@toa.io/extensions.exposition 1.0.0-alpha.28 → 1.0.0-alpha.282

package/source/Factory.ts

@@ -1,51 +1,69 @@
 import assert from 'node:assert'
-import { decode } from '@toa.io/generic'
-import { Tenant } from './Tenant'
-import { Gateway } from './Gateway'
-import { Remotes } from './Remotes'
-import { Tree, syntax } from './RTD'
-import { EndpointsFactory } from './Endpoint'
-import { families, interceptors } from './directives'
-import { DirectivesFactory } from './Directive'
-import { Composition } from './Composition'
-import * as root from './root'
-import { Interception } from './Interception'
-import * as http from './HTTP'
-import type { Broadcast } from './Gateway'
+import { createHash } from 'node:crypto'
+import { console, traces, type LevelName, type TracesOptions } from 'openspan'
+import { Tenant } from './Tenant.js'
+import { Gateway } from './Gateway.js'
+import { Remotes } from './Remotes.js'
+import { Tree } from './RTD/index.js'
+import { EndpointsFactory } from './Endpoint.js'
+import { families, interceptors } from './directives/index.js'
+import { DirectivesFactory } from './Directive.js'
+import { Composition } from './Composition.js'
+import * as root from './root.js'
+import { ATOM_GROUP } from './const.js'
+import { Interception } from './Interception.js'
+import * as http from './HTTP/index.js'
+import type { Branch } from './Branch.js'
+import type { syntax } from './RTD/index.js'
+import type { Broadcast } from './Gateway.js'
 import type { Connector, Locator, extensions } from '@toa.io/core'
 
 export class Factory implements extensions.Factory {
-  private readonly boot: Bootloader
+  private readonly host: Host
 
-  public constructor (boot: Bootloader) {
-    this.boot = boot
+  public constructor (host: Host) {
+    this.host = host
   }
 
-  public tenant (locator: Locator, node: syntax.Node): Connector {
-    const broadcast: Broadcast = this.boot.bindings.broadcast(CHANNEL, locator.id)
+  public async tenant (locator: Locator, node: syntax.Node): Promise<Connector> {
+    const broadcast: Broadcast = await this.host.broadcast(CHANNEL, locator.id)
+    const hash = createHash('sha256').update(JSON.stringify(node)).digest('hex')
 
-    return new Tenant(broadcast, locator, node)
+    // no timestamp: the tenant stamps each announcement with its own start time
+    const branch: Omit<Branch, 'timestamp'> = {
+      namespace: locator.namespace,
+      component: locator.name,
+      isolated: locator.namespace === 'identity',
+      node,
+      version: hash
+    }
+
+    return new Tenant(broadcast, branch)
   }
 
-  public service (): Connector | null {
+  public async service (): Promise<Connector | null> {
     assert.ok(process.env.TOA_EXPOSITION_PROPERTIES,
       'TOA_EXPOSITION_PROPERTIES is undefined')
 
-    const options = decode<http.Options>(process.env.TOA_EXPOSITION_PROPERTIES)
-    const broadcast: Broadcast = this.boot.bindings.broadcast(CHANNEL)
-    const server = http.Server.create({ ...options, methods: syntax.verbs })
-    const remotes = new Remotes(this.boot)
+    configureLogs()
+
+    const options = JSON.parse(process.env.TOA_EXPOSITION_PROPERTIES) as http.Options
+    const broadcast: Broadcast = await this.host.broadcast(CHANNEL)
+    const server = http.Server.create({ ...options })
+    const remotes = new Remotes(this.host)
     const node = root.resolve()
     const methods = new EndpointsFactory(remotes)
-    const directives = new DirectivesFactory(families, remotes)
+    const directives = new DirectivesFactory(families, remotes, this.host, options)
     const interception = new Interception(interceptors)
     const tree = new Tree(node, methods, directives)
 
