From 4bc01aca8ddb2ca6aa1725aa075e028234ebe0f3 Mon Sep 17 00:00:00 2001
From: dgtlmoon <dgtlmoon@gmail.com>
Date: Wed, 11 Feb 2026 17:21:08 +0100
Subject: [PATCH] Price tracker - Use a more memory efficient price scraper,
 use subprocess on linux for cleaner memory management. (#3864)

---
 .../processors/restock_diff/processor.py      | 256 +++++++++++++++-
 .../restock_diff/pure_python_extractor.py     | 286 ++++++++++++++++++
 2 files changed, 541 insertions(+), 1 deletion(-)
 create mode 100644 changedetectionio/processors/restock_diff/pure_python_extractor.py

diff --git a/changedetectionio/processors/restock_diff/processor.py b/changedetectionio/processors/restock_diff/processor.py
index e63f97db..ba89b4e3 100644
--- a/changedetectionio/processors/restock_diff/processor.py
+++ b/changedetectionio/processors/restock_diff/processor.py
@@ -56,6 +56,259 @@ def _deduplicate_prices(data):
     return list(unique_data)
 
 
+# =============================================================================
+# MEMORY MANAGEMENT: Why We Use Multiprocessing (Linux Only)
+# =============================================================================
+#
+# The get_itemprop_availability() function uses 'extruct' to parse HTML metadata
+# (JSON-LD, microdata, OpenGraph, etc). Extruct internally uses lxml, which wraps
+# libxml2 - a C library that allocates memory at the C level.
+#
+# Memory Leak Problem:
+# --------------------
+# 1. lxml's document_fromstring() creates thousands of Python objects backed by
+#    C-level allocations (nodes, attributes, text content)
+# 2. Python's garbage collector can mark these objects as collectible, but
+#    cannot force the OS to reclaim the actual C-level memory
+# 3. malloc/free typically doesn't return memory to OS - it just marks it as
+#    "free in the process address space"
+# 4. With repeated parsing of large HTML (5MB+ pages), memory accumulates even
+#    after Python GC runs
+#
+# Why Multiprocessing Fixes This:
+# --------------------------------
+# When a subprocess exits, the OS forcibly reclaims ALL memory including C-level
+# allocations that Python GC couldn't release. This ensures clean memory state
+# after each extraction.
+#
+# Performance Impact:
+# -------------------
+# - Memray analysis showed 1.2M document_fromstring allocations per page
+# - Without subprocess: memory grows by ~50-500MB per parse and lingers
+# - With subprocess: ~35MB overhead but forces full cleanup after each run
+# - Trade-off: 35MB resource_tracker vs 500MB+ accumulated leak = much better at scale
+#
+# References:
+# -----------
+# - lxml memory issues: https://medium.com/devopss-hole/python-lxml-memory-leak-b8d0b1000dc7
+# - libxml2 caching behavior: https://www.mail-archive.com/lxml@python.org/msg00026.html
+# - GC limitations with C extensions: https://benbernardblog.com/tracking-down-a-freaky-python-memory-leak-part-2/
+#
+# Additional Context:
+# -------------------
+# - jsonpath_ng (used to query the parsed data) is pure Python and doesn't leak
+# - The leak is specifically from lxml's document parsing, not the JSONPath queries
+# - Linux-only because multiprocessing spawn is well-tested there; other platforms
+#   use direct call as fallback
+#
+# Alternative Solution (Future Optimization):
+# -------------------------------------------
+# This entire problem could be avoided by using regex to extract just the machine
+# data blocks (JSON-LD, microdata, OpenGraph tags) BEFORE parsing with lxml:
+#
+#   1. Use regex to extract <script type="application/ld+json">...</script> blocks
+#   2. Use regex to extract <meta property="og:*"> tags
+#   3. Use regex to find itemprop/itemtype attributes and their containing elements
+#   4. Parse ONLY those extracted snippets instead of the entire HTML document
+#
+# Benefits:
+#   - Avoids parsing 5MB of HTML when we only need a few KB of metadata
+#   - Eliminates the lxml memory leak entirely
+#   - Faster extraction (regex is much faster than DOM parsing)
+#   - No subprocess overhead needed
+#
+# Trade-offs:
+#   - Regex for HTML is brittle (comments, CDATA, edge cases)
+#   - Microdata extraction would be complex (need to track element boundaries)
+#   - Would need extensive testing to ensure we don't miss valid data
+#   - extruct is battle-tested; regex solution would need similar maturity
+#
+# For now, the subprocess approach is safer and leverages existing extruct code.
