Fixed instant loading bugs

material/overrides/main.html

@@ -23,5 +23,5 @@
 {% endblock %}
 {% block scripts %}
   {{ super() }}
-  <script src="{{ 'assets/javascripts/custom.054acff4.min.js' | url }}"></script>
+  <script src="{{ 'assets/javascripts/custom.30be3006.min.js' | url }}"></script>
 {% endblock %}

material/templates/base.html

@@ -249,7 +249,7 @@
       </script>
     {% endblock %}
     {% block scripts %}
-      <script src="{{ 'assets/javascripts/bundle.7389ff0e.min.js' | url }}"></script>
+      <script src="{{ 'assets/javascripts/bundle.8bb211bc.min.js' | url }}"></script>
       {% for script in config.extra_javascript %}
         {{ script | script_tag }}
       {% endfor %}

src/templates/assets/javascripts/browser/request/index.ts

@@ -46,20 +46,27 @@ interface Options {
 /**
  * Fetch the given URL
  *
- * If the request fails (e.g. when dispatched from `file://` locations), the
+ * This function returns an observable that emits the response as a blob and
- * observable will complete without emitting a value.
+ * completes, or emits an error if the request failed. The caller can cancel
+ * the request by unsubscribing at any time, which will automatically abort
+ * the inflight request and complete the observable.
+ *
+ * Note that we use `XMLHTTPRequest` not because we're nostalgic, but because
+ * it's the only way to get progress events for downloads and also allow for
+ * cancellation of requests, as the official Fetch API does not support this
+ * yet, even though we're already in 2024.
  *
  * @param url - Request URL
  * @param options - Options
  *
- * @returns Response observable
+ * @returns Data observable
  */
 export function request(
   url: URL | string, options?: Options
 ): Observable<Blob> {
   return new Observable<Blob>(observer => {
     const req = new XMLHttpRequest()
-    req.open("GET",  `${url}`)
+    req.open("GET", `${url}`)
     req.responseType = "blob"
 
     // Handle response
@@ -67,6 +74,8 @@ export function request(
       if (req.status >= 200 && req.status < 300) {
         observer.next(req.response)
         observer.complete()
+
+      // Every response that is not in the 2xx range is considered an error
       } else {
         observer.error(new Error(req.statusText))
       }
@@ -74,12 +83,12 @@ export function request(
 
     // Handle network errors
     req.addEventListener("error", () => {
-      observer.error(new Error("Network Error"))
+      observer.error(new Error("Network error"))
     })
 
     // Handle aborted requests
     req.addEventListener("abort", () => {
-      observer.error(new Error("Request aborted"))
+      observer.complete()
     })
 
     // Handle download progress
@@ -87,9 +96,12 @@ export function request(
       req.addEventListener("progress", event => {
         if (event.lengthComputable) {
           options.progress$!.next((event.loaded / event.total) * 100)
-        } else { // https://bugs.chromium.org/p/chromium/issues/detail?id=463622
+
-          const totalFromHeader = Number(req.getResponseHeader("Content-Length")) || 0
+        // Hack: Chromium doesn't report the total number of bytes if content
-          options.progress$!.next((event.loaded / totalFromHeader) * 100)
+        // is compressed, so we need this fallback - see https://t.ly/ZXofI
+        } else {
+          const length = req.getResponseHeader("Content-Length") ?? 0
+          options.progress$!.next((event.loaded / +length) * 100)
         }
       })
 
@@ -97,8 +109,9 @@ export function request(
       options.progress$.next(5)
     }
 
-    // Send request
+    // Send request and automatically abort request upon unsubscription
     req.send()
+    return () => req.abort()
   })
 }
 
