@greendrake/service-state 0.4.0 → 0.6.0

package/README.md

@@ -44,15 +44,29 @@ const unit = (name: string) => ({
     check: () => systemd.isActive(name).then(on => (on ? 'ON' : 'OFF'))
 })
 
-// Each machine is named for the wire; the registry — any EventSink, and
-// @greendrake/rpc-server's ConnRegistry is one — is where its pushes go.
-const machines = await Promise.all([createServiceStateMachine('web', unit('web.service'), registry), createServiceStateMachine('worker', unit('worker.service'), registry)])
+// Each machine is named for the wire; the registry — any EventSink that is an
+// Audience too, and @greendrake/rpc-server's ConnRegistry is both — is where
+// its pushes go and who is there to hear them; and the last argument is how
+// old a reading may get before the machine takes another, in milliseconds.
+const machines = await Promise.all([createServiceStateMachine('web', unit('web.service'), registry, 30_000), createServiceStateMachine('worker', unit('worker.service'), registry, 30_000)])
 
 const methods = { ...serviceStateMethods(machines), ...whateverElse }
+
+// On shutdown, stop the re-reads and the listening for watchers.
+machines.forEach(machine => machine.close())
 ```
 
 The name is what a call's `{ service }` reaches the machine by and what its every status carries. The factory reads the service before returning, so the machine never reports a state it has not verified — one that assumed `OFF` at boot would have every dashboard showing `OFF` for a service that is up.
 
+Nor does it keep reporting one it verified long ago, to anybody who needs to know. Nothing tells the machine when a service is changed by other hands — a unit stopped from a shell, a schedule that turns it off overnight — so it reads the service again when that is needed, and only then:
+
+- **Asked.** `status` answers from the latest reading while it is younger than `maxAgeMs`; one that old is taken again first, and the answer is what it found. A dashboard opened after a quiet spell is shown what the service is, not what it was.
+- **Watched.** While the sink's audience has anybody in it — a socket open to be pushed to — the machine re-reads whenever its latest reading turns `maxAgeMs` old, and pushes what changed, so a dashboard left open follows the service. The first to arrive has a stale reading taken again at once; the last to leave stops the re-reads.
+
+With nobody asking and nobody watching, the service is left alone. A transition that lands counts as a reading, and so does a `refresh`; one asked while a read is under way joins it, unless a transition has begun since that read went out — which makes it a reading of the past, so a fresh one goes out instead. A transition under way and `ERROR` are the machine's own word rather than a reading, and are not read past — `ERROR` until somebody asks for a `refresh`, because reading past it would take the why away before anyone saw it. A transition begun while a read is out goes ahead, and the read's answer, which describes the service before it, is dropped. `close()` stops the re-reads and the listening for watchers, for a process shutting down.
+
+`check` must settle promptly: a `status` or a `refresh` can wait on it, and the next re-read is timed from when it settles, so a check that never settles leaves the state as it last stood.
+
 `start` and `stop` **return as soon as the transition has begun**, with the in-flight status. Where it ends up is the machine's next state, pushed to everyone — not the caller's answer, because everyone watching needs it equally and one of them happening to have asked changes nothing. A hook that rejects lands the machine in `ERROR` carrying its message.
 
 One transition at a time, and each only from the state it is defined for. Anything else — `start` while already `ON`, `refresh` mid-transition — is refused with `SERVICE_BUSY`. A client showing the state never triggers it, because the state it is showing offers no such button.

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@greendrake/service-state",
-    "version": "0.4.0",
+    "version": "0.6.0",
     "description": "A backend service as a state machine (OFF/STARTING/ON/STOPPING/ERROR): the wire contract both halves share, and the server-side machine behind it",
     "license": "MIT",
     "author": "Eugene Greendrake <eugene@greendrake.nz>",
@@ -38,7 +38,7 @@
     },
     "prettier": "@greendrake/dev-config/prettier",
     "peerDependencies": {
-        "@greendrake/rpc-server": "0.4.0"
+        "@greendrake/rpc-server": "0.6.0"
     },
     "peerDependenciesMeta": {
         "@greendrake/rpc-server": {
@@ -46,8 +46,8 @@
         }
     },
     "devDependencies": {
-        "@greendrake/dev-config": "0.4.0",
-        "@greendrake/rpc-server": "0.4.0",
+        "@greendrake/dev-config": "0.6.0",
+        "@greendrake/rpc-server": "0.6.0",
         "@types/bun": "^1.3.5",
         "typescript": "6.0.3"
     }

package/src/server.ts

@@ -1,12 +1,15 @@
 import { ApiError } from '@greendrake/rpc-server'
-import type { AuthRequirement, EventSink, MethodDef } from '@greendrake/rpc-server'
+import type { Audience, AuthRequirement, EventSink, MethodDef } from '@greendrake/rpc-server'
 import { SERVICE_STATE_EVENT, TRANSITIONS } from './main'
 import type { ServiceCommand, ServiceRef, ServiceState, ServiceStateMethods, ServiceStatus } from './main'
 
 // What a concrete service plugs in. `start` and `stop` resolve once the
 // service is up or down; rejecting puts the machine in ERROR with the
 // rejection's message. `check` reports what the service is actually doing
-// right now, read from the service rather than remembered.
