rebaseline_server: multithreaded loading/diffing of images

gm/rebaseline_server/imagediffdb.py

 #!/usr/bin/python
 
 """
 Copyright 2013 Google Inc.
 
 Use of this source code is governed by a BSD-style license that can be
 found in the LICENSE file.
 
 Calulate differences between image pairs, and store them in a database.
 """
 
 import contextlib
 import csv
+import errno
 import logging
+import Queue
 import os
 import re
 import shutil
 import sys
 import tempfile
+import time
+import threading
 import urllib
 try:
   from PIL import Image, ImageChops
 except ImportError:
   raise ImportError('Requires PIL to be installed; see '
                     + 'http://www.pythonware.com/products/pil/')
 
 # Set the PYTHONPATH to include the tools directory.
 sys.path.append(
     os.path.join(
         os.path.dirname(os.path.realpath(__file__)), os.pardir, os.pardir,
                         'tools'))
 import find_run_binary
 
 SKPDIFF_BINARY = find_run_binary.find_path_to_program('skpdiff')
 
 DEFAULT_IMAGE_SUFFIX = '.png'
 DEFAULT_IMAGES_SUBDIR = 'images'
+DEFAULT_NUM_WORKERS = 8
 
 DISALLOWED_FILEPATH_CHAR_REGEX = re.compile('[^\w\-]')
 
 DIFFS_SUBDIR = 'diffs'
 WHITEDIFFS_SUBDIR = 'whitediffs'
 
 VALUES_PER_BAND = 256
 
 # Keys used within DiffRecord dictionary representations.
 # NOTE: Keep these in sync with static/constants.js
 KEY__DIFFERENCE_DATA__MAX_DIFF_PER_CHANNEL = 'maxDiffPerChannel'
 KEY__DIFFERENCE_DATA__NUM_DIFF_PIXELS = 'numDifferingPixels'
 KEY__DIFFERENCE_DATA__PERCENT_DIFF_PIXELS = 'percentDifferingPixels'
 KEY__DIFFERENCE_DATA__PERCEPTUAL_DIFF = 'perceptualDifference'
 KEY__DIFFERENCE_DATA__WEIGHTED_DIFF = 'weightedDiffMeasure'
 
+# Special values within ImageDiffDB._diff_dict
+DIFFRECORD_FAILED = 'failed'
+DIFFRECORD_PENDING = 'pending'
+
+# TODO(epoger): Temporary(?) list to keep track of how many times we download
+# the same file in multiple threads.
+global_file_collisions = 0
+
 
 class DiffRecord(object):
   """ Record of differences between two images. """
 
   def __init__(self, storage_root,
                expected_image_url, expected_image_locator,
                actual_image_url, actual_image_locator,
                expected_images_subdir=DEFAULT_IMAGES_SUBDIR,
                actual_images_subdir=DEFAULT_IMAGES_SUBDIR,
                image_suffix=DEFAULT_IMAGE_SUFFIX):
     """Download this pair of images (unless we already have them on local disk),
     and prepare a DiffRecord for them.
 
-    TODO(epoger): Make this asynchronously download images, rather than blocking
-    until the images have been downloaded and processed.
-
     Args:
       storage_root: root directory on local disk within which we store all
           images
       expected_image_url: file or HTTP url from which we will download the
           expected image
       expected_image_locator: a unique ID string under which we will store the
           expected image within storage_root (probably including a checksum to
           guarantee uniqueness)
       actual_image_url: file or HTTP url from which we will download the
           actual image
@@ skipping 132 matching lines @@
         KEY__DIFFERENCE_DATA__WEIGHTED_DIFF: self.get_weighted_diff_measure(),
         KEY__DIFFERENCE_DATA__MAX_DIFF_PER_CHANNEL: self._max_diff_per_channel,
         KEY__DIFFERENCE_DATA__PERCEPTUAL_DIFF: self._perceptual_difference,
     }
 
 
 class ImageDiffDB(object):
   """ Calculates differences between image pairs, maintaining a database of
   them for download."""
 
