Pick image_proxy changes from searxng (#2965)

* [mod] /image_proxy: don't decompress images

* [fix] image_proxy: always close the httpx respone

previously, when the content type was not an image and some other error,
the httpx response was not closed

* [mod] /image_proxy: use HTTP/1 instead of HTTP/2

httpx: HTTP/2 is slow when a lot data is downloaded.
https://github.com/dalf/pyhttp-benchmark

also, the usage of HTTP/1 decreases the load average

* [mod] searx.utils.dict_subset: rewrite with comprehension

Co-authored-by: Alexandre Flament <alex@al-f.net>

searx/network/__init__.py

@@ -5,6 +5,7 @@ import threading
 import concurrent.futures
 from time import time
 from queue import SimpleQueue
+from types import MethodType
 
 import httpx
 import h2.exceptions
@@ -134,15 +135,27 @@ async def stream_chunk_to_queue(network, q, method, url, **kwargs):
     try:
         async with await network.stream(method, url, **kwargs) as response:
             q.put(response)
+            # aiter_raw: access the raw bytes on the response without applying any HTTP content decoding
+            # https://www.python-httpx.org/quickstart/#streaming-responses
             async for chunk in response.aiter_bytes(65536):
                 if len(chunk) > 0:
                     q.put(chunk)
+    except httpx.ResponseClosed as e:
+        # the response was closed
+        pass
     except (httpx.HTTPError, OSError, h2.exceptions.ProtocolError) as e:
         q.put(e)
     finally:
         q.put(None)
 
 
+def _close_response_method(self):
+    asyncio.run_coroutine_threadsafe(
+        self.aclose(),
+        get_loop()
+    )
+
+
 def stream(method, url, **kwargs):
     """Replace httpx.stream.
 
@@ -158,10 +171,18 @@ def stream(method, url, **kwargs):
     q = SimpleQueue()
     future = asyncio.run_coroutine_threadsafe(stream_chunk_to_queue(get_network(), q, method, url, **kwargs),
                                               get_loop())
+    # yield response
+    response = q.get()
+    if isinstance(response, Exception):
+        raise response
+    response.close = MethodType(_close_response_method, response)
+    yield response
+
+    # yield chunks
     chunk_or_exception = q.get()
     while chunk_or_exception is not None:
         if isinstance(chunk_or_exception, Exception):
             raise chunk_or_exception
         yield chunk_or_exception
         chunk_or_exception = q.get()
-    return future.result()
+    future.result()

searx/network/network.py

@@ -326,6 +326,14 @@ def initialize(settings_engines=None, settings_outgoing=None):
         if isinstance(network, str):
             NETWORKS[engine_name] = NETWORKS[network]
 
+    # the /image_proxy endpoint has a dedicated network.
+    # same parameters than the default network, but HTTP/2 is disabled.
+    # It decreases the CPU load average, and the total time is more or less the same
+    if 'image_proxy' not in NETWORKS:
+        image_proxy_params = default_params.copy()
+        image_proxy_params['enable_http2'] = False
+        NETWORKS['image_proxy'] = new_network(image_proxy_params)
+
 
 @atexit.register
 def done():

searx/utils.py

@@ -272,11 +272,7 @@ def dict_subset(d, properties):
         >>> >> dict_subset({'A': 'a', 'B': 'b', 'C': 'c'}, ['A', 'D'])
         {'A': 'a'}
     """
-    result = {}
-    for k in properties:
-        if k in d:
-            result[k] = d[k]
-    return result
+    return {k: d[k] for k in properties if k in d}
 
 
 def get_torrent_size(filesize, filesize_multiplier):

searx/webapp.py

@@ -94,7 +94,7 @@ from searx.plugins import plugins
 from searx.plugins.oa_doi_rewrite import get_doi_resolver
 from searx.preferences import Preferences, ValidationException, LANGUAGE_CODES
 from searx.answerers import answerers
-from searx.network import stream as http_stream
+from searx.network import stream as http_stream, set_context_network_name
 from searx.answerers import ask
 from searx.metrology.error_recorder import errors_per_engines
 from searx.settings_loader import get_default_settings_path
@@ -921,6 +921,8 @@ def _is_selected_language_supported(engine, preferences):
 
 @app.route('/image_proxy', methods=['GET'])
 def image_proxy():
+    # pylint: disable=too-many-return-statements, too-many-branches
+
     url = request.args.get('url')
 
     if not url:
@@ -932,14 +934,21 @@ def image_proxy():
         return '', 400
 
     maximum_size = 5 * 1024 * 1024
-
+    forward_resp = False
+    resp = None
     try:
-        headers = dict_subset(request.headers, {'If-Modified-Since', 'If-None-Match'})
-        headers['User-Agent'] = gen_useragent()
+        request_headers = {
+            'User-Agent': gen_useragent(),
+            'Accept': 'image/webp,*/*',
+            'Accept-Encoding': 'gzip, deflate',
+            'Sec-GPC': '1',
+            'DNT': '1',
+        }
+        set_context_network_name('image_proxy')
         stream = http_stream(
             method='GET',
             url=url,
-            headers=headers,
+            headers=request_headers,
             timeout=settings['outgoing']['request_timeout'],
             follow_redirects=True,
             max_redirects=20)
@@ -949,25 +958,37 @@ def image_proxy():
         if content_length and content_length.isdigit() and int(content_length) > maximum_size:
             return 'Max size', 400
 
-        if resp.status_code == 304:
-            return '', resp.status_code
-
         if resp.status_code != 200:
             logger.debug('image-proxy: wrong response code: {0}'.format(resp.status_code))
             if resp.status_code >= 400:
                 return '', resp.status_code
             return '', 400
 
-        if not resp.headers.get('content-type', '').startswith('image/'):
-            logger.debug('image-proxy: wrong content-type: {0}'.format(resp.headers.get('content-type')))
+        if not resp.headers.get('Content-Type', '').startswith('image/'):
+            logger.debug('image-proxy: wrong content-type: %s', resp.headers.get('Content-Type', ''))
             return '', 400
 
-        headers = dict_subset(resp.headers, {'Content-Length', 'Length', 'Date', 'Last-Modified', 'Expires', 'Etag'})
+        forward_resp = True
+    except httpx.HTTPError:
+        logger.exception('HTTP error')
+        return '', 400
+    finally:
+        if resp and not forward_resp:
+            # the code is about to return an HTTP 400 error to the browser
+            # we make sure to close the response between searxng and the HTTP server
+            try:
+                resp.close()
+            except httpx.HTTPError:
+                logger.exception('HTTP error on closing')
 
-        total_length = 0
+    try:
+        headers = dict_subset(
+            resp.headers,
+            {'Content-Type', 'Content-Encoding', 'Content-Length', 'Length'}
+        )
 
         def forward_chunk():
-            nonlocal total_length
+            total_length = 0
             for chunk in stream:
                 total_length += len(chunk)
                 if total_length > maximum_size:
@@ -1148,6 +1169,13 @@ def run():
     )
 
 
+def patch_application(app):
+    # serve pages with HTTP/1.1
+    WSGIRequestHandler.protocol_version = "HTTP/{}".format(settings['server']['http_protocol_version'])
+    # patch app to handle non root url-s behind proxy & wsgi
+    app.wsgi_app = ReverseProxyPathFix(ProxyFix(app.wsgi_app))
+
+
 class ReverseProxyPathFix:
     '''Wrap the application in this middleware and configure the
     front-end server to add these headers, to let you quietly bind