+// right now, read from the service rather than remembered, and must settle
+// promptly: a status or a refresh can wait on it, and the next re-read is
+// timed from when it settles, so a check that never does leaves the state as
+// it last stood.
 export interface ServiceImplementation {
     start(): Promise<void>
     stop(): Promise<void>
@@ -17,15 +20,22 @@ export interface ServiceStateMachine {
     // What the service is called on the wire: what a call names to reach it,
     // and what each of its pushes carries.
     readonly name: string
-    status(): ServiceStatus
+    // What the service is: its last reading, or what its last transition left
+    // it as — read again first when that is `maxAgeMs` old (see the factory).
+    status(): Promise<ServiceStatus>
     // Return as soon as the transition has begun, with the in-flight status.
     // Where it ends up is the machine's next state, pushed to everyone — not
     // this caller's answer, because everyone watching needs it equally and one
     // of them happening to have asked changes nothing.
     start(): ServiceStatus
     stop(): ServiceStatus
-    // Re-read the service and adopt what it reports.
+    // Re-read the service and adopt what it reports, joining a read already
+    // under way — one begun since the last transition — rather than asking
+    // the service again.
     refresh(): Promise<ServiceStatus>
+    // Stop re-reading the service and listening for watchers, for a process
+    // shutting down.
+    close(): void
 }
 
 // The command a state is in no position to take. One code covers both cases —
@@ -41,15 +51,41 @@ const NOT_FOUND = 'SERVICE_NOT_FOUND'
 // The service-state machine over a concrete implementation. Reads the service
 // before returning, so it never reports a state it has not verified — a
 // machine that assumed OFF at boot would have the dashboard show OFF for a
-// service that is up.
-export const createServiceStateMachine = async (name: string, service: ServiceImplementation, sink: EventSink): Promise<ServiceStateMachine> => {
+// service that is up. Nor does it later, when anybody needs to know: nothing
+// tells the machine when a service is changed by other hands — a unit stopped
+// from a shell, a schedule that turns it off overnight — so a reading
+// `maxAgeMs` old is taken again before a status is answered from it, and
+// while the sink has an audience — a socket open to be pushed to — the
+// machine re-reads on that cadence by itself and pushes what changed, so a
+// dashboard left open follows the service. With nobody asking and nobody
+// watching, the service is left alone. A transition that lands counts as a
+// reading. A transition under way and ERROR are the machine's own word rather
+// than a reading, and are not read past — ERROR until somebody asks for a
+// refresh, because reading past it would take the why away before anyone saw
+// it.
+export const createServiceStateMachine = async (name: string, service: ServiceImplementation, sink: EventSink & Audience, maxAgeMs: number): Promise<ServiceStateMachine> => {
     let current: ServiceStatus = {
         service: name,
         state: 'OFF',
         error: null,
         changed_at: Date.now()
     }
-    let inFlight = false
+    // When the service was last read, or a transition last landed — which is
+    // what `maxAgeMs` is counted from. The boot read below is the first.
+    let readAt = Date.now()
+    let transitioning = false
+    // Counts the transitions begun. A read that began before the latest one
+    // reports what the service was before it, so its answer is dropped rather
+    // than adopted over the transition's.
+    let transitions = 0
+    // The read under way and the transition count it began under. A refresh
+    // or re-read asked while it is out joins it rather than asking the service
+    // again — unless a transition has begun since, which makes it a reading of
+    // the past: then a fresh read goes out, and the old one is dropped when it
+    // settles.