-  def __init__(self, storage_root):
+  def __init__(self, storage_root, num_workers=DEFAULT_NUM_WORKERS):
     """
     Args:
       storage_root: string; root path within the DB will store all of its stuff
+      num_workers: integer; number of worker threads to spawn
     """
     self._storage_root = storage_root
 
     # Dictionary of DiffRecords, keyed by (expected_image_locator,
     # actual_image_locator) tuples.
+    # Values can also be DIFFRECORD_PENDING, DIFFRECORD_FAILED.
     self._diff_dict = {}
 
+    # Set up the queue for asynchronously loading DiffRecords, and start the
+    # worker threads reading from it.
+    self._tasks_queue = Queue.Queue(maxsize=2*num_workers)
+    self._workers = []
+    for i in range(num_workers):
+      worker = threading.Thread(target=self.worker, args=(i,))
+      worker.daemon = True
+      worker.start()
+      self._workers.append(worker)
+
+  def worker(self, worker_num):
+    """Launch a worker thread that pulls tasks off self._tasks_queue.
+
+    Args:
+      worker_num: (integer) which worker this is
+    """
+    while True:
+      params = self._tasks_queue.get()
+      key, expected_image_url, actual_image_url = params
+      try:
+        diff_record = DiffRecord(
+            self._storage_root,
+            expected_image_url=expected_image_url,
+            expected_image_locator=key[0],
+            actual_image_url=actual_image_url,
+            actual_image_locator=key[1])
+      except Exception:
+        logging.exception(
+            'exception while creating DiffRecord for key %s' % str(key))
+        diff_record = DIFFRECORD_FAILED
+      self._diff_dict[key] = diff_record
+
   def add_image_pair(self,
                      expected_image_url, expected_image_locator,
                      actual_image_url, actual_image_locator):
     """Download this pair of images (unless we already have them on local disk),
     and prepare a DiffRecord for them.
 
-    TODO(epoger): Make this asynchronously download images, rather than blocking
-    until the images have been downloaded and processed.
-    When we do that, we should probably add a new method that will block
-    until all of the images have been downloaded and processed.  Otherwise,
-    we won't know when it's safe to start calling get_diff_record().
-    jcgregorio notes: maybe just make ImageDiffDB thread-safe and create a
-    thread-pool/worker queue at a higher level that just uses ImageDiffDB?
+    This method will block until the images are downloaded and DiffRecord is
+    available by calling get_diff_record().
 
     Args:
       expected_image_url: file or HTTP url from which we will download the
           expected image
       expected_image_locator: a unique ID string under which we will store the
           expected image within storage_root (probably including a checksum to
           guarantee uniqueness)
       actual_image_url: file or HTTP url from which we will download the
           actual image
       actual_image_locator: a unique ID string under which we will store the
           actual image within storage_root (probably including a checksum to
           guarantee uniqueness)
+
+    Raises:
+      Exception if we are unable to create a DiffRecord for this image pair.
     """
-    expected_image_locator = _sanitize_locator(expected_image_locator)
-    actual_image_locator = _sanitize_locator(actual_image_locator)
-    key = (expected_image_locator, actual_image_locator)
+    key = _generate_key(expected_image_locator, actual_image_locator)
     if not key in self._diff_dict:
       try:
         new_diff_record = DiffRecord(
             self._storage_root,
             expected_image_url=expected_image_url,
             expected_image_locator=expected_image_locator,
             actual_image_url=actual_image_url,
             actual_image_locator=actual_image_locator)
       except Exception:
         # If we can't create a real DiffRecord for this (expected, actual) pair,
         # store None and the UI will show whatever information we DO have.
         # Fixes http://skbug.com/2368 .
         logging.exception(
             'got exception while creating a DiffRecord for '
             'expected_image_url=%s , actual_image_url=%s; returning None' % (
                 expected_image_url, actual_image_url))
         new_diff_record = None
       self._diff_dict[key] = new_diff_record
 
