@eifi1/ui-kit 0.53.0 → 0.55.0

package/src/lib/server-wake.ts

@@ -1,315 +0,0 @@
-/**
- * "Is the server asleep, or is the app broken?" — the cold-start watchdog, lifted
- * from keksdose (feedback #199) so kastlan and Kurvenschmiede get it too.
- *
- * All three backends run on Cloud Run with `--min-instances=0`: after an idle
- * period the container is gone and the first request has to start it, which takes
- * several seconds and occasionally more. Nothing in the UI said so, and none of the
- * axios instances has a timeout, so the honest experience was a spinner that looked
- * stuck. This counts the in-flight requests a cold start would actually delay and
- * escalates in two steps, so a merely slow response says "still loading" and only a
- * genuinely long wait blames the sleeping server.
- *
- * Framework-free on purpose (no zustand, no axios, no React): a counter, two timers
- * and a listener set, read by `useServerWakeStage` through `useSyncExternalStore`.
- * The adapters below take the app's axios instance / fetch as an ARGUMENT, so the
- * kit never imports either.
- *
- * WHAT ARMS IT, and why the default is narrow:
- *   - reads (GET) — what every page does on open — always;
- *   - writes only when the app says so (`watchReadsAnd(/\/auth\/login\b/)`): the
- *     auth POSTs that gate the app are the only writes a user waits on before the
- *     app exists at all, so a cold start there looks exactly like a broken login.
- * An UPLOAD must never be watched — a statement PDF or a photo takes seconds on a
- * perfectly warm server, and "the server is waking up" during a 4 s upload would
- * simply be false. With the default that holds by itself (an upload is a POST);
- * an app widening `shouldWatch` keeps its upload routes out of it.
- *
- * Nor a DOWNLOAD, for the same reason (kastlan 0.18): axios answers only once the
- * whole body is in, so a generated PDF or a large photo — a GET — ran past both
- * thresholds on a warm server, and held the count for every read beside it. The
- * default filter skips a request whose `responseType` is `"blob"`, `"arraybuffer"` or
- * `"stream"`; a custom filter gets the request as its third argument to do the same.
- * (`wrapFetch` ends the wait at the headers, so a download through it is harmless.)
- *
- * It never arms while the browser is offline, whatever `shouldWatch` says: that state
- * has its own, truthful indicator, and a request that never left the device is not
- * evidence of anything about the server.
- */
-
-/** `slow` = longer than a warm server ever takes; `waking` = long enough that a cold
- *  start is the only sensible explanation. */
-export type ServerWakeStage = "idle" | "slow" | "waking";
-
-/** Decides which requests count. `method` arrives lower-cased (`"get"` when the
- *  request named none — axios's and fetch's default); `url` as the client had it;
- *  `request` is what the adapter saw — for axios the request config itself, so
- *  `responseType` or the app's own fields can decide. */
-export type ServerWakeFilter = (method: string, url: string, request: ServerWakeRequest) => boolean;
-
-export interface ServerWakeOptions {
-  /** Past this, a warm round trip is out of the question (a cold database query on a
-   *  free tier can still take a second, so it is not tighter). Default 2000 ms. */
-  slowMs?: number;
-  /** Past this, the container was almost certainly scaled to zero. Default 7000 ms. */
-  wakingMs?: number;
-  /** Which requests count. Default {@link watchReadsAnd}`()` — GETs only, downloads
-   *  excluded. The offline guard runs before it and cannot be overridden. */
-  shouldWatch?: ServerWakeFilter;
-}
-
-/** What a request is, for {@link ServerWakeWatcher.track} and a {@link ServerWakeFilter}. */
-export interface ServerWakeRequest {
-  method?: string;
-  url?: string;
-  /** axios's `responseType`. `"blob"`, `"arraybuffer"` and `"stream"` mark a download,
-   *  which the default filter does not count. */
-  responseType?: string;
-}
-
-export interface ServerWakeWatcher {
-  /**
-   * Count a request that is going out, and arm the escalation if it is the first.
-   * Returns whether it was counted — the caller MUST call {@link end} exactly once
-   * for every `true`, or the counter never returns to zero and the notice sticks.
-   * (The axios adapter stashes it on the request config for that reason.)
-   * `request` reaches the filter as its third argument; default `{ method, url }`.
-   */
-  start(method?: string, url?: string, request?: ServerWakeRequest): boolean;
-  /** A counted request settled — either way: an error still proves the server
-   *  answered or gave up. The stage then follows the oldest request still out (0.51), so
-   *  the notice goes away when the last one does, or sooner when the rest are recent.