-    const composition = new Composition(this.boot)
+    const composition = new Composition(this.host)
     const gateway = new Gateway(broadcast, tree, interception)
 
     gateway.depends(remotes)
     gateway.depends(composition)
+    // what the directives meter through; one atom per process, connected once
+    gateway.depends(this.host.atom(ATOM_GROUP))
 
     server.attach(gateway.process.bind(gateway))
     server.depends(gateway)
@@ -55,6 +73,35 @@ export class Factory implements extensions.Factory {
 }
 
 const CHANNEL = 'exposition'
+const LOGS_PREFIX = 'TOA_TELEMETRY_LOGS'
+const TRACES_ENV = 'TOA_TELEMETRY_TRACES'
+
+function configureLogs (): void {
+  const globEnv = process.env[LOGS_PREFIX]
+  const level: LevelName = process.env.TOA_DEV === '1' ? 'trace' : 'info'
+  const options = globEnv === undefined ? { level } : JSON.parse(globEnv) as { level?: LevelName }
+
+  console.configure({ level: options.level ?? level })
+
+  const tracesEnv = process.env[TRACES_ENV]
+
+  traces(tracesEnv === undefined ? development() : JSON.parse(tracesEnv) as TracesOptions)
+}
+
+/**
+ * Tracing is off unless it is configured. The console exporter is a local development
+ * mechanism, so it is turned on for `toa dev` and for a boot trace the CLI has already
+ * asked for (`runtime/boot/src/span.js`), and nowhere else — a deployment that wants
+ * traces annotates `telemetry.traces.exporters`.
+ *
+ * The gateway boots without the telemetry extension, hence the copy of
+ * `extensions/telemetry/source/extension.ts`.
+ */
+function development (): TracesOptions {
+  const local = process.env.TOA_DEV === '1' || process.env.TOA_BOOT_TRACE === '1'
+
+  return local ? { exporters: { console: {} } } : {}
+}
 
 // eslint-disable-next-line @typescript-eslint/consistent-type-imports
-export type Bootloader = typeof import('@toa.io/boot')
+export type Host = extensions.Host

package/source/Gateway.ts

@@ -1,15 +1,25 @@
+import assert from 'node:assert'
+import { setTimeout } from 'node:timers/promises'
+import { console } from 'openspan'
 import { type bindings, Connector } from '@toa.io/core'
-import * as http from './HTTP'
-import { rethrow } from './exceptions'
-import type { Interception } from './Interception'
-import type { Method, Parameter, Tree } from './RTD'
-import type { Label } from './discovery'
-import type { Branch } from './Branch'
+import * as http from './HTTP/index.js'
+import { rethrow } from './exceptions.js'
+import { decide } from './Branch.js'
+import type { Interception } from './Interception.js'
+import type { Method, Node, Parameter, Tree, Match } from './RTD/index.js'
+import type { Label } from './discovery.js'
+import type { Branch, Exposed } from './Branch.js'
 
 export class Gateway extends Connector {
   private readonly broadcast: Broadcast
   private readonly tree: Tree
   private readonly interceptor: Interception
+  private readonly branches = new Map<string, Exposed>()
+  private lastMerge = 0
+  private widestGap = 0
+  private lastPing = 0
+  private stopped = false
+  private resolveFirstMerge: (() => void) | null = null
 
   public constructor (broadcast: Broadcast, tree: Tree, interception: Interception) {
     super()
@@ -28,17 +38,20 @@ export class Gateway extends Connector {
     if (interception !== null)
       return interception
 
-    const match = this.tree.match(context.url.pathname)
+    const { node, parameters } = this.match(context)
 
-    if (match === null)
-      throw new http.NotFound('Route not found')
+    if (context.request.method === 'OPTIONS')
+      return await this.explain(node, parameters)
+
+    let verb = context.request.method
 
-    const { node, parameters } = match
+    if (!(verb in node.methods) && verb === 'HEAD' && 'GET' in node.methods)
+      verb = 'GET'
 
-    if (!(context.request.method in node.methods))
+    if (!(verb in node.methods))
       throw new http.MethodNotAllowed()
 