+# =============================================================================
+
+
+def _extract_itemprop_availability_worker(pipe_conn):
+    """
+    Subprocess worker for itemprop extraction (Linux memory management).
+
+    Uses spawn multiprocessing to isolate extruct/lxml memory allocations.
+    When the subprocess exits, the OS reclaims ALL memory including lxml's
+    C-level allocations that Python's GC cannot release.
+
+    Args:
+        pipe_conn: Pipe connection to receive HTML and send result
+    """
+    import json
+    import gc
+
+    html_content = None
+    result_data = None
+
+    try:
+        # Receive HTML as raw bytes (no pickle)
+        html_bytes = pipe_conn.recv_bytes()
+        html_content = html_bytes.decode('utf-8')
+
+        # Explicitly delete html_bytes to free memory
+        del html_bytes
+        gc.collect()
+
+        # Perform extraction in subprocess (uses extruct/lxml)
+        result_data = get_itemprop_availability(html_content)
+
+        # Convert Restock object to dict for JSON serialization
+        result = {
+            'success': True,
+            'data': dict(result_data) if result_data else {}
+        }
+        pipe_conn.send_bytes(json.dumps(result).encode('utf-8'))
+
+        # Clean up before exit
+        del result_data, html_content, result
+        gc.collect()
+
+    except MoreThanOnePriceFound:
+        # Serialize the specific exception type
+        result = {
+            'success': False,
+            'exception_type': 'MoreThanOnePriceFound'
+        }
+        pipe_conn.send_bytes(json.dumps(result).encode('utf-8'))
+
+    except Exception as e:
+        # Serialize other exceptions
+        result = {
+            'success': False,
+            'exception_type': type(e).__name__,
+            'exception_message': str(e)
+        }
+        pipe_conn.send_bytes(json.dumps(result).encode('utf-8'))
+
+    finally:
+        # Final cleanup before subprocess exits
+        # Variables may already be deleted in try block, so use try/except
+        try:
+            del html_content
+        except (NameError, UnboundLocalError):
+            pass
+        try:
+            del result_data
+        except (NameError, UnboundLocalError):
+            pass
+        gc.collect()
+        pipe_conn.close()
+
+
+def extract_itemprop_availability_safe(html_content) -> Restock:
+    """
+    Extract itemprop availability with hybrid approach for memory efficiency.
+
+    Strategy (fastest to slowest, least to most memory):
+    1. Try pure Python extraction (JSON-LD, OpenGraph, microdata) - covers 80%+ of cases
+    2. Fall back to extruct with subprocess isolation on Linux for complex cases
+
+    Args:
+        html_content: HTML string to parse
+
+    Returns:
+        Restock: Extracted availability data
+
+    Raises:
+        MoreThanOnePriceFound: When multiple prices detected
+        Other exceptions: From extruct/parsing
+    """
+    import platform
+
+    # Step 1: Try pure Python extraction first (fast, no lxml, no memory leak)
+    try:
+        from .pure_python_extractor import extract_metadata_pure_python, query_price_availability
+
+        logger.trace("Attempting pure Python metadata extraction (no lxml)")
+        extracted_data = extract_metadata_pure_python(html_content)
+        price_data = query_price_availability(extracted_data)
+
+        # If we got price AND availability, we're done!