-   *  Never takes the counter below zero. */
-  end(): void;
-  /** `start` + `end` around a promise (or a function returning one), for a request
-   *  that does not go through an adapter. Resolves and rejects as the work does. */
-  track<T>(work: PromiseLike<T> | (() => PromiseLike<T>), request?: ServerWakeRequest): Promise<T>;
-  /** Back to a pristine watchdog — for tests, which share an instance across cases. */
-  reset(): void;
-  /** `useSyncExternalStore`'s pair. Listeners hear stage changes only. */
-  subscribe(listener: () => void): () => void;
-  getStage(): ServerWakeStage;
-  /** Counted requests currently outstanding. */
-  getInFlight(): number;
-}
-
-/** Response types that make a request a download — see the module comment. */
-const DOWNLOAD_RESPONSE_TYPES = new Set(["blob", "arraybuffer", "stream"]);
-
-/**
- * The usual filter: every GET, plus any other request whose URL matches `writes` —
- * keksdose's `/\/auth\/(login|register|demo-session)\b/` — but never a download
- * (`responseType` `"blob"`, `"arraybuffer"` or `"stream"`). Keep upload routes out of
- * the pattern (see the module comment).
- */
-export function watchReadsAnd(writes?: RegExp): ServerWakeFilter {
-  return (method, url, request) =>
-    !DOWNLOAD_RESPONSE_TYPES.has(request?.responseType ?? "") &&
-    (method === "get" || (writes !== undefined && writes.test(url)));
-}
-
-function isOffline(): boolean {
-  return typeof navigator !== "undefined" && navigator.onLine === false;
-}
-
-export function createServerWake({
-  slowMs = 2000,
-  wakingMs = 7000,
-  shouldWatch = watchReadsAnd(),
-}: ServerWakeOptions = {}): ServerWakeWatcher {
-  // When each counted request went out, oldest first; its length is the count.
-  //
-  // The stage is the age of the OLDEST outstanding request (0.51). It used to be the age of
-  // the first request since the count was last zero, so overlapping reads that were each
-  // fast — one going out before the last came back, on a busy page — chained past 2 s and
-  // showed "taking longer than usual" with nothing slow at all (keksdose live #454, found
-  // there while chasing a read that really was slow).
-  //
-  // `end()` names no request, so it retires the OLDEST start. When a newer request comes
-  // back first, the start left standing for the older one is later than its own, so the
-  // age read is never MORE than the oldest request's real age: a notice can come late,
-  // never falsely.
-  const starts: number[] = [];
-  let stage: ServerWakeStage = "idle";
-  let slowTimer: ReturnType<typeof setTimeout> | null = null;
-  let wakingTimer: ReturnType<typeof setTimeout> | null = null;
-  const listeners = new Set<() => void>();
-
-  const setStage = (next: ServerWakeStage) => {
-    if (next === stage) return;
-    stage = next;
-    for (const listener of listeners) listener();
-  };
-
-  const clearTimers = () => {
-    if (slowTimer !== null) clearTimeout(slowTimer);
-    if (wakingTimer !== null) clearTimeout(wakingTimer);
-    slowTimer = null;
-    wakingTimer = null;
-  };
-
-  // The stage for the oldest outstanding request's age, and timers for what is still to
-  // come. With none outstanding: idle.
-  const arm = () => {
-    clearTimers();
-    if (starts.length === 0) {
-      setStage("idle");
-      return;
-    }
-    const age = Date.now() - starts[0];
-    setStage(age >= wakingMs ? "waking" : age >= slowMs ? "slow" : "idle");
-    if (age < slowMs) slowTimer = setTimeout(() => setStage("slow"), slowMs - age);
-    if (age < wakingMs) wakingTimer = setTimeout(() => setStage("waking"), wakingMs - age);
-  };
-
-  const start = (method?: string, url?: string, request: ServerWakeRequest = { method, url }): boolean => {
-    if (isOffline()) return false;
-    if (!shouldWatch((method ?? "get").toLowerCase(), url ?? "", request)) return false;
-    starts.push(Date.now());
-    // A request joining others does not move the clock: a page firing a second read at
-    // 1.9 s is still waiting on the same sleeping server, not starting afresh.
-    if (starts.length === 1) arm();
-    return true;
-  };
-
-  const end = () => {
-    if (starts.length === 0) return;
-    // The server answered (or gave up on) one of them, so the oldest's clock is over; the
-    // next oldest's age decides, which may well put the notice away again.
-    starts.shift();
-    arm();
-  };
-
-  return {
-    start,
-    end,
-    track<T>(work: PromiseLike<T> | (() => PromiseLike<T>), request: ServerWakeRequest = {}) {
-      const counted = start(request.method, request.url, request);
-      let promise: Promise<T>;
-      try {
-        promise = Promise.resolve(typeof work === "function" ? work() : work);
-      } catch (error) {
-        // A thunk that throws before it returns a promise settled too.