@@ -125,6 +138,26 @@ export function requestJSON<T>(
     )
 }
 
+/**
+ * Fetch HTML from the given URL
+ *
+ * @param url - Request URL
+ * @param options - Options
+ *
+ * @returns Data observable
+ */
+export function requestHTML(
+  url: URL | string, options?: Options
+): Observable<Document> {
+  const dom = new DOMParser()
+  return request(url, options)
+    .pipe(
+      switchMap(res => res.text()),
+      map(res => dom.parseFromString(res, "text/html")),
+      shareReplay(1)
+    )
+}
+
 /**
  * Fetch XML from the given URL
  *

src/templates/assets/javascripts/integrations/instant/index.ts

@@ -24,23 +24,24 @@ import {
   EMPTY,
   Observable,
   Subject,
-  bufferCount,
   catchError,
+  combineLatestWith,
   concat,
   debounceTime,
+  distinct,
+  distinctUntilChanged,
   distinctUntilKeyChanged,
   endWith,
-  filter,
+  exhaustMap,
   fromEvent,
   ignoreElements,
   map,
+  merge,
   of,
-  sample,
   share,
-  skip,
-  startWith,
   switchMap,
   take,
+  tap,
   withLatestFrom
 } from "rxjs"
 
@@ -50,13 +51,13 @@ import {
   getElements,
   getLocation,
   getOptionalElement,
-  request,
+  requestHTML,
   setLocation,
   setLocationHash
 } from "~/browser"
 import { getComponentElement } from "~/components"
 
-import { fetchSitemap } from "../sitemap"
+import { Sitemap, fetchSitemap } from "../sitemap"
 
 /* ----------------------------------------------------------------------------
  * Helper types
@@ -75,55 +76,164 @@ interface SetupOptions {
  * Helper functions
  * ------------------------------------------------------------------------- */
 
+/**
+ * Handle clicks on internal URLs while skipping external URLs
+ *
+ * @param ev - Mouse event
+ * @param sitemap - Sitemap
+ *
+ * @returns URL observable
+ */
+function handle(
+  ev: MouseEvent, sitemap: Sitemap
+): Observable<URL> {
+  if (!(ev.target instanceof Element))
+    return EMPTY
+
+  // Skip, as target is not within a link - clicks on non-link elements are
+  // also captured, which we need to exclude from processing
+  const el = ev.target.closest("a")
+  if (el === null)
+    return EMPTY
+
+  // Skip, as link opens in new window - we now know we have captured a click
+  // on a link, but the link either has a `target` property defined, or the
+  // user pressed the `meta` or `ctrl` key to open it in a new window. Thus,
+  // we need to filter this event as well.
+  if (el.target || ev.metaKey || ev.ctrlKey)
+    return EMPTY
+
+  // Next, we must check if the URL is relevant for us, i.e., if it's an
+  // internal link to a page that is managed by MkDocs. Only then we can be
+  // sure that the structure of the page to be loaded adheres to the current
+  // document structure and can subsequently be injected into it without doing
+  // a full reload. For this reason, we must canonicalize the URL by removing
+  // all search parameters and hash fragments.
+  const url = new URL(el.href)
+  url.search = url.hash = ""
+
+  // Skip, if URL is not included in the sitemap - this could be the case when
+  // linking between versions or languages, or to another page that the author
+  // included as part of the build, but that is not managed by MkDocs. In that
+  // case we must not continue with instant navigation.
+  if (!sitemap.has(`${url}`))
+    return EMPTY
+
+  // We now know that we have a link to an internal page, so we prevent the
+  // browser from navigation and emit the URL for instant navigation. Note that
+  // this also includes anchor links, which means we need to implement anchor
+  // positioning ourselves. The reason for this is that if we wouldn't manage
+  // anchor links as well, scroll restoration will not work correctly (e.g.
+  // following an anchor link and scrolling).
+  ev.preventDefault()
+  return of(new URL(el.href))
+}
+
 /**
  * Create a map of head elements for lookup and replacement
  *
- * @param head - Document head
+ * @param document - Document
  *
- * @returns Element map
+ * @returns Tag map
  */
-function lookup(head: HTMLHeadElement): Map<string, HTMLElement> {
+function head(document: Document): Map<string, HTMLElement> {
-
-  // @todo When resolving URLs, we must make sure to use the correct base for
-  // resolution. The next time we refactor instant loading, we should use the
-  // location subject as a source, which is also used for anchor links tracking,
-  // but for now we just rely on canonical.
-  const canonical = getOptionalElement<HTMLLinkElement>("[rel=canonical]", head)
-  if (typeof canonical !== "undefined")
-    canonical.href = canonical.href.replace("//localhost:", "//127.0.0.1:")
-
-  // Create tag map and index elements in head
   const tags = new Map<string, HTMLElement>()
-  for (const el of getElements(":scope > *", head)) {
+  for (const el of getElements(":scope > *", document.head))
-    let html = el.outerHTML
+    tags.set(el.outerHTML, el)
-
-    // If the current element is a style sheet or script, we must resolve the
-    // URL relative to the current location and make it absolute, so it's easy
-    // to deduplicate it later on by comparing the outer HTML of tags. We must
-    // keep identical style sheets and scripts without replacing them.
-    for (const key of ["href", "src"]) {
-      const value = el.getAttribute(key)!
-      if (value === null)
-        continue
-
-      // Resolve URL relative to current location
-      const url = new URL(value, canonical?.href)
-      const ref = el.cloneNode() as HTMLElement
-
-      // Set resolved URL and retrieve HTML for deduplication
-      ref.setAttribute(key, `${url}`)
-      html = ref.outerHTML
-      break
-    }
-
-    // Index element in tag map
-    tags.set(html, el)
-  }
 
