - matches: request.url.query.shipmentTrackingNumber
      pattern: '[\w\d\.]+'
      use: getShipment
  default:
    inline: 'Please provide a query'
getShipment:
  resolver: template
  engine: mustache
  template:
    resolver: file
    file:
      resolver: inline
        inline: './renderShipment.mst'
  provide:
    shipment:
      resolver: service
      url: env.SHIPMENTS_SVC
      query:
        resolver: file
        file:
          resolver: inline
            inline: './getShipment.graphql'
      variables:
        id: request.url.query.shipmentId
        name: request.url.query.shipmentName
        trackingNumber: request.url.query.trackingNumber

Setting top-level context properties for reusable values is an important way to reduce verbosity and code duplication in the definition file. See Reducing boilerplate for more practices to reduce verbosity in a definition file.

Context persistence and size

When writing UPWARD definitions, no distinction needs to be made between "persistent" context values that should be the same for each request, and other values that may differ for each request. An UPWARD server must do topological sorting to determine the order in which to run Resolvers, so an efficient implementation can identify the parts of the root context that are not dependent on the incoming request and cache or discard them.

Resolver Reference

A Resolver is an object which describes how a value is obtained. There are five kinds of Resolver:

• InlineResolver adds hardcoded values
• FileResolver loads files from a filesystem
• ServiceResolver places GraphQL queries and loads the result set
• TemplateResolver renders a template string against the context
• ConditionalResolver does branch logic using pattern matching on context values
• ProxyResolver delegates request/response handling to a proxy
• DirectoryResolver delegates request/response handling to a static file directory
• UrlResolver builds a URL from strings and other URLs
• ComputedResolver resolves to a PHP class to get data

Each Resolver takes different configuration parameters. Like a context lookup string, a resolver represents an operation which will execute and then deliver its results upward in the tree, until all dependencies of the top-level status, headers, and body definitions are resolved.

InlineResolver

There is sometimes a need to put literal values in the context. However, a string property value performs a context lookup:

body: 'Hello world!'

The above expression uses Hello world! as a property name and tries to substitute the value of the context property named Hello world!. This is likely not the intent of the user. To set the body to the literal string 'Hello world!', use an InlineResolver:

body:
  resolver: inline
  inline: 'Hello world!'

The inline of an InlineResolver may be of any type; it may be a primitive value, a list, or an object of arbitrary depth. List values and Object properties may be context lookups or resolvers themselves, so when building lists or objects, make sure to use InlineResolvers again at every other level of depth. (A consequence of this is that it is difficult and verbose to set a literal value resembling a resolver configuration, such as { "resolver": "inline" }, but this seems to be a minor inconvenience.)

InlineResolver Configuration Options

Property	Type	Default	Description
inline	any		Required. The value to be assigned to context.

FileResolver

Queries and templates are often large, and often reused by other systems. They can be inline strings, as in the first example, but they should usually be separate files, referenced with a FileResolver.

query:
  resolver: file
  file:
    resolver: inline
    inline: './productDetail.graphql'
  encoding:
    resolver: inline
    inline: 'utf-8'

The above expression loads the content of the file ./productDetail.graphql and sets it as the property query. The file path is resolved relative to the location of the definition file. The encoding property is optional.

FileResolver Configuration Options

Property	Type	Default	Description
file	Resolved		Required. Path to the file to be read. Resolved relative to the definition file.
encoding	Resolved	utf-8	Character set to use when reading the file as text. Can be utf-8, latin-1, or binary.
parse	Resolved	auto	Attempt to parse the file as a given file type. The default of auto should attempt to determine the file type from its extension. The value text will effectively disable parsing.

Parsing

An UPWARD server must support pre-parsing of graphql, json, and mustache files according to their respective specifications, but should support as many filetypes as necessary and may support custom filetypes.

FileResolver Error Handling

If the file cannot be found or there were any other failures reading the file, the resolver must resolve an object with a single errors property instead of a string or a parsed value. Errors must be formatted like [GraphQL errors][graphql spec errors property].

ℹ️ (File contents are not expected to change during server runtime, so an UPWARD-compliant server should cache file contents on startup.)

