witchcraft-go-server

README

witchcraft-go-server

witchcraft-go-server is a Go implementation of a Witchcraft server. It provides a way to quickly and easily create servers that work in the Witchcraft ecosystem.

Implementation

Configuration

A witchcraft server is provided with install configuration and runtime configuration. Install configuration specifies configuration values that are static -- they are read in once at server startup and are known to never change. Runtime configuration values are considered refreshable. When file-based configuration is used, whenever the runtime configuration file is updated its contents are loaded and the corresponding values are refreshed. See the section on refreshable configuration for more information on this.

The default configuration uses file-based configuration with the install configuration at var/conf/install.yml and runtime configuration at var/conf/runtime.yml. It is possible to use code to specify different sources of configuration (for example, in-memory providers).

witchcraft-server also supports using encrypted-config-value to automatically decrypt encrypted configuration values. The default configuration expects a key file to be at var/conf/encrypted-config-value.key. It is possible to use code to specify a different source for the key (or to specify that no key should be used). If the configuration does not contain encrypted values, any specified ECV key will not be read. If the install configuration contains encrypted values but the encryption key is missing or malformed, the server will fail to start. If the runtime config contains encrypted values but fails to decrypt them, a warning will be logged and the encrypted values passed to the server.

witchcraft-server defines base configuration for its install and runtime configuration. Servers that want to provide their own install and/or runtime configuration should embed the base configuration structs within the definition of their configuration structs.

Route registration

A witchcraft server is backed by a wrouter.Router and allows authors to register route handlers on the server. The router uses a specific format for path templates to specify path parameters and has rules around the kinds of paths that can be matched. witchcraft is opinionated about the path formats and does not support registering paths that cannot be expressed using its template rules. All witchcraft routes are configured to emit request logs and trace logs and update metrics for the requests using built-in middleware. The context.Context for the http.Request provided to the handlers is configured with all of the standard loggers (service logger, event logger, trace logger, etc.).

When registering routes on the router, it is also possible to specify path/header/query param keys that should be considered "safe" or "forbidden" when used as parameters in logging. These are combined with the default set of safe and forbidden header parameters defined by the req2log package in witchcraft-go-logging.

Liveness, readiness, and health

witchcraft-server registers the endpoints /status/liveness, /status/readiness and /status/health to report the server's liveness, readiness and health. By default, these endpoints use a built-in provider that reports liveness, readiness and health based on the state of the server. It is possible to configure the liveness and readiness providers in code, and health status providers can also be added via code (health supports specifying multiple sources to report health, and the server's built-in health status provider will always be one of them).

The default behavior serves both the user-registered endpoints and the status endpoints from the same server. However, if a "management port" is specified in the server's install configuration and its value differs from the "port" value in configuration, then witchcraft-server starts a second management server on the specified port and serves the status endpoints on that port. This can be useful in scenarios where all of the traffic to the main endpoints require client certificates for TLS but the status endpoints need to be served without requiring client TLS certificates.

Context path

If context-path is specified in the install configuration, all of the routes registered on the server will be prefixed with the specified context-path.

Security

witchcraft-server only supports HTTPS. The TLS client authentication type is configurable in code. The base install configuration has fields to specify the location of server key and certificate material for TLS connections.

Although it is not possible to run witchcraft-server using HTTP, it is possible to configure the server in code to use a generated self-signed certificate on start-up. Running the server in this mode and connecting to it using TLS without server certificate verification (equivalent of curl -k or an http.Transport with TLSClientConfig: &tls.Config{InsecureSkipVerify: true}) provides an analog to using HTTP, with the benefit that the traffic itself is still encrypted.

Logging

witchcraft-server is configured with service, event, metric, request and trace loggers from the witchcraft-go-logging project and emits structured JSON logs using zap as the logger implementation. The default behavior emits logs to the var/log directory (var/log/service.log, var/log/request.log, etc.) unless the server is run in a Docker container, in which case the logs are always emitted to stdout. The use-console-log property in the install configuration can also be set to "true" to always output logs to stdout. The runtime configuration supports configuring the log output level for service logs.

The context.Context provided to request handlers is configured with all of the standard loggers (service logger, event logger, trace logger, etc.). All of the handlers are also configured to emit request logs and trace logs.