+        if price_data.get('price') and price_data.get('availability'):
+            result = Restock(price_data)
+            logger.debug(f"Pure Python extraction successful: {dict(result)}")
+            return result
+
+        # If we got some data but not everything, still try extruct for completeness
+        if price_data.get('price') or price_data.get('availability'):
+            logger.debug(f"Pure Python extraction partial: {price_data}, will try extruct for completeness")
+
+    except Exception as e:
+        logger.debug(f"Pure Python extraction failed: {e}, falling back to extruct")
+
+    # Step 2: Fall back to extruct (uses lxml, needs subprocess on Linux)
+    logger.trace("Falling back to extruct (lxml-based) with subprocess isolation")
+
+    # Only use subprocess isolation on Linux
+    # Other platforms may have issues with spawn or don't need the aggressive memory management
+    if platform.system() == 'Linux':
+        import multiprocessing
+        import json
+        import gc
+
+        try:
+            ctx = multiprocessing.get_context('spawn')
+            parent_conn, child_conn = ctx.Pipe()
+            p = ctx.Process(target=_extract_itemprop_availability_worker, args=(child_conn,))
+            p.start()
+
+            # Send HTML as raw bytes (no pickle)
+            html_bytes = html_content.encode('utf-8')
+            parent_conn.send_bytes(html_bytes)
+
+            # Explicitly delete html_bytes copy immediately after sending
+            del html_bytes
+            gc.collect()
+
+            # Receive result as JSON
+            result_bytes = parent_conn.recv_bytes()
+            result = json.loads(result_bytes.decode('utf-8'))
+
+            # Wait for subprocess to complete
+            p.join()
+
+            # Close pipes
+            parent_conn.close()
+            child_conn.close()
+
+            # Clean up all subprocess-related objects
+            del p, parent_conn, child_conn, result_bytes
+            gc.collect()
+
+            # Handle result or re-raise exception
+            if result['success']:
+                # Reconstruct Restock object from dict
+                restock_obj = Restock(result['data'])
+                # Clean up result dict
+                del result
+                gc.collect()
+                return restock_obj
+            else:
+                # Re-raise the exception that occurred in subprocess
+                exception_type = result['exception_type']
+                exception_msg = result.get('exception_message', '')
+                del result
+                gc.collect()
+
+                if exception_type == 'MoreThanOnePriceFound':
+                    raise MoreThanOnePriceFound()
+                else:
+                    raise Exception(f"{exception_type}: {exception_msg}")
+
+        except Exception as e:
+            # If multiprocessing itself fails, log and fall back to direct call
+            logger.warning(f"Subprocess extraction failed: {e}, falling back to direct call")
+            gc.collect()
+            return get_itemprop_availability(html_content)
+    else:
+        # Non-Linux: direct call (no subprocess overhead needed)
+        return get_itemprop_availability(html_content)
+
+
 # should return Restock()
 # add casting?
 def get_itemprop_availability(html_content) -> Restock:
@@ -196,8 +449,9 @@ class perform_site_check(difference_detection_processor):
         multiple_prices_found = False
 
         # Try built-in extraction first, this will scan metadata in the HTML
+        # On Linux, this runs in a subprocess to prevent lxml/extruct memory leaks
         try:
-            itemprop_availability = get_itemprop_availability(self.fetcher.content)
+            itemprop_availability = extract_itemprop_availability_safe(self.fetcher.content)
         except MoreThanOnePriceFound as e:
             # Don't raise immediately - let plugins try to handle this case
             # Plugins might be able to determine which price is correct
diff --git a/changedetectionio/processors/restock_diff/pure_python_extractor.py b/changedetectionio/processors/restock_diff/pure_python_extractor.py
new file mode 100644
index 00000000..b46920fd
--- /dev/null
+++ b/changedetectionio/processors/restock_diff/pure_python_extractor.py
@@ -0,0 +1,286 @@
+"""
+Pure Python metadata extractor - no lxml, no memory leaks.
+
+This module provides a fast, memory-efficient alternative to extruct for common
+e-commerce metadata extraction. It handles:
+- JSON-LD (covers 80%+ of modern sites)
+- OpenGraph meta tags
+- Basic microdata attributes
+
+Uses Python's built-in html.parser instead of lxml/libxml2, avoiding C-level
+memory allocation issues. For edge cases, the main processor can fall back to
+extruct (with subprocess isolation on Linux).
+"""
+
+from html.parser import HTMLParser
+import json
+import re
+from loguru import logger
+
+
+class JSONLDExtractor(HTMLParser):
+    """
+    Extract JSON-LD structured data from HTML.
+
+    Finds all <script type="application/ld+json"> tags and parses their content.