   // Return tag map
   return tags
 }
 
+/**
+ * Resolve relative URLs in the given document
+ *
+ * @param document - Document
+ *
+ * @returns Document observable
+ */
+function resolve(document: Document): Observable<Document> {
+  for (const el of getElements<HTMLLinkElement>("[href], [src]", document))
+    for (const key in ["href", "src"]) {
+      const value = el.getAttribute(key)
+      if (!/^(?:[a-z]+:)?\/\//i.test(value!))
+        el.href = el.href
+    }
+
+  // Return document observable
+  return of(document)
+}
+
+/**
+ * Create a map of head elements for lookup and replacement
+ *
+ * @param next - Next document
+ *
+ * @returns Document observable
+ */
+function inject(next: Document): Observable<Document> {
+  for (const selector of [
+    "[data-md-component=announce]",
+    "[data-md-component=container]",
+    "[data-md-component=header-topic]",
+    "[data-md-component=outdated]",
+    "[data-md-component=logo]",
+    "[data-md-component=skip]",
+    ...feature("navigation.tabs.sticky")
+      ? ["[data-md-component=tabs]"]
+      : []
+  ]) {
+    const source = getOptionalElement(selector)
+    const target = getOptionalElement(selector, next)
+    if (
+      typeof source !== "undefined" &&
+      typeof target !== "undefined"
+    ) {
+      source.replaceWith(target)
+    }
+  }
+
+  // Update meta tags
+  const tags = head(document)
+  for (const [html, el] of head(next))
+    if (tags.has(html))
+      tags.delete(html)
+    else
+      document.head.appendChild(el)
+
+  // Remove meta tags that are not present in the new document
+  for (const el of tags.values())
+    el.remove()
+
+  // After components and meta tags were replaced, re-evaluate scripts
+  // that were provided by the author as part of Markdown files
+  const container = getComponentElement("container")
+  return concat(getElements("script", container))
+    .pipe(
+      switchMap(el => {
+        const script = next.createElement("script")
+        if (el.src) {
+          for (const name of el.getAttributeNames())
+            script.setAttribute(name, el.getAttribute(name)!)
+          el.replaceWith(script)
+
+          // Complete when script is loaded
+          return new Observable(observer => {
+            script.onload = () => observer.complete()
+          })
+
+        // Complete immediately
+        } else {
+          script.textContent = el.textContent
+          el.replaceWith(script)
+          return EMPTY
+        }
+      }),
+      ignoreElements(),
+      endWith(next)
+    )
+}
+
 /* ----------------------------------------------------------------------------
  * Functions
  * ------------------------------------------------------------------------- */
@@ -146,292 +256,168 @@ export function setupInstantNavigation(
     return EMPTY
 
   // Load sitemap immediately, so we have it available when the user initiates
-  // the first instant navigation request, and canonicalize URLs to the current
+  // the first navigation request, so there's no perceived delay.
-  // base URL. The base URL will remain stable in between loads, as it's only
+  const sitemap$ = fetchSitemap(config.base)
-  // read at the first initialization of the application.
-  const sitemap$ = fetchSitemap()
-    .pipe(
-      map(paths => paths.map(path => `${new URL(path, config.base)}`))
-    )
 