FileResolver shorthand

Filenames are usually not variable; they will usually be specified as literals, rather than dynamically resolved out of context. For readability and convenience, a "shorthand syntax" must be available to imply a FileResolver that loads and parses a UTF-8 encoded file from a string filesystem path. Instead of explicitly resolving a filepath as an inline string:

query:
  resolver: file
  file:
    resolver: inline
    inline: './path/to/file.graphql'

The shorthand syntax should allow the following to be equivalent:

query: './path/to/file.graphql'

An UPWARD-compliant server must treat a barestring as a literal filepath instead of a context lookup when:

• The string is being argued to a configuration parameter that accepts a FileResolver for the implied type of the file after parsing

  AND

• The string begins with one of the following prefixes:

  • a relative path prefix, ./ and/or one or more ../
  • an absolute path prefix, / or C:\ or any other Windows drive letter
  • a file URI scheme, file://

  AND

• Upon checking the filesystem, the string resolves to a [regular file][regular file] (not a symlink, device, or directory)

In the latter case, if the shorthand string neither resolves to a legal file, nor exists as a resolvable context value, an UPWARD-compliant server should raise a detailed error message explaining this.

ServiceResolver

An UPWARD server uses a ServiceResolver to obtain live data from a GraphQL backing service. URL, method, and headers can be specified manually; the query itself is constructed using query and variables parameters.

documentResult:
  resolver: service
  endpoint:
    resolver: url
    baseUrl:
      inline: https://example.com
    pathname:
      inline: graphql
  method:
    resolver: inline
    inline: POST
  headers:
    resolver: inline
    inline:
      'content-type': 'application/json'
      accept: 'application/json'
      # A small inline template renders the full syntax of the OAuth Bearer
      # token header, instead of just the plain token.
      authorization:
        resolver: template
        engine: 'mustache'
        template:
          resolver: 'inline'
          inline: 'Bearer {{env.BEARER_TOKEN}}'
  query:
    resolver: inline
    inline: 'query getDocument($id: String!) {
      document(id: $id) {
        title
        contents
      }
    }'
  variables:
    resolver: inline
    inline:
      # This is a barestring indicating a context lookup. It resolves to the
      # `id` value in the URL query string of the request, using the builtin
      # `request` context object.
      id: request.url.query.id

ℹ️ (For the purposes of demonstration, the query here is resolved inline. The best practice is to store queries in files and use FileResolvers to obtain them.)

ServiceResolver Configuration Options

Property	Type	Default	Description
endpoint	Resolved	https://localhost/graphql	Required. The URL of the service endpoint to call.
url	Resolved	https://localhost/graphql	🚫 Deprecated. Synonymous with endpoint. Replaced with endpoint for readability. This parameter is still supported for now, but should be replaced with endpoint. If both are present, an error will be thrown.
method	Resolved	POST	The HTTP method to use. While GraphQL queries are typically POSTS, some services expose GraphQL over GET instead.
headers	Resolved>		Additional HTTP headers to send with the GraphQL request. Some headers are set automatically, but the headers configuration can append to headers that can have multiple values.
query	Resolved		Required. The GraphQL query object. Can either be a parsed query, or a string that can be parsed as a valid query.
variables	Resolved>	{}	Variables to use with the GraphQL query. Must resolve to an object with keys and values, almost always with an InlineResolver.

ServiceResolvers always use GraphQL. To obtain data from a non-GraphQL service, an UPWARD server may implement client-side directives which change the behavior of a GraphQL query, such as [apollo-link-rest][apollo-link-rest], and place the directives in the query itself. This should be transparent to the UPWARD server itself, which delegates the service call to a GraphQL client. If an UPWARD server's GraphQL client has no implementation for such a directive, then it must pass the query unmodified to the backing service to handle the directive.

Example REST service call

The below behavior is not standard or normative: it is an example of how one might query a REST service using an UPWARD server whose GraphQL client implementation has a @rest directive like [apollo-link-rest][apollo-link-rest].

documentResult:
  resolver: service
  url: env.REST_API_ENDPOINT
  method:
    resolver: inline
    inline: GET
  headers:
    # In contrast to the template above, this assumes that en environment
    # variable is already set with the "Bearer " format.
    inline:
      authorization: env.BEARER_TOKEN_STRING
  query: 'query getDocument($id: String!) {
      document(id: $id) @rest(type: "Document", path: "/documents/{args.id}") {
        title
        contents
      }
      links(documentId: $id) {
        name
        url
      }
    }'
  variables:
    inline:
      id: request.url.query.id

Response Assignment

The GraphQL specification requires that a successful response to a GraphQL query have a [root "data" property][graphql spec data property], an object whose properties correspond to the entities returned, and an that a failed response have a [root "errors" property][graphql spec errors property]. An UPWARD server must assign the entire root object to the named context value. The query above would, if successful, result in a context with documentResult.data.document and documentResult.data.links properties. If the query failed, the context would contain documentResult.errors.