Service logger origin

By default, the origin field of the service logger is set to be the package path of the package in which the witchcraft-server is started. For example, if the server is started in the file github.com/palantir/project/server/server.go, the origin for all service log lines will be github.com/palantir/project/server.

It is possible to configure the origin to be a different value using code. The origin can be specified to be a string constant or a function can be used that returns a specific package path based on supplied parameters (for example, the function can specify that the caller package's parent package should be used as the origin). The origin can also be set to empty, in which case it is omitted from the log output.

Trace IDs and instrumentation

witchcraft-server supports zipkin-compatible tracing and ensures that every request is instrumented for tracing. witchcraft-server also recognizes that some code will use trace IDs without necessarily using full zipkin-compatible spans, so some allowances are made to support this scenario.

The built-in witchcraft-server middleware that registers loggers on the context also ensures that a zipkin span is started. If the incoming request header has valid zipkin span information (that is, it specifies both a X-B3-TraceId and X-B3-SpanId in the header), then the span created by the middleware is a child span of the incoming span. If the incoming request does not have a trace ID header, a new root span is created. If the header specifies a trace ID but not a span ID, the middleware creates a new root zipkin span, but ensures that the trace ID of the created span matches what is specified in the header. If an incoming request is routed to a registered endpoint, the built-in router middleware will create another span (which is a child span of the one created by the request middleware) whose span name is the HTTP method and template for the endpoint.

The trace information generated by the middleware is set on the header and will be visible to subsequent handlers. If the request specifies a X-B3-Sampled header, the value specified in that header is used to determine sampling. If this header is not present, whether or not the trace is sampled is determined by the sampling source configured for the witchcraft-server (by default, all traces are sampled). If a trace is not sampled, witchcraft-server will not generate any trace log output for it. However, the infrastructure will still perform all of the trace-related operations (such as creating child spans and setting span information on headers). The install configuration field trace-sample-rate represents a float between 0 and 1 (inclusive) to control the proportion of traces sampled by default. If the WithTraceSampler server option is provided, it overrides this configuration.

witchcraft-server also ensures that the context for every request has a trace ID. After the logging middleware executes, the request is guaranteed to have a trace ID (either from the incoming request or from the newly generated root span), and that trace ID is registered on the context. The witchcraft.TraceIDFromContext(context.Context) string function can be used to retrieve the trace ID from the context.

Creating new spans/trace log entries

Use the wtracing.StartSpanFromContext function to start a new span. This function will create a new span that is a child span of the span in the provided context. Defer the Finish() function of the returned span to ensure that the span is properly marked as finished (the "finish" operation will also generate a trace log entry if the span is sampled).

Middleware

witchcraft-server supports registering middleware to perform custom handling/augmenting of incoming requests. There are 2 different kinds of middleware: request and route middleware.

Request middleware is executed on every request received by the server. The function signature for request middleware is func(rw http.ResponseWriter, r *http.Request, next http.Handler). Request middleware is the most common kind of middleware. The server has built-in request middleware that adds a panic handler, sets the loggers and trade ID on the request context and updates request-related metrics. Any user-supplied request middleware is run after the built-in request middleware in the order in which they were added (which means that the context has all of the loggers configured). Request middleware is run before the request is handled by the router, which means that it is possible to rewrite the URL and other properties of the request and the router will route the modified request. However, note that the built-in logging middleware extracts the UID, SID, TokenID and TraceID from the request and sets them on the loggers before user-provided middleware is invoked, so if the user-defined middleware modifies the header in a manner that would change any of these values, the middleware should also update the request context to have loggers that use the updated values.

Route middleware is only executed on the routes that are registered on the router -- they wrap the handler registered on the route, so they are executed after the path has been matched and the handler for the router has been located and the path parameters have been extracted and set on the context. The function signature for route middleware is func(rw http.ResponseWriter, r *http.Request, reqVals RequestVals, next RouteRequestHandler). The RequestVals struct stores the path template for the route along with the path parameters and their values. The server has built-in route middleware that records a request log entry after the request has completed and creates a trace log span and logs a trace log entry after the request has completed. Route middleware is run after all of the request middleware has run, and any user-supplied route middleware is run after the built-in route middleware in the order in which they were added. Because router middleware is executed after the routing has been determined, changing the URL of the request will not change the handler that is invoked or the path parameter values that have been extracted/stored (although it may still impact behavior based on the content of the actual handler that is registered). In general, most users will likely use request middleware rather than route middleware. However, if users want to only execute middleware on matched routes and want route-specific information such as the unrendered path template and the path parameter values, then route middleware should be used.