-  // Intercept inter-site navigation - to keep the number of event listeners
+  // Since we might be on a slow connection, the user might trigger multiple
-  // low we use the fact that uncaptured events bubble up to the body. This also
+  // instant navigation events that overlap. MkDocs produces relative URLs for
-  // has the nice property that we don't need to detach and then again attach
+  // all internal links, which becomes a problem in this case, because we need
-  // event listeners when instant navigation occurs.
+  // to change the base URL the moment the user clicks a link that should be
-  const instant$ = fromEvent<MouseEvent>(document.body, "click")
+  // intercepted in order to be consistent with popstate, which means that the
-    .pipe(
+  // base URL would now be incorrect when resolving another relative link from
-      withLatestFrom(sitemap$),
+  // the same site. For this reason we always resolve all relative links to
-      switchMap(([ev, sitemap]) => {
+  // absolute links, so we can be sure this never happens.
-        if (!(ev.target instanceof Element))
+  of(document)
-          return EMPTY
+    .subscribe(resolve)
 
-        // Skip, as target is not within a link - clicks on non-link elements
+  // --------------------------------------------------------------------------
-        // are also captured, which we need to exclude from processing
+  // Navigation interception
-        const el = ev.target.closest("a")
+  // --------------------------------------------------------------------------
-        if (el === null)
-          return EMPTY
 
-        // Skip, as link opens in new window - we now know we have captured a
+  // Intercept navigation - to keep the number of event listeners down we use
-        // click on a link, but the link either has a `target` property defined,
+  // the fact that uncaptured events bubble up to the body. This has the nice
-        // or the user pressed the `meta` or `ctrl` key to open it in a new
+  // property that we don't need to detach and then re-attach event listeners
-        // window. Thus, we need to filter those events, too.
+  // when the document is replaced after a navigation event.
-        if (el.target || ev.metaKey || ev.ctrlKey)
+  const instant$ =
-          return EMPTY
+    fromEvent<MouseEvent>(document.body, "click")
+      .pipe(
+        combineLatestWith(sitemap$),
+        switchMap(([ev, sitemap]) => handle(ev, sitemap)),
+        share()
+      )
 
-        // Next, we must check if the URL is relevant for us, i.e., if it's an
+  // Intercept history change events, e.g. when the user uses the browser's
-        // internal link to a page that is managed by MkDocs. Only then we can
+  // back or forward buttons, and emit new location for fetching and parsing
-        // be sure that the structure of the page to be loaded adheres to the
+  const history$ =
-        // current document structure and can subsequently be injected into it
+    fromEvent<PopStateEvent>(window, "popstate")
-        // without doing a full reload. For this reason, we must canonicalize
+      .pipe(
-        // the URL by removing all search parameters and hash fragments.
+        map(getLocation),
-        const url = new URL(el.href)
+        share()
-        url.search = url.hash = ""
+      )
 
-        // Skip, if URL is not included in the sitemap - this could be the case
+  // While it would be better UX to defer navigation events until the document
-        // when linking between versions or languages, or to another page that
+  // is fully fetched and parsed, we must schedule it here to synchronize with
-        // the author included as part of the build, but that is not managed by
+  // popstate events, as they are emitted immediately. Moreover we need to
-        // MkDocs. In that case we must not continue with instant navigation.
+  // store the current viewport offset for scroll restoration later on.
-        if (!sitemap.includes(`${url}`))
-          return EMPTY
-
-        // We now know that we have a link to an internal page, so we prevent
-        // the browser from navigation and emit the URL for instant navigation.
-        // Note that this also includes anchor links, which means we need to
-        // implement anchor positioning ourselves. The reason for this is that
-        // if we wouldn't manage anchor links as well, scroll restoration will
-        // not work correctly (e.g. following an anchor link and scrolling).
-        ev.preventDefault()
-        return of(new URL(el.href))
-      }),
-      share()
-    )
-
-  // Before fetching for the first time, resolve the absolute favicon position,
-  // as the browser will try to fetch the icon immediately
-  instant$.pipe(take(1))
-    .subscribe(() => {
-      const favicon = getOptionalElement<HTMLLinkElement>("link[rel=icon]")
-      if (typeof favicon !== "undefined")
-        favicon.href = favicon.href
-    })
-
-  // Enable scroll restoration before window unloads - this is essential to
-  // ensure that full reloads (F5) restore the viewport offset correctly. If
-  // only popstate events wouldn't reset the scroll position prior to their
-  // emission, we could just reset this in popstate. Meh.
-  fromEvent(window, "beforeunload")
-    .subscribe(() => {
-      history.scrollRestoration = "auto"
-    })
-
-  // When an instant navigation event occurs, disable scroll restoration, since
-  // we must normalize and synchronize the behavior across all browsers. For
-  // instance, when the user clicks the back or forward button, the browser
-  // would immediately jump to the position of the previous document.
   instant$.pipe(withLatestFrom(viewport$))
     .subscribe(([url, { offset }]) => {
-      history.scrollRestoration = "manual"
-
-      // While it would be better UX to defer the history state change until the
-      // document was fully fetched and parsed, we must schedule it here, since
-      // popstate events are emitted when history state changes happen. Moreover
-      // we need to back up the current viewport offset, so we can restore it
-      // when popstate events occur, e.g., when the browser's back and forward
-      // buttons are used for navigation.
       history.replaceState(offset, "")
       history.pushState(null, "", url)
     })
 