ServiceResolver Error Handling

The GraphQL specification [defines error behavior][graphql spec errors] clearly, and UPWARD servers should pass the errors collection from a GraphQL response as described above. Other errors can occur during resolution, such as:

• Network errors: the URL is unresolvable or unresponsive
• Parse errors: a dynamically supplied query could not be parsed
• Validation errors: required variables were absent or the wrong type

An UPWARD server should format these errors the same way that GraphQL services do; it should return an object with an errors array containing all errors encountered.

TemplateResolver

Once the UPWARD server assembles the data for a response, it must turn that data into a response body. The TemplateResolver renders a template into a string, executing it with the context object as its root value by default. UPWARD servers must provide a Mustache template renderer that adheres to the [Mustache specification][mustache spec]. UPWARD servers can also provide other template renderers, which must evaluate context objects into strings.

status:
  resolver: inline
  inline: 200
headers:
  resolver: inline
  inline:
    'content-type':
      resolver: inline
      inline: 'text/html'
body:
  resolver: template
  engine:
    resolver: inline
    inline: mustache
  root: documentResult.data.document
  template:
    resolver: inline
    inline: |
      {{> headtag}}
      {{> header}}
      <div class="document">
        <h1>{{title}}</h1>
        <div class="document-body">
          {{&amp; contents}}
        </div>
      </div>
      {{&gt; footer}}

The above configuration resolves into an HTML document displaying content from the documentResult.data.document context value. Its use of [Mustache partials][mustache partials] implies that additional files called headtag.mst, header.mst, and footer.mst exist in the directory containing the definition file. Attempting to include a missing partial should raise an error as soon as the template is resolved, ideally at server startup time.

ℹ️ (For illustrative purposes, the above uses an InlineResolver where it would be more appropriate to use a FileResolver to obtain the template, as with the query in the ServiceResolver example.)

TemplateResolver Configuration Options

Property	Type	Default	Description
engine	Resolved		Required. The label of the template engine to use.
provide	Resolved>		Required. A list, or an object mapping, of values to make available in the template. Passing the entire context to a template for evaluation can cause cyclic dependencies.
template	Resolved		The template to render.

Template Context

The entire context cannot be available for template render; that would cause an immediate circular dependency, since the template's output is added to context! Instead, use the provide argument to select what values the template actually needs. They will be available, at root, inside the template.

The provide argument can be a list:

provide:
  - env
  - articleResult

The resulting template eval context might look like this:

"env": {
  "envVars": "here"
},
"articleResult": {
  "data": {
    "article": {
      articleContents: 'here'
    }
  }
}

Lists may only inject "base" context properties. The above articleResult could not be articleResult.data.article when using the list format.

The other, more powerful option for the provide argument is to provide a mapping, as a simple object. A mapping must resolve to a plain object of string keys and context values. It might appear as:

provide:
  inline:
    article: articleResult.data.article

This would give the template a single root property "article", thus flatting out the template tree and making templates more readable.

"article": {
  articleContents: 'here'
}

Template Engines

The engine property must resolve to a string labeling a supported template engine. The only required template engine is Mustache, and its label must be mustache. An UPWARD server may support additional template engines. For instance, an UPWARD server may support [ReactJS server-side rendering][react dom server].

Example React DOM Server support

body:
  resolver: template
  engine:
    resolver: inline
    inline: react
  provide:
    inline:
      document: documentResult.data.document
      query: request.url.query
  template:
    resolver: file
    file:
      resolver: inline
      inline: './build/RootComponents/Document.js'

The above configuration assumes support for a template engine labeled react. The underlying template engine could be a simple Node module:

const { createElement } = require('react');
const { renderToString } = require('react-dom/server');

module.exports = (props, component) =>
  renderToString(createElement(require(component), props));

Attaching such a template engine to a Resolver would be trivial.

TemplateResolver Error Handling

If the engine is unknown or not supported, the UPWARD server should detect this as soon as possible and send a 500 error.

If there were any errors parsing the template, evaluating, or serializing the data, the resolver must resolve an object with an errors array instead of a rendered string. Errors must be formatted like [GraphQL errors][graphql spec errors property].

ConditionalResolver

A ConditionalResolver tests a context value for a particular pattern, and then yields to another Resolver depending on the match results.