-    const method = node.methods[context.request.method]
+    const method = node.methods[verb]
 
     const interruption = await context.timing.capture('preflight',
       method.directives.preflight(context, parameters)).catch(rethrow)
@@ -55,13 +68,54 @@ export class Gateway extends Connector {
   protected override async open (): Promise<void> {
     await this.discover()
 
-    console.info('Gateway has started and is awaiting resource branches')
+    console.info('Gateway started')
+  }
+
+  /**
+   * Both of these reach into components, and a dependency is torn down only once this has
+   * returned — where `dispose` runs after every one of them already has. The throttling
+   * ticker firing in between calls an endpoint that has just been unbound, and reports the
+   * refusal as a failure to reconcile.
+   */
+  protected override async close (): Promise<void> {
+    this.stopped = true
+
+    this.tree.dispose()
   }
 
   protected override dispose (): void {
     console.info('Gateway is closed')
   }
 
+  private match (context: http.Context): Match {
+    const match = this.tree.match(context.url.pathname)
+
+    if (match === null) {
+      // the route may be missing because an expose has been lost
+      this.reping()
+
+      throw new http.NotFound('Route not found')
+    }
+
+    if (match.node.forward === null)
+      return match
+
+    const destination = match.node.forward.replace(/\/:([^/]+)/g,
+      (_, name) => {
+        const value = match.parameters.find((parameter) => parameter.name === name)?.value
+
+        assert.ok(value !== undefined, `Forwarded parameter '${name}' not found`)
+
+        return `/${value}`
+      })
+
+    const forward = this.tree.match(destination)
+
+    assert.ok(forward !== null, 'Forwarded route not found')
+
+    return forward
+  }
+
   private async call (method: Method, context: http.Context, parameters: Parameter[]): Promise<http.OutgoingMessage> {
     if (context.url.pathname[context.url.pathname.length - 1] !== '/')
       throw new http.NotFound('Trailing slash is required')
@@ -77,21 +131,157 @@ export class Gateway extends Connector {
       .catch(rethrow) as http.OutgoingMessage
   }
 
+  private async explain (node: Node, parameters: Parameter[]): Promise<http.OutgoingMessage> {
+    const body = await node.explain(parameters)
+    const allow = [...Object.keys(node.methods)].join(', ')
+    const headers = new Headers({ allow })
+
+    return { body, headers }
+  }
+
   private async discover (): Promise<void> {
+    const first = new Promise<void>((resolve) => {
+      this.resolveFirstMerge = resolve
+    })
+
     await this.broadcast.receive<Branch>('expose', this.merge.bind(this))
+
+    void this.knock()
+
+    await this.settled(first)
+  }
+
+  /**
+   * A single ping is enough only if every tenant is listening by then, which is
+   * not the case while the deployment is still rolling out.
+   */
+  private async knock (): Promise<void> {
+    for (const delay of KNOCK_DELAYS) {
+      if (this.stopped)
+        return
+
+      if (delay > 0)
+        await setTimeout(delay, undefined, { ref: false })
+
+      await this.ping()
+    }
+  }
+
+  private async ping (): Promise<void> {
+    if (this.stopped)
+      return
+
+    this.lastPing = Date.now()
+
     await this.broadcast.transmit<null>('ping', null)
+      .catch((exception: Error) => console.error('Discovery ping failed',
+        { message: exception.message }))
+  }
+
+  private reping (): void {
+    if (Date.now() - this.lastPing < PING_COOLDOWN)
+      return
+
+    void this.ping()
+  }
+
+  private async settled (first: Promise<void>): Promise<void> {
+    const deadline = Date.now() + SETTLE_TIMEOUT
+    const abort = new AbortController()
+
+    // an uncancelled timer keeps the process alive long after the race is won
+    await Promise.race([first, setTimeout(SETTLE_TIMEOUT, undefined, { signal: abort.signal })])
+      .finally(() => { abort.abort() })
+      .catch(() => {})
+
+    if (this.lastMerge === 0) {
+      console.warn('Discovery timed out waiting for the first expose')
+
+      return
+    }
+
+    while (Date.now() - this.lastMerge < this.quiet()) {
+      if (Date.now() >= deadline)
+        break
+
+      await setTimeout(SETTLE_POLL)
+    }
+  }
+
+  /**
+   * How long the branches must stay quiet before discovery counts as settled.
+   *
+   * Every tenant of a local composition answers the first ping at once, so a short window
+   * is enough; a rolling deployment brings them up seconds apart, and the window grows with
+   * the widest gap seen so far to keep waiting for the ones still starting.
+   */
+  private quiet (): number {
+    const adaptive = this.widestGap * SETTLE_QUIET_FACTOR
+
+    return Math.min(Math.max(adaptive, SETTLE_QUIET_MIN), SETTLE_QUIET_MAX)
   }
 