+    let reading: { promise: Promise<ServiceStatus>; under: number } | undefined
+    let nextRead: ReturnType<typeof setTimeout> | undefined
+    let closed = false
 
     const set = (state: ServiceState, error: string | null): ServiceStatus => {
         current = {
@@ -66,44 +102,104 @@ export const createServiceStateMachine = async (name: string, service: ServiceIm
     // a poll against a steady service is silent.
     const adopt = (state: ServiceState, error: string | null): ServiceStatus => (state === current.state && error === current.error ? current : set(state, error))
 
+    // The next re-read, for when the latest reading turns `maxAgeMs` old —
+    // straight away if it already has. Only while somebody is watching; none
+    // while a transition is under way, whose landing is the next reading; and
+    // none in ERROR, which stands until somebody refreshes.
+    const schedule = (): void => {
+        clearTimeout(nextRead)
+        if (!closed && sink.size > 0 && !transitioning && current.state !== 'ERROR') {
+            nextRead = setTimeout(() => void read(), Math.max(0, readAt + maxAgeMs - Date.now()))
+        }
+    }
+
+    // The first watcher starts the re-reads and the last one to leave stops
+    // them.
+    const stopListening = sink.onPresence(present => {
+        if (present) {
+            schedule()
+        } else {
+            clearTimeout(nextRead)
+        }
+    })
+
+    // A transition landing or a read settling: the service as it now stands,
+    // which is a reading — and a push, wherever it moved. A transition lands
+    // away from the state it passed through, so it always pushes.
+    const reached = (state: ServiceState, error: string | null): ServiceStatus => {
+        readAt = Date.now()
+        const status = adopt(state, error)
+        schedule()
+        return status
+    }
+
     const run = (command: ServiceCommand): ServiceStatus => {
         const { from, during, to } = TRANSITIONS[command]
-        if (inFlight || current.state !== from) {
+        if (transitioning || current.state !== from) {
             throw new ApiError(BUSY)
         }
-        inFlight = true
+        transitioning = true
+        transitions += 1
+        clearTimeout(nextRead)
         const answer = set(during, null)
-        void service[command]()
+        // Over the moment it lands, because its landing is a reading and a
+        // reading schedules the next only with no transition under way.
+        const landed = (state: ServiceState, error: string | null): void => {
+            transitioning = false
+            reached(state, error)
+        }
+        void service[command]().then(
+            () => landed(to, null),
+            (e: Error) => landed('ERROR', e.message)
+        )
+        return answer
+    }
+
+    const read = (): Promise<ServiceStatus> => {
+        if (reading?.under === transitions) {
+            return reading.promise
+        }
+        clearTimeout(nextRead)
+        const under = transitions
+        // A transition begun since the read went out has the last word — its
+        // own landing schedules the next read — and the read's answer, from
+        // before it, is dropped.
+        const land = (state: ServiceState, error: string | null): ServiceStatus => (under === transitions ? reached(state, error) : current)
+        const promise = service
+            .check()
             .then(
-                () => set(to, null),
-                (e: Error) => set('ERROR', e.message)
+                state => land(state, null),
+                (e: Error) => land('ERROR', e.message)
             )
             .finally(() => {
-                inFlight = false
+                if (reading?.promise === promise) {
+                    reading = undefined
+                }
             })
-        return answer
+        reading = { promise, under }
+        return promise
     }
 
     const refresh = async (): Promise<ServiceStatus> => {
-        if (inFlight) {
+        if (transitioning) {
             throw new ApiError(BUSY)
         }
-        inFlight = true
-        try {
-            return adopt(await service.check(), null)
-        } catch (e) {
-            return adopt('ERROR', (e as Error).message)
-        } finally {
-            inFlight = false
-        }
+        return read()
     }
 
     const machine: ServiceStateMachine = {
         name,
-        status: () => current,
+        // Answered from the latest reading while it is fresh. A transition under
+        // way and ERROR are answered as they stand, being the machine's own word.
+        status: () => (transitioning || current.state === 'ERROR' || Date.now() - readAt < maxAgeMs ? Promise.resolve(current) : read()),
         start: () => run('start'),
         stop: () => run('stop'),
-        refresh
+        refresh,
+        close: () => {
+            closed = true
+            clearTimeout(nextRead)
+            stopListening()
+        }
     }
     await machine.refresh()
     return machine