+    Handles multiple JSON-LD blocks on the same page.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.in_jsonld = False
+        self.data = []  # List of all parsed JSON-LD objects
+        self.current_script = []
+
+    def handle_starttag(self, tag, attrs):
+        if tag == 'script':
+            # Check if this is a JSON-LD script tag
+            for attr, value in attrs:
+                if attr == 'type' and value == 'application/ld+json':
+                    self.in_jsonld = True
+                    self.current_script = []
+                    break
+
+    def handle_data(self, data):
+        if self.in_jsonld:
+            self.current_script.append(data)
+
+    def handle_endtag(self, tag):
+        if tag == 'script' and self.in_jsonld:
+            # Parse the accumulated script content
+            script_content = ''.join(self.current_script)
+            if script_content.strip():
+                try:
+                    # Parse JSON (handles both objects and arrays)
+                    parsed = json.loads(script_content)
+                    if isinstance(parsed, list):
+                        self.data.extend(parsed)
+                    else:
+                        self.data.append(parsed)
+                except json.JSONDecodeError as e:
+                    logger.debug(f"Failed to parse JSON-LD: {e}")
+                    pass
+
+            self.in_jsonld = False
+            self.current_script = []
+
+
+class OpenGraphExtractor(HTMLParser):
+    """
+    Extract OpenGraph meta tags from HTML.
+
+    Finds <meta property="og:*"> tags commonly used for social media sharing.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.og_data = {}
+
+    def handle_starttag(self, tag, attrs):
+        if tag == 'meta':
+            attrs_dict = dict(attrs)
+            prop = attrs_dict.get('property', '')
+
+            # Extract OpenGraph properties
+            if prop.startswith('og:'):
+                content = attrs_dict.get('content', '')
+                if content:
+                    self.og_data[prop] = content
+
+
+class MicrodataExtractor(HTMLParser):
+    """
+    Extract basic microdata attributes from HTML.
+
+    Finds elements with itemprop attributes. This is a simplified extractor
+    that doesn't handle nested itemscope/itemtype hierarchies - for complex
+    cases, use extruct as fallback.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.microdata = {}
+        self.current_itemprop = None
+
+    def handle_starttag(self, tag, attrs):
+        attrs_dict = dict(attrs)
+
+        if 'itemprop' in attrs_dict:
+            itemprop = attrs_dict['itemprop']
+
+            # Price/currency/availability can be in content/href attributes
+            if itemprop == 'price':
+                if 'content' in attrs_dict:
+                    self.microdata['price'] = attrs_dict['content']
+                else:
+                    self.current_itemprop = 'price'
+
+            elif itemprop == 'priceCurrency':
+                if 'content' in attrs_dict:
+                    self.microdata['currency'] = attrs_dict['content']
+                else:
+                    self.current_itemprop = 'priceCurrency'
+
+            elif itemprop == 'availability':
+                # Can be in href (link) or content (meta)
+                if 'href' in attrs_dict:
+                    self.microdata['availability'] = attrs_dict['href']
+                elif 'content' in attrs_dict:
+                    self.microdata['availability'] = attrs_dict['content']
+                else:
+                    self.current_itemprop = 'availability'
+
+    def handle_data(self, data):
+        # Capture text content for itemprop elements
+        if self.current_itemprop == 'price':
+            # Try to extract numeric price from text
+            try:
+                price_text = re.sub(r'[^\d.]', '', data.strip())
+                if price_text:
+                    self.microdata['price'] = float(price_text)
+            except ValueError:
+                pass
+        elif self.current_itemprop == 'priceCurrency':
+            currency = data.strip()
+            if currency:
+                self.microdata['currency'] = currency
+        elif self.current_itemprop == 'availability':
+            availability = data.strip()
+            if availability:
+                self.microdata['availability'] = availability
+
+    def handle_endtag(self, tag):
+        # Reset current itemprop after closing tag
+        self.current_itemprop = None
+
+
+def extract_metadata_pure_python(html_content):
+    """
+    Extract structured metadata from HTML using pure Python parsers.