   private merge (branch: Branch): void {
+    const id = branch.namespace + '.' + branch.component
+
+    const attributes = {
+      namespace: branch.namespace,
+      component: branch.component,
+      version: branch.version
+    }
+
+    const exposed = this.branches.get(id)
+
+    if (exposed !== undefined)
+      switch (decide(exposed, branch)) {
+        // rebuilding an identical branch would only tear down its live endpoints
+        case 'refresh':
+          this.tree.refresh(exposed.nodes)
+          console.trace('Branch refreshed', attributes)
+
+          return
+
+        case 'superseded':
+          console.debug('Branch superseded', {
+            ...attributes,
+            timestamp: branch.timestamp,
+            exposed: exposed.timestamp
+          })
+
+          return
+      }
+
+    let nodes: Node[]
+
     try {
-      this.tree.merge(branch.node, branch)
+      nodes = this.tree.merge(branch.node, branch)
+    } catch (exception: unknown) {
+      const message = exception instanceof Error ? exception.message : 'Unknown error'
 
-      console.info('Resource branch of ' +
-        `'${branch.namespace}.${branch.component}' has been merged`)
-    } catch (exception) {
-      console.error(exception)
+      console.error('Branch merge exception', { message, ...attributes })
+
+      return
     }
+
+    this.branches.set(id, { version: branch.version, timestamp: branch.timestamp, nodes })
+
+    if (this.lastMerge !== 0)
+      this.widestGap = Math.max(this.widestGap, Date.now() - this.lastMerge)
+
+    this.lastMerge = Date.now()
+    this.resolveFirstMerge?.()
+    this.resolveFirstMerge = null
+
+    console.info('Branch merged', attributes)
   }
 }
 
 export type Broadcast = bindings.Broadcast<Label>
+
+const SETTLE_QUIET_MIN = 500
+const SETTLE_QUIET_MAX = 10_000
+const SETTLE_QUIET_FACTOR = 2
+const SETTLE_TIMEOUT = 30_000
+const SETTLE_POLL = 50
+const KNOCK_DELAYS = [0, 500, 1000, 1500]
+const PING_COOLDOWN = 5_000

package/source/HTTP/Context.ts

@@ -1,13 +1,21 @@
+import crypto from 'node:crypto'
+import { buffer } from 'node:stream/consumers'
 import Negotiator from 'negotiator'
-import { Timing } from './Timing'
-import { type Format, formats, types } from './formats'
-import { read } from './messages'
-import type { OutgoingMessage } from './messages'
-import type * as http from 'node:http'
+import { console } from 'openspan'
+import { Timing } from './Timing.js'
+import { type Format, formats, types } from './formats/index.js'
+import { read } from './messages.js'
+import { address } from './address.js'
+import type { OutgoingMessage } from './messages.js'
+import type { IncomingMessage } from './types.js'
 
 export class Context {
+  public readonly id: string
   public readonly authority: string
   public readonly request: IncomingMessage
+
+  /** the client address, read as the deployment says; none otherwise, see `documentation/ip.md` */
+  public readonly ip: string | undefined
   public readonly url: URL
   public readonly subtype: string | null = null
   public readonly encoder: Format | null = null
@@ -19,19 +27,26 @@ export class Context {
     response: []
   }
 