The ConditionalResolver is the only branch logic operation available in the UPWARD specification. It performs pattern matching on a context value using [Perl compatible regular expressions][pcre]. It can only test a single context value for each potential match, and it has no Boolean operators such as AND, OR, or NOT. All of these logical operations can be achieved through pattern matching, however. It takes two configuration values: when must be a list of Matcher>>, and default must be a resolver to use if none of the when conditions are true.

monkey:
  resolver: conditional
  when:
    # All values are coerced to strings for regex matching.
    - matches: request.url.query.grab
      pattern: '^(true|1)$'
      use:
        resolver: inline
        inline: "do anyway"
    - matches: status
      pattern: '^403$'
      use:
        resolver: inline
        inline: "see"
  default:
    resolver: inline
    inline: "do"
status:
  resolver: inline
  inline: 403
body:
  resolver: template
  engine:
    resolver: inline
    inline: mustache
  template:
    resolver: inline
    inline: "<p>monkey <b>{{ monkey }}</b>.</p>"

The above configuration uses a ConditionalResolver to create a context value monkey. The when list contains two matchers. They are executed in order. The first matcher tests if the request query string has a value grab matching the regular expression ^(true|1)$, If the request query string contains grab=true or grab=1, this matcher succeeds and the ConditionalResolver yields to the resolver specified in the matcher's use property. The context now matches the object:

{
  "status": 403,
  "monkey": "do anyway",
  "body": "<p>monkey <b>do anyway</b></p>"
}

If the request query string does not contain a qualifying grab property, the second matcher runs. This matcher tests if the status context value matches the regular expression ^403$, effectively checking if the response status code has already been set to HTTP 403 Forbidden. The status resolver runs before the monkey resolver, though it is defined later in the document (see Concurrency), so the matcher tests against the value 403 (cast to a string) and succeeds. The ConditionalResolver uses the matcher's resolver, and the context now matches the object:

{
  "status": 403,
  "monkey": "see",
  "body": "<p>monkey <b>see</b></p>"
}

If some other configuration provides a different status, such as 200, then neither matcher in the when list succeeds, and the ConditionalResolver uses its default resolver. The context now matches the object:

{
  "status": 200,
  "monkey": "do",
  "body": "<p>monkey <b>do</b></p>"
}

ConditionalResolver Configuration Options

Property	Type	Default	Description
when	Matcher[]
default	Resolved		Required. The default resolver to use if no matcher succeeds.

Matchers

A Matcher is an object which can only be used as an item in a ConditionalResolver's when list. It must have the following properties:

Property	Type	Default	Description
matches	string		Required. The context value to match. Must be a bare string context lookup.
pattern	string		Required. [PCRE][pcre] regular expression to use to test against the value in matches.
use	Resolved		Required. Resolver to use if the match succeeds.

Match context

During the resolution of a matcher's use resolver, properties from the match object are temporarily added to the context. Using these temporary context values, resolvers can extract matching text, or capture groups, from the match itself.

• $match.$0 - The full text of the last matched value.
• $match.$1 - The string captured in the first backreference the regex declared.

The $match object must have additional numbered properties for each backreference.

ConditionalResolver notes

• Matchers run in sequence, top to bottom. They do not "fall through"; the first successful matcher will exit the conditional.
• Matchers can only test against context properties; they cannot use a resolver as the matches value.
• Each matcher can test only one context property against one pattern.
  • However, the pattern can use regex alternation to achieve OR-type logic.
  • Additionally, a list of matchers need not each test the same property.
• The default resolver is required.
• To achieve AND-like logic, nest ConditionalResolvers to arbitrary depth.
• To achieve OR-like logic, repeat the same resolver configuration in several subsequent matchers.
• If the conditional becomes verbose, consult Reducing Boilerplate for ways to simplify it.
• Though template engines with logical operators can be also be used to perform branch logic, this is not recommended; it can prevent static analysis of context value dependencies.

ProxyResolver

The ProxyResolver acts as a "passthrough" to another service. It is guaranteed to resolve into an object with status, headers, and body properties from a logical point of view. In implementation, it may handle request objects more directly, for performance purposes.

A ProxyResolver is an important part of the UPWARD philosophy: a PWA's UPWARD file must describe all of the network behavior that a PWA expects at runtime, and that includes proxies to various backing APIs.

ProxyResolver Example

proxy:
  target: env.MAGENTO_BACKEND_URL
  ignoreSSLErrors: true

ProxyResolver Configuration Options

Property	Type	Default	Description
target	Resolved
ignoreSSLErrors	Resolved	false	Ignore remnote SSL certificate errors (useful for internal communication among containers).