-  // Emit URL that should be fetched via instant navigation on location subject,
+  // Emit URLs that should be fetched via instant navigation on location subject
-  // which was passed into this function. Instant navigation can be intercepted
+  // which was passed into this function. The state of instant navigation can be
-  // by other parts of the application, which can synchronously back up or
+  // intercepted by other parts of the application, which can synchronously back
-  // restore state before instant navigation happens.
+  // up or restore state before or after instant navigation happens.
-  instant$.subscribe(location$)
+  merge(instant$, history$)
+    .subscribe(location$)
 
-  // Fetch document - when fetching, we could use `responseType: document`, but
+  // --------------------------------------------------------------------------
-  // since all MkDocs links are relative, we need to make sure that the current
+  // Fetching and parsing
-  // location matches the document we just loaded. Otherwise any relative links
+  // --------------------------------------------------------------------------
-  // in the document might use the old location. If the request fails for some
+
-  // reason, we fall back to regular navigation and set the location explicitly,
+  // Fetch document - we deduplicate requests to the same location, so we don't
-  // which will force-load the page. Furthermore, we must pre-warm the buffer
+  // end up with multiple requests for the same page. We use `switchMap`, since
-  // for the duplicate check, or the first click on an anchor link will also
+  // we want to cancel the previous request when a new one is triggered, which
-  // trigger an instant navigation event, which doesn't make sense.
+  // is automatically handled by the observable returned by `request`. This is
-  const response$ = location$
+  // essential to ensure a good user experience, as we don't want to load pages
+  // that are not needed anymore, e.g., when the user clicks multiple links in
+  // quick succession or on slow connections. If the request fails for some
+  // reason, we fall back and use regular navigation, forcing a reload.
+  const document$ = location$
     .pipe(
-      startWith(getLocation()),
       distinctUntilKeyChanged("pathname"),
-      skip(1),
+      switchMap(url => requestHTML(url, { progress$ })
-      switchMap(url => request(url, { progress$ })
         .pipe(
           catchError(() => {
             setLocation(url, true)
             return EMPTY
           })
         )
-      )
+      ),
-    )
 