Long-running execution not associated with a route

In some instances, a server may want a long-running task not associated with an endpoint. For example, the server may want a long-running goroutine that performs an operation at some interval for the lifetime of the server.

It is recommended that such goroutines be launched in the initialization function provided to witchcraft.With and use the ctx Context as its context. This context has the same lifecycle as the server and has all of the configured loggers (service loggers, metric loggers, etc.) already configured on it.

The provided context does not have a span or trace ID associated with it. If a trace ID is desired, create a new span with wtracing.StartSpanFromContext and the provided context to derive a new context that has a new root span associated with it. This function also updates any loggers in the context to use the new trace ID (for example, service loggers will include the trace ID).

Metrics

witchcraft-server initializes a metrics registry that uses the github.com/palantir/pkg/metrics package (which uses github.com/rcrowley/go-metrics internally) to track metrics for the server. All of the tracked metrics are emitted as metric log entries once every metric emit interval, which is 60 seconds by default (and can be configured to be a custom interval in the install configuration).

By default, witchcraft-server captures various Go runtime metrics (such as allocations, number of running goroutines, etc.) at the same frequency as the metric emit frequency. The collection of Go runtime statistics can be disabled with the WithDisableGoRuntimeMetrics server method.

SIGQUIT handling

witchcraft-server sets up a SIGQUIT handler such that, if the program is terminated using a SIGQUIT signal (kill -3), a goroutine dump is written as a diagnostic.1 log. This behavior can be disabled using server.WithDisableSigQuitHandler. If server.WithSigQuitHandlerWriter is used, the stacks will also be written in their unparsed form to the provided writer.

Shutdown signal handling

witchcraft-server attempts to drain active connections and gracefully shut down by calling server.Shutdown upon receiving a SIGTERM or SIGINT signal. This behavior can be disabled using server.WithDisableShutdownSignalHandler.

Example server initialization

Basic production server

The following is an example program that launches a witchcraft-server that registers a GET /myNum endpoint that returns a randomly generated number encoded as JSON:

package main

import (
	"context"
	"math/rand"
	"net/http"

	"github.com/palantir/witchcraft-go-server/rest"
	"github.com/palantir/witchcraft-go-server/witchcraft"
	"github.com/palantir/witchcraft-go-server/witchcraft/refreshable"
	"github.com/palantir/witchcraft-go-server/wrouter"
)

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			if err := registerMyNumEndpoint(info.Router); err != nil {
				return nil, err
			}
			return nil, nil
		}).
		Start(); err != nil {
		panic(err)
	}
}

func registerMyNumEndpoint(router wrouter.Router) error {
	return router.Get("/myNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, rand.Intn(100), http.StatusOK)
	}))
}

Creating a witchcraft-server starts with the witchcraft.NewServer function, which returns a new witchcraft server with default configuration. The *witchcraft.Server struct has various With* functions that can be used to configure the server, and the Start() function starts the server using the specified configuration.

The WithInitFunc(InitFunc) function is used to register routes on the server. The initialization function provided to WithInitFunc is of the type witchcraft.InitFunc, which has the following definition: type InitFunc func(ctx context.Context, info InitInfo) (cleanup func(), rErr error).

The ctx provided to the function is valid for the duration of the server and has loggers configured on it. The info struct contains fields that can be used to initialize various state and configuration for the server -- refer to the InitInfo documentation for more information.

In this example, a "GET" endpoint is registered on the router using the "/myNum" path, and rest package is used to write a JSON response.

This example server uses all of the witchcraft defaults -- it looks for install configuration in var/conf/install.yml and uses config.Install as its type, looks for runtime configuration in var/conf/runtime.yml and uses config.Runtime as its type, and looks for an encrypted-config-value key in var/conf/encrypted-config-value.key. The install configuration must also specify paths to key and certificate files to use for TLS.