ProxyResolver notes

ProxyResolvers are special targets for static analysis. Using simple techniques, it should be possible for an analysis system to determine proxying rules by walking the UPWARD tree. This is effectively a declarative proxy config, and deployment tools can be enhanced to create proxy servers in front of UPWARD where appropriate.

DirectoryResolver

Like the ProxyResolver, the DirectoryResolver delegates request and response handling to a static server. Unlike the ProxyResolver, it needs access to a local directory, which it will then sere as a public assets folder. Much like the ProxyResolver, this resolver exists to bolster the notion that an UPWARD file can describe the expected behavior of a PWA server-side site, in detail.

DirectoryResolver Example

static:
  directory:
    inline: './dist'

DirectoryResolver Configuration Options

Property	Type	Default	Description
directory	Resolved		Required. The local directory path to be served.

UrlResolver

The UrlResolver is a utility for building WHATWG Spec compliant URLs out of strings and/or other URL objects. Using the UrlResolver is easier and makes more well-formed URLs than string templates.

UrlResolver Example

Consider a server which places authorized calls to a REST API running in a local container. The container is always hosted at a virtual domain admin.local, but its port, API version, and auth token may vary from deployment to deployment. For the following example, assume these environment variables:

ADMIN_PORT=8081
ADMIN_API_VERSION=1
ADMIN_REFRESH_TOKEN=a1b2c3

The following configuration defines a base URL for the REST API, and a URL to a particular resource in that API.

adminRESTTokenRefresh:
  baseUrl: adminRESTApiBase
  pathname:
    inline: adminToken
  query:
    refreshToken: env.ADMIN_REFRESH_TOKEN
    role:
      inline: owner

adminRESTApiBase:
  baseUrl:
    inline: https://admin.host/api/rest
  port: env.ADMIN_PORT
  pathname: apiVersion

apiVersion:
  engine: mustache
  provide:
    versionNumber: env.ADMIN_API_VERSION
  template:
    inline: 'v{{versionNumber}}'

URLResolvers resolve to strings. The context value adminRESTTokenRefresh evaluates to:

https://admin.host:8081/api/rest/v1/adminToken?refreshToken=a1b2c3&role=owner

UrlResolver Configuration Options

Property	Type	Default	Description
baseUrl	Resolved		Required. The base URL to use for constructing the new URL. Must be false if no base URL is required.
hash	Resolved		Hash fragment of the URL, beginning with #.
hostname	Resolved		Domain or IP address of the URL.
password	Resolved		A password to be specified before the hostname.
pathname	Resolved		Slash-delimited path from the root of the host to a resource.
port	Resolved		Port number of the URL.
protocol	Resolved	https:	Protocol scheme of the URL, including the final :.
query	Resolved>		Object to encode into a query string. Keys are query parameter names, and values must resolve to primitives (strings, numbers, booleans, etc) that will be URL-encoded into query parameter values.
search	Resolved		Serialized query string. Values must be URL (percent) encoded.

UrlResolver Notes

• If no base URL is required, either because the defined URL should be relative or the other parameters build an absolute URL, baseUrl must be explicitly set to false.

• Pathnames must join based on leading and trailing slashes.

  • If a pathname has a leading slash, it must overwrite any pathname on the baseUrl.

  baseUrl: https://fleet.local/ships/hood/
  pathname: /admiral
  

  evaluates to https://fleet.local/admiral.

  • If a pathname has no leading slash, and the pathname of the baseUrl has a trailing slash, then the pathname must append to the last segment of the baseUrl path.

  baseUrl: https://fleet.local/ships/hood/
  pathname: captain/name
  

  evaluates to https://fleet.local/ships/hood/captain/name.

  • If a pathname has no leading slash, and the pathname of the baseUrl has no trailing slash, then a pathname must replace the last segment of the baseUrl path.

  baseUrl: https://fleet.local/ships/hood
  pathname: yamato/
  

  evaluates to https://fleet.local/ships/yamato/

• If both search and query are present, the parameters must be merged, giving preference to query where there are conflicts. "Array" query parameters are not defined by this specification, since their behavior is inconsistent across platforms.

ComputedResolver

The ComputedResolver is used for the UPWARD PHP implementation to inline key page data in the initial page response without needing to make external or round trip calls. This essentially provides a way to add more resolvers than are available in this spec. It can be noted that this resolver breaks spec and its JS implementation is only to avoid crashing during build.