-  // Initialize the DOM parser, parse the returned HTML, and replace selected
+      // The document was successfully fetched and parsed, so we can inject its
-  // components before handing control down to the application
+      // contents into the currently active document
-  const dom = new DOMParser()
+      switchMap(resolve),
-  const document$ = response$
+      switchMap(inject),
-    .pipe(
-      switchMap(res => res.text()),
-      switchMap(res => {
-        const next = dom.parseFromString(res, "text/html")
-        for (const selector of [
-          "[data-md-component=announce]",
-          "[data-md-component=container]",
-          "[data-md-component=header-topic]",
-          "[data-md-component=outdated]",
-          "[data-md-component=logo]",
-          "[data-md-component=skip]",
-          ...feature("navigation.tabs.sticky")
-            ? ["[data-md-component=tabs]"]
-            : []
-        ]) {
-          const source = getOptionalElement(selector)
-          const target = getOptionalElement(selector, next)
-          if (
-            typeof source !== "undefined" &&
-            typeof target !== "undefined"
-          ) {
-            source.replaceWith(target)
-          }
-        }
-
-        // Update meta tags
-        const source = lookup(document.head)
-        const target = lookup(next.head)
-        for (const [html, el] of target) {
-
-          // Hack: skip stylesheets and scripts until we manage to replace them
-          // entirely in order to omit flashes of white content @todo refactor
-          if (
-            el.getAttribute("rel") === "stylesheet" ||
-            el.hasAttribute("src")
-          )
-            continue
-
-          if (source.has(html)) {
-            source.delete(html)
-          } else {
-            document.head.appendChild(el)
-          }
-        }
-
-        // Remove meta tags that are not present in the new document
-        for (const el of source.values())
-
-          // Hack: skip stylesheets and scripts until we manage to replace them
-          // entirely in order to omit flashes of white content @todo refactor
-          if (
-            el.getAttribute("rel") === "stylesheet" ||
-            el.hasAttribute("src")
-          )
-            continue
-          else
-            el.remove()
-
-        // After components and meta tags were replaced, re-evaluate scripts
-        // that were provided by the author as part of Markdown files
-        const container = getComponentElement("container")
-        return concat(getElements("script", container))
-          .pipe(
-            switchMap(el => {
-              const script = next.createElement("script")
-              if (el.src) {
-                for (const name of el.getAttributeNames())
-                  script.setAttribute(name, el.getAttribute(name)!)
-                el.replaceWith(script)
-
-                // Complete when script is loaded
-                return new Observable(observer => {
-                  script.onload = () => observer.complete()
-                })
-
-              // Complete immediately
-              } else {
-                script.textContent = el.textContent
-                el.replaceWith(script)
-                return EMPTY
-              }
-            }),
-            ignoreElements(),
-            endWith(next)
-          )
-      }),
       share()
     )
 
-  // Intercept popstate events, e.g. when using the browser's back and forward
+  // --------------------------------------------------------------------------
-  // buttons, and emit new location for fetching and parsing
+  // Scroll restoration
-  const popstate$ = fromEvent<PopStateEvent>(window, "popstate")
+  // --------------------------------------------------------------------------
-  popstate$.pipe(map(getLocation))
-    .subscribe(location$)
 
-  // Intercept clicks on anchor links, and scroll document into position - as
+  // Handle scroll restoration - we must restore the viewport offset after the
-  // we disabled scroll restoration, we need to do this manually here
+  // document has been fetched and injected, and every time the user clicks an
-  location$
+  // anchor that leads to an element on the same page, which might also happen
-    .pipe(
+  // when the user uses the back or forward button.
-      startWith(getLocation()),
+  merge(
-      bufferCount(2, 1),
+    document$.pipe(withLatestFrom(location$, (_, url) => url)),
-      filter(([prev, next]) => (
+
-        prev.pathname === next.pathname &&
+    // Handle instant navigation events that are triggered by the user clicking
-        prev.hash     !== next.hash
+    // on an anchor link with a hash fragment different from the current one, as
+    // well as from popstate events, which are emitted when the user navigates
+    // back and forth between pages. We use a two-layered subscription to scope
+    // the scroll restoration to the current page, as we don't need to restore
+    // the viewport offset when the user navigates to a different page, as this
+    // is already handled by the previous observable.
+    location$.pipe(
+      distinctUntilKeyChanged("pathname"),
+      switchMap(() => location$),
+      distinctUntilKeyChanged("hash"),
+    ),
+
+    // Handle instant navigation events that are triggered by the user clicking
+    // on an anchor link with the same hash fragment as the current one in the
+    // URL. Is is essential that we only intercept those from instant navigation
+    // events and not from history change events, or we'll end up in and endless
+    // loop. The top-level history entry must be removed, as it will be replaced
+    // with a new one, which would otherwise lead to a duplicate entry.
+    location$.pipe(
+      distinctUntilChanged((a, b) => (
+        a.pathname === b.pathname &&
+        a.hash     === b.hash
       )),
-      map(([, next]) => next)
+      switchMap(() => instant$),
+      tap(() => history.back())
     )
-      .subscribe(url => {
+  )
-        if (history.state !== null || !url.hash) {
+    .subscribe(url => {
-          window.scrollTo(0, history.state?.y ?? 0)
-        } else {
-          history.scrollRestoration = "auto"
-          setLocationHash(url.hash)
-          history.scrollRestoration = "manual"
-        }
-      })
 