-  public constructor (authority: string, request: IncomingMessage, properties: Properties) {
+  private consumed = false
+
+  // eslint-disable-next-line max-params
+  public constructor (authority: string, request: IncomingMessage, properties: Properties,
+    url: URL) {
     this.authority = authority
     this.request = request
+    this.ip = address(request, properties.ip)
 
-    this.url = new URL(request.url, `https://${request.headers.host}`)
-    this.timing = new Timing(properties.trace)
+    this.id = crypto.randomUUID()
+    // parsed by the server, which had to parse it anyway to know the request is valid
+    this.url = url
+    this.timing = new Timing()
     this.debug = properties.debug
+    this.log(request)
 
-    if (this.debug)
-      this.log(request)
+    const accept = this.request.headers.accept
 
-    if (this.request.headers.accept !== undefined) {
-      const match = SUBTYPE.exec(this.request.headers.accept)
+    if (accept !== undefined) {
+      const match = SUBTYPE.exec(accept)
 
       if (match !== null) {
         const {
@@ -45,45 +60,77 @@ export class Context {
       }
     }
 
-    const negotiator = new Negotiator(this.request)
-    const mediaType = negotiator.mediaType(types)
+    const encoder = negotiate(this.request)
+
+    if (encoder !== undefined)
+      this.encoder = encoder
+  }
+
+  public async buffer (): Promise<Buffer> {
+    this.consumed = true
 
-    if (mediaType !== undefined)
-      this.encoder = formats[mediaType]
+    return await buffer(this.request)
   }
 
-  public async body<T> (): Promise<T> {
-    const value = await read(this)
+  public async body<T>(): Promise<T> {
+    let value = this.consumed ? null : await read(this)
+
+    for (const transform of this.pipelines.body)
+      value = await transform(value)
 
-    return this.pipelines.body.length === 0
-      ? value
-      : this.pipelines.body.reduce((value, transform) => transform(value), value)
+    return value
   }
 
   private log (request: IncomingMessage): void {
-    const message = `${request.method} ${request.url}`
-    const { authorization, ...headers } = request.headers
+    const headers = { ...request.headers }
 
-    if (authorization !== undefined)
-      headers.authorization = authorization.slice(0, authorization.indexOf(' '))
+    if (headers.authorization !== undefined)
+      headers.authorization = SCHEME.exec(headers.authorization)?.[1] ?? '[malformed]'
 
-    console.debug(message, headers)
+    console.debug('Received request', { method: request.method, url: request.url, headers })
   }
 }
 
-export interface IncomingMessage extends http.IncomingMessage {
-  url: string
-  method: string
-}
+/** The scheme of an `authorization` header; the value after it is a credential. */
+const SCHEME = /^(\w{1,32})(?: |$)/
 
 interface Pipelines {
   body: Array<(input: unknown) => unknown>
-  response: Array<(output: OutgoingMessage) => void>
+  response: Array<(output: OutgoingMessage) => void | Promise<void>>
 }
 
 interface Properties {
   debug: boolean
-  trace: boolean
+
+  /** the header the client address is read from; the connection's without one */
+  ip?: string
 }
 
+/**
+ * Negotiation parses the header and sorts the candidates, and the value repeats:
+ * clients send one of a handful of `accept` strings. Bounded, the header is theirs.
+ */
+function negotiate (request: IncomingMessage): Format | undefined {
+  const accept = request.headers.accept ?? ''
+  const known = NEGOTIATED.get(accept)
+
+  if (known !== undefined)
+    return known === NONE ? undefined : known
+
+  const mediaType = new Negotiator(request).mediaType(types)
+  const encoder = mediaType === undefined ? undefined : formats[mediaType]
+
+  if (NEGOTIATED.size >= NEGOTIATED_LIMIT)
+    NEGOTIATED.clear()
+
+  NEGOTIATED.set(accept, encoder ?? NONE)
+
+  return encoder
+}
+
+/** distinguishes "negotiated to nothing" from "not negotiated yet" */
+const NONE = Symbol('not acceptable') as unknown as Format
+const NEGOTIATED = new Map<string, Format>()
+const NEGOTIATED_LIMIT = 1024
+
 const SUBTYPE = /^(?<type>\w{1,32})\/(vnd\.toa\.(?<subtype>\S{1,32})\+)(?<suffix>\S{1,32})$/