-        promise = Promise.reject(error);
-      }
-      return counted ? promise.finally(end) : promise;
-    },
-    reset() {
-      starts.length = 0;
-      clearTimers();
-      setStage("idle");
-    },
-    subscribe(listener) {
-      listeners.add(listener);
-      return () => {
-        listeners.delete(listener);
-      };
-    },
-    getStage: () => stage,
-    getInFlight: () => starts.length,
-  };
-}
-
-/**
- * The shared instance: what `<ServerWakeNotice />`, `useServerWakeStage()`,
- * {@link attachServerWake} and {@link wrapFetch} use when no watcher is passed. GETs
- * only, default timings — an app that also watches its login POST creates its own
- * with {@link createServerWake} and passes it to both ends.
- */
-export const serverWake: ServerWakeWatcher = createServerWake();
-
-// ── axios ─────────────────────────────────────────────────────────────────────
-
-/** The config fields the watchdog reads — what axios's request config has. The whole
- *  config reaches the filter, so an app's filter can read the rest of it. */
-export interface ServerWakeAxiosConfig {
-  method?: string;
-  url?: string;
-  responseType?: string;
-}
-
-/** The slice of an interceptor manager the adapter uses. A method (not a function
- *  property) so axios's wider `use(onFulfilled?, onRejected?, options?)` fits it. */
-interface InterceptorManagerLike<V> {
-  use(onFulfilled: (value: V) => V | Promise<V>, onRejected: (error: unknown) => unknown): number;
-  eject(id: number): void;
-}
-
-/** An axios instance, structurally — so the kit types the adapter without importing
- *  axios. `axios.create()`'s result fits it as is. */
-export interface AxiosLikeInstance<C extends ServerWakeAxiosConfig, R extends { config: C }> {
-  interceptors: {
-    request: InterceptorManagerLike<C>;
-    response: InterceptorManagerLike<R>;
-  };
-}
-
-/** Whether the watchdog counted this request — the response side owes a matching
- *  `end()` only then. On the config, as keksdose's `_wake` was: it travels with the
- *  request onto the response and onto an error. Cleared once settled, so a 401 replay
- *  of the same config re-enters the request interceptor and is counted afresh — the
- *  pairing stays balanced. */
-const COUNTED = "_serverWake";
-type Marked = { [COUNTED]?: boolean };
-
-/**
- * Drive a watcher from an axios instance's interceptors: `attachServerWake(api)`.
- * Returns a function that ejects both interceptors again.
- *
- * Attach it BEFORE the app's own response interceptors: axios runs those in the order
- * they were added, so the wake one ends the wait before a 401 handler awaits a token
- * refresh — otherwise the notice would keep counting a request that has already been
- * answered. (Request-side order does not matter.)
- */
-export function attachServerWake<C extends ServerWakeAxiosConfig, R extends { config: C }>(
-  instance: AxiosLikeInstance<C, R>,
-  watcher: ServerWakeWatcher = serverWake,
-): () => void {
-  const settle = (config: unknown) => {
-    const marked = config as Marked | undefined;
-    if (marked?.[COUNTED]) {
-      marked[COUNTED] = false;
-      watcher.end();
-    }
-  };
-  const request = instance.interceptors.request.use(
-    (config) => {
-      (config as C & Marked)[COUNTED] = watcher.start(config.method, config.url, config);
-      return config;
-    },
-    (error) => Promise.reject(error),
-  );
-  const response = instance.interceptors.response.use(
-    (res) => {
-      settle(res.config);
-      return res;
-    },
-    (error) => {
-      settle((error as { config?: unknown } | null)?.config);
-      return Promise.reject(error);
-    },
-  );
-  return () => {
-    instance.interceptors.request.eject(request);
-    instance.interceptors.response.eject(response);
-  };
-}
-
-// ── fetch ─────────────────────────────────────────────────────────────────────
-
-/**
- * A `fetch` that drives a watcher: `const apiFetch = wrapFetch(fetch)`. The wait ends
- * when the response's HEADERS arrive (or the call fails) — the server is awake by then,
- * however long the body takes.
- */
-export function wrapFetch(
-  fetchImpl: typeof fetch = globalThis.fetch,
-  watcher: ServerWakeWatcher = serverWake,
-): typeof fetch {
-  return (input, init) => {
-    const request = typeof Request !== "undefined" && input instanceof Request ? input : undefined;
-    const method = init?.method ?? request?.method;
-    const url = typeof input === "string" ? input : input instanceof URL ? input.href : (request?.url ?? "");
-    return watcher.track(() => fetchImpl(input, init), { method, url });
-  };
-}