-  // Intercept clicks on the same anchor link - we must use a distinct pipeline
+      // Check if the current history entry has a state, which happens when the
-  // for this, or we'd end up in a loop, setting the hash again and again
+      // user presses the back or forward button to visit a page we've already
-  location$
+      // seen. If there's no state, it means a new page was visited and we must
-    .pipe(
+      // scroll to the top, unless an anchor is given.
-      sample(instant$),
-      startWith(getLocation()),
-      bufferCount(2, 1),
-      filter(([prev, next]) => (
-        prev.pathname === next.pathname &&
-        prev.hash     === next.hash
-      )),
-      map(([, next]) => next)
-    )
-      .subscribe(url => {
-        history.scrollRestoration = "auto"
-        setLocationHash(url.hash)
-        history.scrollRestoration = "manual"
-
-        // Hack: we need to make sure that we don't end up with multiple history
-        // entries for the same anchor link, so we just remove the last entry
-        history.back()
-      })
-
-  // After parsing the document, check if the current history entry has a state.
-  // This may happen when users press the back or forward button to visit a page
-  // that was already seen. If there's no state, it means a new page was visited
-  // and we should scroll to the top, unless an anchor is given.
-  document$.pipe(withLatestFrom(location$))
-    .subscribe(([, url]) => {
       if (history.state !== null || !url.hash) {
         window.scrollTo(0, history.state?.y ?? 0)
       } else {
+        history.scrollRestoration = "auto"
         setLocationHash(url.hash)
+        history.scrollRestoration = "manual"
       }
     })
 
-  // If the current history is not empty, register an event listener updating
+  // Disable scroll restoration when an instant navigation event occurs, so the
-  // the current history state whenever the scroll position changes. This must
+  // browser does not immediately set the viewport offset to the prior history
-  // be debounced and cannot be done in popstate, as popstate has already
+  // entry, scrolling to the position on the same page, which would look odd.
-  // removed the entry from the history.
+  // Instead, we manually restore the position once the page has loaded.
+  location$.subscribe(() => {
+    history.scrollRestoration = "manual"
+  })
+
+  // Enable scroll restoration before window unloads - this is essential to
+  // ensure that full reloads (F5) restore the viewport offset correctly. If
+  // only popstate events wouldn't reset the viewport offset prior to their
+  // emission, we could just reset this in popstate. Meh.
+  fromEvent(window, "beforeunload")
+    .subscribe(() => {
+      history.scrollRestoration = "auto"
+    })
+
+  // Track viewport offset, so we can restore it when the user navigates back
+  // and forth between pages. Note that this must be debounced and cannot be
+  // done in popstate, as popstate has already removed the entry from the
+  // history, which means it is too late.
   viewport$
     .pipe(
       distinctUntilKeyChanged("offset"),

src/templates/assets/javascripts/integrations/sitemap/.eslintrc

@@ -0,0 +1,5 @@
+{
+  "rules": {
+    "no-null/no-null": "off"
+  }
+}

src/templates/assets/javascripts/integrations/sitemap/index.ts

@@ -21,17 +21,17 @@
  */
 
 import {
-  EMPTY,
   Observable,
   catchError,
-  defaultIfEmpty,
   map,
-  of,
+  of
-  tap
 } from "rxjs"
 
-import { configuration } from "~/_"
+import {
-import { getElements, requestXML } from "~/browser"
+  getElement,
+  getElements,
+  requestXML
+} from "~/browser"
 
 /* ----------------------------------------------------------------------------
  * Types
@@ -40,41 +40,75 @@ import { getElements, requestXML } from "~/browser"
 /**
  * Sitemap, i.e. a list of URLs
  */
-export type Sitemap = string[]
+export type Sitemap = Map<string, URL[]>
 
 /* ----------------------------------------------------------------------------
  * Helper functions
  * ------------------------------------------------------------------------- */
 