+
+    Returns a dict with three keys:
+    - 'json-ld': List of parsed JSON-LD objects
+    - 'opengraph': Dict of OpenGraph properties
+    - 'microdata': Dict of microdata properties
+
+    Args:
+        html_content: HTML string to parse
+
+    Returns:
+        dict: Extracted metadata in three formats
+    """
+    result = {
+        'json-ld': [],
+        'opengraph': {},
+        'microdata': {}
+    }
+
+    # Extract JSON-LD
+    try:
+        jsonld_extractor = JSONLDExtractor()
+        jsonld_extractor.feed(html_content)
+        result['json-ld'] = jsonld_extractor.data
+        logger.trace(f"Pure Python: Found {len(jsonld_extractor.data)} JSON-LD blocks")
+    except Exception as e:
+        logger.debug(f"JSON-LD extraction failed: {e}")
+
+    # Extract OpenGraph
+    try:
+        og_extractor = OpenGraphExtractor()
+        og_extractor.feed(html_content)
+        result['opengraph'] = og_extractor.og_data
+        if result['opengraph']:
+            logger.trace(f"Pure Python: Found {len(og_extractor.og_data)} OpenGraph tags")
+    except Exception as e:
+        logger.debug(f"OpenGraph extraction failed: {e}")
+
+    # Extract Microdata
+    try:
+        microdata_extractor = MicrodataExtractor()
+        microdata_extractor.feed(html_content)
+        result['microdata'] = microdata_extractor.microdata
+        if result['microdata']:
+            logger.trace(f"Pure Python: Found microdata: {result['microdata']}")
+    except Exception as e:
+        logger.debug(f"Microdata extraction failed: {e}")
+
+    return result
+
+
+def query_price_availability(extracted_data):
+    """
+    Query extracted metadata for price and availability information.
+
+    Uses jsonpath_ng to query JSON-LD data (same approach as extruct).
+    Falls back to OpenGraph and microdata if JSON-LD doesn't have the data.
+
+    Args:
+        extracted_data: Dict from extract_metadata_pure_python()
+
+    Returns:
+        dict: {'price': float, 'currency': str, 'availability': str}
+    """
+    from jsonpath_ng import parse
+
+    result = {}
+
+    # 1. Try JSON-LD first (most reliable and common)
+    for data in extracted_data.get('json-ld', []):
+        try:
+            # Use jsonpath to find price/availability anywhere in the structure
+            price_parse = parse('$..(price|Price)')
+            availability_parse = parse('$..(availability|Availability)')
+            currency_parse = parse('$..(priceCurrency|currency|priceCurrency)')
+
+            price_results = [m.value for m in price_parse.find(data)]
+            if price_results and not result.get('price'):
+                # Handle various price formats
+                price_val = price_results[0]
+                if isinstance(price_val, (int, float)):
+                    result['price'] = float(price_val)
+                elif isinstance(price_val, str):
+                    # Extract numeric value from string
+                    try:
+                        result['price'] = float(re.sub(r'[^\d.]', '', price_val))
+                    except ValueError:
+                        pass
+
+            avail_results = [m.value for m in availability_parse.find(data)]
+            if avail_results and not result.get('availability'):
+                result['availability'] = str(avail_results[0])
+
+            curr_results = [m.value for m in currency_parse.find(data)]
+            if curr_results and not result.get('currency'):
+                result['currency'] = str(curr_results[0])
+
+            # If we found price, this JSON-LD block is good
+            if result.get('price'):
+                logger.debug(f"Pure Python: Found price data in JSON-LD: {result}")
+                break
+
+        except Exception as e:
+            logger.debug(f"Error querying JSON-LD: {e}")
+            continue
+
+    # 2. Try OpenGraph if JSON-LD didn't provide everything
+    og_data = extracted_data.get('opengraph', {})
+    if not result.get('price') and 'og:price:amount' in og_data:
+        try:
+            result['price'] = float(og_data['og:price:amount'])
+        except ValueError:
+            pass
+    if not result.get('currency') and 'og:price:currency' in og_data:
+        result['currency'] = og_data['og:price:currency']
+    if not result.get('availability') and 'og:availability' in og_data:
+        result['availability'] = og_data['og:availability']
+
+    # 3. Use microdata as last resort
+    microdata = extracted_data.get('microdata', {})
+    if not result.get('price') and 'price' in microdata:
+        result['price'] = microdata['price']
+    if not result.get('currency') and 'currency' in microdata:
+        result['currency'] = microdata['currency']
+    if not result.get('availability') and 'availability' in microdata:
+        result['availability'] = microdata['availability']
+
+    return result