Basic local/test server

The defaults for the server make sense for a production environment, but can make running the server locally (or in tests) cumbersome. We can modify the main function as follows to configure the witchcraft server to use in-memory defaults:

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			if err := registerMyNumEndpoint(info.Router); err != nil {
				return nil, err
			}
			return nil, nil
		}).
		WithSelfSignedCertificate().
		WithECVKeyProvider(witchcraft.ECVKeyNoOp()).
		WithRuntimeConfig(config.Runtime{}).
		WithInstallConfig(config.Install{
			ProductName: "example-app",
			Server: config.Server{
				Port: 8100,
			},
			UseConsoleLog: true,
		}).
		Start(); err != nil {
		panic(err)
	}
}

The WithSelfSignedCertificate() function configures the server to start using a generated self-signed certificate, which removes the need to specify server TLS material. The WithECVKeyProvider(witchcraft.ECVKeyNoOp()) function configures the server to use an empty ECV key source. The WithRuntimeConfig(config.Runtime{}) function configures the server to use the provided runtime configuration (in this case, it is empty), and the WithInstallConfig function specifies the install configuration that should be used (it specifies that port 8100 should be used and that log output should go to STDOUT).

With this configuration, the program can be run using go run:

➜ go run main.go
{"level":"INFO","time":"2018-11-27T05:47:02.013456Z","message":"Listening to https","type":"service.1","origin":"github.com/palantir/witchcraft-go-server/app_example","params":{"address":":8100","server":"example-app"}}

Issuing a request to this server using curl produces the expected response (note that the -k option is used to skip certificate verification because the server is using a self-signed certificate):

➜ curl -k https://localhost:8100/myNum
81

You can also observe that the server emits trace and request logs based on receiving this request:

{"time":"2018-11-27T05:47:28.313585Z","type":"trace.1","span":{"traceId":"7e43bde2647413fc","id":"01228e628b3b3d22","name":"GET /myNum","parentId":"7e43bde2647413fc","timestamp":1543297648313551,"duration":29000}}
{"time":"2018-11-27T05:47:28.313719Z","type":"request.2","method":"GET","protocol":"HTTP/2.0","path":"/myNum","status":200,"requestSize":0,"responseSize":3,"duration":146,"traceId":"7e43bde2647413fc","params":{"Accept":"*/*","User-Agent":"curl/7.54.0","X-B3-Parentspanid":"7e43bde2647413fc","X-B3-Sampled":"1","X-B3-Spanid":"01228e628b3b3d22","X-B3-Traceid":"7e43bde2647413fc"}}
{"time":"2018-11-27T05:47:28.313802Z","type":"trace.1","span":{"traceId":"7e43bde2647413fc","id":"7e43bde2647413fc","name":"witchcraft-go-server request middleware","timestamp":1543297648313496,"duration":304000}}

Server using install configuration

The previous examples used the built-in install configuration. Most real servers will use custom install configuration that specifies configuration for the server. Any struct can be used as install configuration, but it must support being unmarshaled as YAML and must embed the config.Install struct. The install configuration is loaded once when the server starts (it is never reloaded), so only values that are static for the lifetime of the server should be specified in this configuration.

The following example modifies the previous example so that the endpoint returns the number defined in the install configuration instead of a random number:

package main

import (
	"context"
	"net/http"

	"github.com/palantir/witchcraft-go-server/config"
	"github.com/palantir/witchcraft-go-server/rest"
	"github.com/palantir/witchcraft-go-server/witchcraft"
	"github.com/palantir/witchcraft-go-server/witchcraft/refreshable"
	"github.com/palantir/witchcraft-go-server/wrouter"
)

type AppInstallConfig struct {
	config.Install `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			if err := registerMyNumEndpoint(info.Router, info.InstallConfig.(AppInstallConfig).MyNum); err != nil {
				return nil, err
			}
			return nil, nil
		},
		).
		WithSelfSignedCertificate().
		WithECVKeyProvider(witchcraft.ECVKeyNoOp()).
		WithRuntimeConfig(config.Runtime{}).
		WithInstallConfigType(AppInstallConfig{}).
		WithInstallConfig(AppInstallConfig{
			Install: config.Install{
				ProductName: "example-app",
				Server: config.Server{
					Port: 8100,
				},
				UseConsoleLog: true,
			},
			MyNum: 13,
		}).
		Start(); err != nil {
		panic(err)
	}
}

func registerMyNumEndpoint(router wrouter.Router, num int) error {
	return router.Get("/myNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, num, http.StatusOK)
	}))
}

This example defines the AppInstallConfig struct, which embeds config.Install and also defines a MyNum field. The WithInstallConfigType(AppInstallConfig{}) function call is added to specify AppInstallConfig{} as the install struct and the initialization function logic is modified to convert the provided installConfig interface{} into an AppInstallConfig and uses the MyNum value as the value that is returned by the endpoint. The WithInstallConfig function is also updated to use configuration that specifies a value for MyNum.

