[fd/fragment] Improve progress calculation (#8241)

This uses the download speed from all threads and also adds smoothing to speed and eta

Authored by: Grub4K

yt_dlp/downloader/fragment.py

@@ -14,6 +14,7 @@
 from ..networking.exceptions import HTTPError, IncompleteRead
 from ..utils import DownloadError, RetryManager, encodeFilename, traverse_obj
 from ..utils.networking import HTTPHeaderDict
+from ..utils.progress import ProgressCalculator
 
 
 class HttpQuietDownloader(HttpFD):
@@ -226,8 +227,7 @@ def _start_frag_download(self, ctx, info_dict):
         resume_len = ctx['complete_frags_downloaded_bytes']
         total_frags = ctx['total_frags']
         ctx_id = ctx.get('ctx_id')
-        # This dict stores the download progress, it's updated by the progress
+        # Stores the download progress, updated by the progress hook
-        # hook
         state = {
             'status': 'downloading',
             'downloaded_bytes': resume_len,
@@ -237,14 +237,8 @@ def _start_frag_download(self, ctx, info_dict):
             'tmpfilename': ctx['tmpfilename'],
         }
 
-        start = time.time()
+        ctx['started'] = time.time()
-        ctx.update({
+        progress = ProgressCalculator(resume_len)
-            'started': start,
-            'fragment_started': start,
-            # Amount of fragment's bytes downloaded by the time of the previous
-            # frag progress hook invocation
-            'prev_frag_downloaded_bytes': 0,
-        })
 
         def frag_progress_hook(s):
             if s['status'] not in ('downloading', 'finished'):
@@ -259,38 +253,35 @@ def frag_progress_hook(s):
             state['max_progress'] = ctx.get('max_progress')
             state['progress_idx'] = ctx.get('progress_idx')
 
-            time_now = time.time()
+            state['elapsed'] = progress.elapsed
-            state['elapsed'] = time_now - start
             frag_total_bytes = s.get('total_bytes') or 0
             s['fragment_info_dict'] = s.pop('info_dict', {})
+
+            # XXX: Fragment resume is not accounted for here
             if not ctx['live']:
                 estimated_size = (
                     (ctx['complete_frags_downloaded_bytes'] + frag_total_bytes)
                     / (state['fragment_index'] + 1) * total_frags)
-                state['total_bytes_estimate'] = estimated_size
+                progress.total = estimated_size
+                progress.update(s.get('downloaded_bytes'))
+                state['total_bytes_estimate'] = progress.total
+            else:
+                progress.update(s.get('downloaded_bytes'))
 
             if s['status'] == 'finished':
                 state['fragment_index'] += 1
                 ctx['fragment_index'] = state['fragment_index']
-                state['downloaded_bytes'] += frag_total_bytes - ctx['prev_frag_downloaded_bytes']
+                progress.thread_reset()
-                ctx['complete_frags_downloaded_bytes'] = state['downloaded_bytes']
+
-                ctx['speed'] = state['speed'] = self.calc_speed(
+            state['downloaded_bytes'] = ctx['complete_frags_downloaded_bytes'] = progress.downloaded
-                    ctx['fragment_started'], time_now, frag_total_bytes)
+            state['speed'] = ctx['speed'] = progress.speed.smooth
-                ctx['fragment_started'] = time.time()
+            state['eta'] = progress.eta.smooth
-                ctx['prev_frag_downloaded_bytes'] = 0
+
-            else:
-                frag_downloaded_bytes = s['downloaded_bytes']
-                state['downloaded_bytes'] += frag_downloaded_bytes - ctx['prev_frag_downloaded_bytes']
-                ctx['speed'] = state['speed'] = self.calc_speed(
-                    ctx['fragment_started'], time_now, frag_downloaded_bytes - ctx.get('frag_resume_len', 0))
-                if not ctx['live']:
-                    state['eta'] = self.calc_eta(state['speed'], estimated_size - state['downloaded_bytes'])
-                ctx['prev_frag_downloaded_bytes'] = frag_downloaded_bytes
             self._hook_progress(state, info_dict)
 
         ctx['dl'].add_progress_hook(frag_progress_hook)
 