package/source/HTTP/Probe.ts

@@ -0,0 +1,112 @@
+import * as http from 'node:http'
+import { console } from 'openspan'
+
+/**
+ * The readiness probe, on a port of its own.
+ *
+ * It cannot share the port the gateway serves traffic on: kubelet speaks HTTP/1.1 only, and a
+ * cleartext HTTP/2 server answers nothing else on its port — there is no ALPN to negotiate with.
+ *
+ * It is the gateway's own readiness, not the process's. The identity composition nested in this
+ * process connects before route discovery has settled, so a probe tied to that composition would
+ * report ready while the gateway still had no routes and no open port.
+ */
+export class Probe {
+  private readonly server = http.createServer()
+  private readonly port: number
+  private readonly path: string
+
+  private ready = false
+  private startedAt = 0
+  private listening = false
+  private skipped = false
+
+  public constructor (port: number, path: string = PATH) {
+    this.port = port
+    this.path = path
+
+    this.server.on('request', (request, response) => this.listener(request, response))
+  }
+
+  public async listen (): Promise<void> {
+    this.startedAt = Date.now()
+
+    try {
+      await new Promise<void>((resolve, reject) => {
+        this.server.once('error', reject)
+        this.server.listen(this.port, () => {
+          this.server.removeListener('error', reject)
+          resolve()
+        })
+      })
+    } catch (error: any) {
+      // processes sharing a host share the port; in a pod nothing else claims it
+      if (error?.code === 'EADDRINUSE') {
+        this.skipped = true
+
+        console.warn('Ready probe port already in use, skipping', { port: this.port })
+
+        return
+      }
+
+      throw error
+    }
+
+    this.listening = true
+
+    // a probe answers while the process runs; it must never be the reason it keeps running
+    this.server.unref()
+  }
+
+  /** The gateway is listening and has routes: answer `200`. */
+  public complete (): void {
+    this.ready = true
+
+    // the IPC signal is not tied to the probe: a process that gave up the shared port is
+    // still ready, and pm2 `wait_ready` would otherwise block until `listen_timeout`
+    process.send?.('ready')
+  }
+
+  public async close (): Promise<void> {
+    this.ready = false
+
+    if (!this.listening || this.skipped)
+      return
+
+    this.listening = false
+
+    // keep-alive connections would otherwise hold the server handle, delaying the exit
+    this.server.closeAllConnections()
+
+    await new Promise<void>((resolve) => this.server.close(() => resolve()))
+  }
+
+  private listener (request: http.IncomingMessage, response: http.ServerResponse): void {
+    if (request.url !== this.path) {
+      response.writeHead(404).end()
+
+      return
+    }
+
+    if (this.ready)
+      response.writeHead(200, { 'cache-control': 'no-store' }).end()
+    else {
+      const remaining = Math.ceil((Date.now() - this.startedAt) / 1000).toString()
+
+      response.writeHead(503, { 'retry-after': remaining }).end()
+    }
+  }
+}
+
+/**
+ * Reserved for this probe. `8001` is the Telemetry readiness probe's, and `toa export` refuses a
+ * port claimed twice — `toa mono` and a local run put every service in one process.
+ */
+export const PROBE = 8004
+export const PATH = '/.ready'
+
+/**
+ * The initial delay of the readiness probe. The server does not sleep for it: whoever
+ * probes is the one that waits, and doing it here as well only delayed the process twice.
+ */
+export const DELAY = 3 // seconds