Running the updated program using go run main.go and issuing curl -k https://localhost:8100/myNum returns 13.

A real program will generally read runtime configuration from disk rather than specifying it directly in code. We can modify the example above to do this by simply removing the WithInstallConfig call:

package main

import (
	"context"
	"net/http"

	"github.com/palantir/witchcraft-go-server/config"
	"github.com/palantir/witchcraft-go-server/rest"
	"github.com/palantir/witchcraft-go-server/witchcraft"
	"github.com/palantir/witchcraft-go-server/witchcraft/refreshable"
	"github.com/palantir/witchcraft-go-server/wrouter"
)

type AppInstallConfig struct {
	config.Install `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			if err := registerMyNumEndpoint(info.Router, info.InstallConfig.(AppInstallConfig).MyNum); err != nil {
				return nil, err
			}
			return nil, nil
		},
		).
		WithSelfSignedCertificate().
		WithECVKeyProvider(witchcraft.ECVKeyNoOp()).
		WithInstallConfigType(AppInstallConfig{}).
		Start(); err != nil {
		panic(err)
	}
}

func registerMyNumEndpoint(router wrouter.Router, num int) error {
	return router.Get("/myNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, num, http.StatusOK)
	}))
}

By default, the install configuration is read from var/conf/install.yml. Create a file at that path relative to the Go file and provide it with the YAML content for the configuration:

product-name: "example-app"
use-console-log: true
server:
  port: 8100
my-num: 77

Running the updated program using go run main.go and issuing curl -k https://localhost:8100/myNum returns 77.

Server using runtime configuration

Runtime configuration is similar to install configuration. The main difference is that runtime configuration supports reloading configuration. When file-based runtime configuration is used, whenever the configuration file is updated, the associated values are updated as well.

The following example defines a custom runtime configuration struct and returns the refreshable int value in the runtime from its endpoint (the example uses a basic in-memory install configuration for simplicity):

package main

import (
	"context"
	"net/http"

	"github.com/palantir/witchcraft-go-server/config"
	"github.com/palantir/witchcraft-go-server/rest"
	"github.com/palantir/witchcraft-go-server/witchcraft"
	"github.com/palantir/witchcraft-go-server/witchcraft/refreshable"
	"github.com/palantir/witchcraft-go-server/wrouter"
)

type AppRuntimeConfig struct {
	config.Runtime `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			myNumRefreshable := refreshable.NewInt(info.RuntimeConfig.Map(func(in interface{}) interface{} {
				return in.(AppRuntimeConfig).MyNum
			}))
			if err := registerMyNumEndpoint(info.Router, myNumRefreshable); err != nil {
				return nil, err
			}
			return nil, nil
		},
		).
		WithSelfSignedCertificate().
		WithECVKeyProvider(witchcraft.ECVKeyNoOp()).
		WithInstallConfig(config.Install{
			ProductName: "example-app",
			Server: config.Server{
				Port: 8100,
			},
			UseConsoleLog: true,
		}).
		WithRuntimeConfigType(AppRuntimeConfig{}).
		Start(); err != nil {
		panic(err)
	}
}

func registerMyNumEndpoint(router wrouter.Router, numProvider refreshable.Int) error {
	return router.Get("/myNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, numProvider.CurrentInt(), http.StatusOK)
	}))
}

The refreshable configuration warrants some closer examination. Note that the registerMyNumEndpoint takes a numProvider refreshable.Int as an argument rather than an int and returns the result of CurrentInt(). Conceptually, the numProvider is guaranteed to always return the current value of the number specified in the runtime configuration. Using this pattern removes the need for writing code that listens for updates -- the code can simply assume that the provider always returns the most recent value. refreshable.Int and refreshable.String are helper types that provide functions that return the current value of the correct type. For types without helper functions, the general refreshable.Refreshable should be used, and the interface{} returned by Current() must be explicitly converted to the proper target type (this is required because Go does not support generics/templatization).