ComputedResolver Example

someDataPoint:
  resolver: computed
  type:
    resolver: inline
    inline: typeKey

The ComputedResolver's resolved value can vary depending on the class used. The typeKey maps to a PHP class, who takes responsibility for returning a value.

ComputedResolver Configuration Options

Property	Type	Default	Description
type	Resolved		Required. This value is used to map to a PHP class that will return a value.
``	Resolved		Since this is essentially a way to add more resolvers, any data defined in the yaml is passed to the class

ComputedResolver Notes

In its PHP implementation, the definition and resolver is passed to the class, so any information the class needs can be passed in this manner. The type is mapped to a PHP class through the di.xml and the class has access to any Magento dependency injection it needs. The JS implementation is only present to not break when the resolver type is defined.

Reducing boilerplate

The above examples are fairly verbose, to make the workings of UPWARD configuration especially clear. The UPWARD specification is intentionally verbose in its canonical format, to enable maximal static analysis. However, An UPWARD-compliant server must also include features to reduce boilerplate.

Default parameters

Many resolver configuration parameters have default values, so they can be omitted from configuration if the default is appropriate.

In the first example of a ServiceResolver above, some config parameters are actually unnecessary. The default URL for a service call is https://localhost/graphql, and the default method is POST. MIME headers are automatically generated, so the following configuration would be equivalent:

documentResult:
  resolver: service
  headers:
    resolver: inline
    inline:
      Authorization: env.BEARER_TOKEN
  query:
    resolver: inline
    inline: 'query getDocument($id: String!) {
      document(id: $id) {
        title
        contents
      }
    }'
  variables:
    id: request.url.query.id

Using the best practice of storing queries in separate files and storing reusable objects as separate context values, the configuration would likely look like this:

documentQuery: './queries/document.graphql'
oauthHeaders:
  resolver: inline
  inline:
    authorization: env.BEARER_TOKEN

documentResult:
  resolver: service
  headers: oauthHeaders
  query: documentQuery
  variables:
    id: request.url.query.id

Builtin constants

The default context must contain some builtin constants for common strings, to avoid repetitive use of InlineResolvers.

In the first example of a TemplateResolver above, some inline resolvers are unnecessary. HTTP response codes, common MIME types, common template engine labels, and other useful values must be preset in the initial context, so the following, shorter configuration is equivalent to the TemplateResolver example:

status: 200
headers:
  resolver: inline
  inline:
    'content-type': 'text/html'
body:
  resolver: template
  engine: 'mustache'
  provide:
    model: documentResult.data.document
  template:
    resolver: inline
    inline: |
      {{> headtag}}
      {{> header}}
      <div class="document">
        <h1>{{model.title}}</h1>
        <div class="document-body">
          {{&amp; contents}}
        </div>
      </div>
      {{&gt; footer}}

Resolver type inference

For any value which must be a Resolver or a context lookup, an UPWARD-compatible server should attempt to infer resolver types from a supplied resolver configuration object, rather than requiring a resolver name to be specified. Each resolver has required parameters that are mutually exclusive with one another, so a resolver type can be inferred from the presence of those parameters.

If parameter exists...	Then infer resolver type:
inline	InlineResolver
file	FileResolver
query	ServiceResolver
engine	TemplateResolver
when	ConditionalResolver
target	ProxyResolver
directory	DirectoryResolver

Resolver type inference allows configuration to omit resolver: parameters, which makes it possible to be far more terse. The example optimized configuration in Builtin constants could be further reduced:

status: 200
headers:
  inline:
    'content-type': 'text/html'
body:
  engine: 'mustache'
  root: documentResult.data.document
  template:
    inline: |
      {{> headtag}}
      {{> header}}
      <div class="document">
        <h1>{{title}}</h1>
        <div class="document-body">
          {{&amp; contents}}
        </div>
      </div>
      {{&gt; footer}}

And the example optimized configuration in Default parameters could be further reduced:

documentResult:
  headers:
    inline:
      authorization: env.BEARER_TOKEN
  query: './queries/document.graphql'
  variables:
    id: request.url.query.id

ℹ️ (Note that the variables object cannot be a resolver, as specified in ServiceResolver configuration options — it must be an object whose keys are expected query variable names and whose values are resolvers, so there is no ambiguity if a GraphQL query expects a variable named, for example, file.)

YAML anchors

The YAML specificationsupports an [anchor and reference syntax][yaml anchors] which can also be used to shorten configuration files. While this can be used as part of legal YAML parsing, its use is discouraged by UPWARD files, since context resolut