 /**
- * Preprocess a list of URLs
+ * Resolve URL to the given base URL
  *
- * This function replaces the `site_url` in the sitemap with the actual base
+ * When serving the site with instant navigation, MkDocs will set the hostname
- * URL, to allow instant navigation to work in occasions like Netlify previews.
+ * to the value as specified in `dev_addr`, but the browser allows for several
+ * hostnames to be used: `localhost`, `127.0.0.1` or even `0.0.0.0`, depending
+ * on configuration. This function resolves the URL to the given hostname.
  *
- * @param urls - URLs
+ * @param url - URL
+ * @param base - Base URL
  *
- * @returns URL path parts
+ * @returns Resolved URL
  */
-function preprocess(urls: Sitemap): Sitemap {
+function resolve(url: URL, base: URL) {
-  if (urls.length < 2)
+  url.protocol = base.protocol
-    return [""]
+  url.hostname = base.hostname
+  return url
+}
 
-  /* Take the first two URLs and remove everything after the last slash */
+/**
-  const [root, next] = [...urls]
+ * Extract sitemap from document
-    .sort((a, b) => a.length - b.length)
+ *
-    .map(url => url.replace(/[^/]+$/, ""))
+ * This function extracts the URLs and alternate links from the document, and
+ * associates alternate links to the original URL as found in `loc`, allowing
+ * the browser to navigate to the correct page when switching languages. The
+ * format of the sitemap is expected to adhere to:
+ *
+ * ``` xml
+ * <urlset>
+ *   <url>
+ *     <loc>...</loc>
+ *     <xhtml:link rel="alternate" hreflang="en" href="..."/>
+ *     <xhtml:link rel="alternate" hreflang="de" href="..."/>
+ *     ...
+ *   </url>
+ *   ...
+ * </urlset>
+ * ```
+ *
+ * @param document - Document
+ * @param base - Base URL
+ *
+ * @returns Sitemap
+ */
+function extract(document: Document, base: URL): Sitemap {
+  const sitemap: Sitemap = new Map()
+  for (const el of getElements("url", document)) {
+    const url = getElement("loc", el)
 
-  /* Compute common prefix */
+    // Create entry for location and add it to the list of links
-  let index = 0
+    const links = [resolve(new URL(url.textContent!), base)]
-  if (root === next)
+    sitemap.set(`${links[0]}`, links)
-    index = root.length
-  else
-    while (root.charCodeAt(index) === next.charCodeAt(index))
-      index++
 
-  /* Remove common prefix and return in original order */
+    // Attach alternate links to current entry
-  return urls.map(url => url.replace(root.slice(0, index), ""))
+    for (const link of getElements("[rel=alternate]", el)) {
+      const href = link.getAttribute("href")
+      if (href != null)
+        links.push(resolve(new URL(href), base))
+    }
+  }
+
+  // Return sitemap
+  return sitemap
 }
 
 /* ----------------------------------------------------------------------------
@@ -84,24 +118,17 @@ function preprocess(urls: Sitemap): Sitemap {
 /**
  * Fetch the sitemap for the given base URL
  *
+ * If a network or parsing error occurs, we just default to an empty sitemap,
+ * which means the caller should fall back to regular navigation.
+ *
  * @param base - Base URL
  *
  * @returns Sitemap observable
  */
-export function fetchSitemap(base?: URL): Observable<Sitemap> {
+export function fetchSitemap(base: URL | string): Observable<Sitemap> {
-  const cached = __md_get<Sitemap>("__sitemap", sessionStorage, base)
+  return requestXML(new URL("sitemap.xml", base))
-  if (cached) {
+    .pipe(
-    return of(cached)
+      map(document => extract(document, new URL(base))),
-  } else {
+      catchError(() => of(new Map())),
-    const config = configuration()
+    )
-    return requestXML(new URL("sitemap.xml", base || config.base))
-      .pipe(
-        map(sitemap => preprocess(getElements("loc", sitemap)
-          .map(node => node.textContent!)
-        )),
-        catchError(() => EMPTY), // @todo refactor instant loading
-        defaultIfEmpty([]),
-        tap(sitemap => __md_set("__sitemap", sitemap, sessionStorage, base))
-      )
-  }
 }

src/templates/assets/javascripts/integrations/version/index.ts

@@ -134,7 +134,7 @@ export function setupVersionSelector(
                 map(sitemap => {
                   const location = getLocation()
                   const path = location.href.replace(config.base, "")
-                  return sitemap.includes(path.split("#")[0])
+                  return sitemap.has(path.split("#")[0])
                     ? new URL(`../${version}/${path}`, config.base)
                     : new URL(url)
                 })