The numProvider provided to registerMyNumEndpoint is derived by applying a mapping function to the runtimeConfig refreshable.Refreshable parameter. runtimeConfig.Map is provided with a function that, given an updated runtime configuration, returns the portion of the configuration that is required. The input to the mapping function must be explicitly cast to the runtime configuration type (in this case, in.(AppRuntimeConfig)), and then the relevant section can be accessed (or derived) and returned. The result of the Map function is a Refreshable that returns the mapped portion. In this case, because we know the result will always be an int, we wrap the returned Refreshable in a refreshable.NewInt call, which provides the convenience function CurrentInt() that performs the type conversion of the result to an int.

By default, the runtime configuration is read from var/conf/runtime.yml. Create a file at that path relative to the Go file and provide it with the YAML content for the configuration:

my-num: 99

Running the updated program using go run main.go and issuing curl -k https://localhost:8100/myNum returns 99. While the program is still running, update the content of the file to be my-num: 88, save it, then run the curl command again. The output is 88.

Full server example

The following is an example of a server that defines and uses both custom install and runtime configuration:

package main

import (
	"context"
	"net/http"

	"github.com/palantir/witchcraft-go-server/config"
	"github.com/palantir/witchcraft-go-server/rest"
	"github.com/palantir/witchcraft-go-server/witchcraft"
	"github.com/palantir/witchcraft-go-server/witchcraft/refreshable"
	"github.com/palantir/witchcraft-go-server/wrouter"
)

type AppInstallConfig struct {
	config.Install `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

type AppRuntimeConfig struct {
	config.Runtime `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

func main() {
	if err := witchcraft.NewServer().
		WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
			if err := registerInstallNumEndpoint(info.Router, info.InstallConfig.(AppInstallConfig).MyNum); err != nil {
				return nil, err
			}

			myNumRefreshable := refreshable.NewInt(info.RuntimeConfig.Map(func(in interface{}) interface{} {
				return in.(AppRuntimeConfig).MyNum
			}))
			if err := registerRuntimeNumEndpoint(info.Router, myNumRefreshable); err != nil {
				return nil, err
			}
			return nil, nil
		},
		).
		WithInstallConfigType(AppInstallConfig{}).
		WithRuntimeConfigType(AppRuntimeConfig{}).
		WithSelfSignedCertificate().
		WithECVKeyProvider(witchcraft.ECVKeyNoOp()).
		Start(); err != nil {
		panic(err)
	}
}

func registerInstallNumEndpoint(router wrouter.Router, num int) error {
	return router.Get("/installNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, num, http.StatusOK)
	}))
}

func registerRuntimeNumEndpoint(router wrouter.Router, numProvider refreshable.Int) error {
	return router.Get("/runtimeNum", http.HandlerFunc(func(rw http.ResponseWriter, req *http.Request) {
		rest.WriteJSONResponse(rw, numProvider.CurrentInt(), http.StatusOK)
	}))
}

With var/conf/install.yml:

product-name: "example-app"
use-console-log: true
server:
  port: 8100
my-num: 7

And var/conf/runtime.yml:

my-num: 13

Querying installNum returns 7, while querying runtimeNum returns 13:

➜ curl -k https://localhost:8100/installNum
7
➜ curl -k https://localhost:8100/runtimeNum
13

In a production server, WithSelfSignedCertificate() and WithECVKeyProvider(witchcraft.ECVKeyNoOp()) would not be called and the proper security and key material would exist in their expected locations.

Refreshable configuration

The runtime configuration for witchcraft-server uses the refreshable.Refreshable interface. Conceptually, a Refreshable is a container that holds a value of a specific type that may be updated/refreshed. The following is the interface definition for Refreshable:

type Refreshable interface {
	// Current returns the most recent value of this Refreshable.
	Current() interface{}

	// Subscribe subscribes to changes of this Refreshable. The provided function is called with the value of Current()
	// whenever the value changes.
	Subscribe(consumer func(interface{})) (unsubscribe func())

	// Map returns a new Refreshable based on the current one that handles updates based on the current Refreshable.
	Map(func(interface{}) interface{}) Refreshable
}

The runtimeConfig refreshable.Refreshable parameter provided to the initialization function specified using WithInitFunc stores the latest unmarshaled runtime configuration as its current value, and the type of the value is specified using the WithRuntimeConfigType function (if this function is not called, config.Runtime is used as the default type).