-        return start
+        return ctx['started']
 
     def _finish_frag_download(self, ctx, info_dict):
         ctx['dest_stream'].close()
@@ -500,7 +491,6 @@ def _download_fragment(fragment):
                 download_fragment(fragment, ctx_copy)
                 return fragment, fragment['frag_index'], ctx_copy.get('fragment_filename_sanitized')
 
-            self.report_warning('The download speed shown is only of one thread. This is a known issue')
             with tpe or concurrent.futures.ThreadPoolExecutor(max_workers) as pool:
                 try:
                     for fragment, frag_index, frag_filename in pool.map(_download_fragment, fragments):

yt_dlp/utils/progress.py

@@ -0,0 +1,109 @@
+from __future__ import annotations
+
+import bisect
+import threading
+import time
+
+
+class ProgressCalculator:
+    # Time to calculate the speed over (seconds)
+    SAMPLING_WINDOW = 3
+    # Minimum timeframe before to sample next downloaded bytes (seconds)
+    SAMPLING_RATE = 0.05
+    # Time before showing eta (seconds)
+    GRACE_PERIOD = 1
+
+    def __init__(self, initial: int):
+        self._initial = initial or 0
+        self.downloaded = self._initial
+
+        self.elapsed: float = 0
+        self.speed = SmoothValue(0, smoothing=0.7)
+        self.eta = SmoothValue(None, smoothing=0.9)
+
+        self._total = 0
+        self._start_time = time.monotonic()
+        self._last_update = self._start_time
+
+        self._lock = threading.Lock()
+        self._thread_sizes: dict[int, int] = {}
+
+        self._times = [self._start_time]
+        self._downloaded = [self.downloaded]
+
+    @property
+    def total(self):
+        return self._total
+
+    @total.setter
+    def total(self, value: int | None):
+        with self._lock:
+            if value is not None and value < self.downloaded:
+                value = self.downloaded
+
+            self._total = value
+
+    def thread_reset(self):
+        current_thread = threading.get_ident()
+        with self._lock:
+            self._thread_sizes[current_thread] = 0
+
+    def update(self, size: int | None):
+        if not size:
+            return
+
+        current_thread = threading.get_ident()
+
+        with self._lock:
+            last_size = self._thread_sizes.get(current_thread, 0)
+            self._thread_sizes[current_thread] = size
+            self._update(size - last_size)
+
+    def _update(self, size: int):
+        current_time = time.monotonic()
+
+        self.downloaded += size
+        self.elapsed = current_time - self._start_time
+        if self.total is not None and self.downloaded > self.total:
+            self._total = self.downloaded
+
+        if self._last_update + self.SAMPLING_RATE > current_time:
+            return
+        self._last_update = current_time
+
+        self._times.append(current_time)
+        self._downloaded.append(self.downloaded)
+
+        offset = bisect.bisect_left(self._times, current_time - self.SAMPLING_WINDOW)
+        del self._times[:offset]
+        del self._downloaded[:offset]
+        if len(self._times) < 2:
+            self.speed.reset()
+            self.eta.reset()
+            return
+
+        download_time = current_time - self._times[0]
+        if not download_time:
+            return
+
+        self.speed.set((self.downloaded - self._downloaded[0]) / download_time)
+        if self.total and self.speed.value and self.elapsed > self.GRACE_PERIOD:
+            self.eta.set((self.total - self.downloaded) / self.speed.value)
+        else:
+            self.eta.reset()
+
+
+class SmoothValue:
+    def __init__(self, initial: float | None, smoothing: float):
+        self.value = self.smooth = self._initial = initial
+        self._smoothing = smoothing
+
+    def set(self, value: float):
+        self.value = value
+        if self.smooth is None:
+            self.smooth = self.value
+        else:
+            self.smooth = (1 - self._smoothing) * value + self._smoothing * self.smooth
+
+    def reset(self):
+        self.value = self.smooth = self._initial