For example, for the call:

witchcraft.NewServer().
    WithInitFunc(func(ctx context.Context, info witchcraft.InitInfo) (func(), error) {
        return nil, nil
    }).
    WithRuntimeConfigType(AppRuntimeConfig{})

The WithRuntimeConfigType(AppRuntimeConfig{}) function specifies that the type of the runtime configuration is AppRuntimeConfig, so the value returned by runtimeConfig.Current() in WithInitFunc will have the type AppRuntimeConfig. Because Go does not have a notion of generics, the author must make this association manually and perform the conversion of the current value into the desired type when using it (for example, runtimeConfig.Current().(AppRuntimeConfig)).

The Refreshable interface supports using the Map function to derive a new refreshable based on the value of the current refreshable. This allows downstream functions that are only interested in a subset of the refreshable to observe just the relevant portion.

For example, consider the AppRuntimeConfig definition:

type AppRuntimeConfig struct {
	config.Runtime `yaml:",inline"`

	MyNum int `yaml:"my-num"`
}

A downstream function may only be interested in updates to the MyNum variable -- if updates to config.Runtime are not relevant to the function, there is no need to subscribe to it. The following code derives a new Refreshable from the runtimeConfig refreshable:

myNumRefreshable := runtimeConfig.Map(func(in interface{}) interface{} {
    return in.(AppRuntimeConfig).MyNum
})

The Current() function for myNumRefreshable returns the MyNum field of in.(AppRuntimeConfig), and the derived Refreshable is only updated when the derived value changes. Accessing a field is the most common usage of Map, but any arbitrary logic can be performed in the mapping function. Just note that the mapping will be performed whenever the parent refreshable is updated and the result will be compared using reflect.DeepEqual.

The general Refreshable interface returns an interface{} and its result must always be converted to the actual underlying type. However, if a Refreshable is known to return an int, string or bool, convenience wrapper types are provided to return typed values. For example, refreshable.NewInt(in Refreshable) returns a refreshable.Int, which is defined as:

type Int interface {
	Refreshable
	CurrentInt() int
}

The CurrentInt() function returns the current value converted to an int, which makes it easier to use in code and alleviates the need for clients to manually remember the type stored in the Refreshable.

If a Refreshable with a particular value/type is used widely throughout a code base, it may make sense to define a similar interface so that clients do not have to manually track the type information. For example, a typed Refreshable for AppRuntimeConfig can be defined as follows:

type RefreshableAppRuntimeConfig interface {
	Refreshable
	CurrentAppRuntimeConfig() AppRuntimeConfig
}

type refreshableAppRuntimeConfig struct {
	Refreshable
}

func (r refreshableAppRuntimeConfig) CurrentAppRuntimeConfig() AppRuntimeConfig {
	return rt.Current().(AppRuntimeConfig)
} 

Updating refreshable configuration: provider-based vs. push-based

The "provider" model of configuration updates takes the philosophy that executing code simply needs the most up-to-date value of a Refreshable when it executes. This model makes the most sense when the value is read whenever an endpoint is executed or when a long-running or periodically executed background task executes. In these scenarios, the latest value of the Refreshable is only needed when the logic executes. This update model is typically the most common, and is achieved by passing down specific Refreshable providers for the required values to the handlers/routines.

However, in some cases, an application may want to be notified of every update to a field and react to that update immediately -- for example, if updating a specific configuration field triggers an expensive computation that should happen immediately, the logic wants to be notified as soon as the update is made.

In this scenario, the Subscribe function should be used for the Refreshable that has the value for which updates are needed. For example, consider the following configuration:

type AppRuntimeConfig struct {
	config.Runtime `yaml:",inline"`

	AssetURLs []string `yaml:"asset-urls"`
}

The AssetURLs field specifies URLs that should be downloaded by the program whenever the value is updated. This can be handled as follows:

unsubscribe := runtimeConfig.Map(func(in interface{}) interface{} {
    return in.(AppRuntimeConfig).AssetURLs
}).Subscribe(func(in interface{}) {
	assetURLs := in.([]string)
	// perform work
})
// unsubscribe should be deferred or stored and run at shutdown 

The Map function returns a new Refreshable that updates only when the AssetURLs field is updated, and the Subscribe function subscribes a listener that performs work as soon as the value is updated. This ensures that the logic is run as soon as the value is refreshed every time the value is updated.

License

This project is made available under the Apache 2.0 License.