+  def add_image_pair_async(self,
+                           expected_image_url, expected_image_locator,
+                           actual_image_url, actual_image_locator):
+    """Download this pair of images (unless we already have them on local disk),
+    and prepare a DiffRecord for them.
+
+    This method will return quickly; calls to get_diff_record() will block
+    until the DiffRecord is available (or we have given up on creating it).
+
+    Args:
+      expected_image_url: file or HTTP url from which we will download the
+          expected image
+      expected_image_locator: a unique ID string under which we will store the
+          expected image within storage_root (probably including a checksum to
+          guarantee uniqueness)
+      actual_image_url: file or HTTP url from which we will download the
+          actual image
+      actual_image_locator: a unique ID string under which we will store the
+          actual image within storage_root (probably including a checksum to
+          guarantee uniqueness)
+    """
+    key = _generate_key(expected_image_locator, actual_image_locator)
+    if not key in self._diff_dict:
+      # If we have already requested a diff between these two images,
+      # we don't need to request it again.
+      #
+      # Threading note: If multiple threads called into this method with the
+      # same key at the same time, there will be multiple tasks on the queue
+      # with the same key.  But that's OK; they will both complete successfully,
+      # and just waste a little time in the process. Nothing will break.

[rmistry, 2014/04/14 17:59:06]

Can we use semaphores/locks here? Based on your tests how many
global_file_collisions show up?

Also, you said it runs about 60% faster, does that mean it now takes about 10-15
mins during startup (with gpu cpu comparison)?

[epoger, 2014/04/14 18:10:51]

In my tests, I didn't see any global_file_collisions yet.  Thus, I concluded we
were better off not incurring the cost of an additional lock.  (Given that the
downside is not disastrous, just takes some extra time.)
Well..... no.  I haven't tried with the gpu/cpu comparison yet.

This is the command I was running, on my Ubiquity instance:

$ rm -rf gm/rebaseline_server/static/generated-images/ &&
gm/rebaseline_server/server.py --export --actuals-repo ""

(Clear out disk cache to force download of actual/expected images for all with
differing checksums, but don't update the actuals checksums from skia-autogen so
we have consistent results.)

It consistenly took ~139 seconds before this change, and ~58 seconds after.

I'm sure there is plenty more speedup to be found.  I will keep looking...

+      self._diff_dict[key] = DIFFRECORD_PENDING
+      self._tasks_queue.put((key, expected_image_url, actual_image_url))
+
   def get_diff_record(self, expected_image_locator, actual_image_locator):
     """Returns the DiffRecord for this image pair.
 
-    Raises a KeyError if we don't have a DiffRecord for this image pair.
+    Args:
+      expected_image_locator: a unique ID string under which we will store the
+          expected image within storage_root (probably including a checksum to
+          guarantee uniqueness)
+      actual_image_locator: a unique ID string under which we will store the
+          actual image within storage_root (probably including a checksum to
+          guarantee uniqueness)
+
+    Returns the DiffRecord for this image pair, or None if we were unable to
+    generate one.
     """
-    key = (_sanitize_locator(expected_image_locator),
-           _sanitize_locator(actual_image_locator))
-    return self._diff_dict[key]
+    key = _generate_key(expected_image_locator, actual_image_locator)
+    diff_record = self._diff_dict[key]
+
+    # If we have no results yet, block until we do.
+    while diff_record == DIFFRECORD_PENDING:
+      time.sleep(1)
+      diff_record = self._diff_dict[key]
+
+    # Once we have the result...
+    if diff_record == DIFFRECORD_FAILED:
+      logging.error(
+          'failed to create a DiffRecord for expected_image_locator=%s , '
+          'actual_image_locator=%s' % (
+              expected_image_locator, actual_image_locator))
+      return None
+    else:
+      return diff_record
 
 
 # Utility functions
 
 def _calculate_weighted_diff_metric(histogram, num_pixels):
   """Given the histogram of a diff image (per-channel diff at each
   pixel between two images), calculate the weighted diff metric (a
   stab at how different the two images really are).
 
   TODO(epoger): Delete this function, now that we have perceptual diff?
@@ skipping 68 matching lines @@
 def _download_and_open_image(local_filepath, url):
   """Open the image at local_filepath; if there is no file at that path,
   download it from url to that path and then open it.
 
   Args:
     local_filepath: path on local disk where the image should be stored
     url: URL from which we can download the image if we don't have it yet
 
   Returns: a PIL image object
   """
+  global global_file_collisions
   if not os.path.exists(local_filepath):
     _mkdir_unless_exists(os.path.dirname(local_filepath))
     with contextlib.closing(urllib.urlopen(url)) as url_handle:
-      with open(local_filepath, 'wb') as file_handle:
+
+      # First download the file contents into a unique filename, and
+      # then rename that file.  That way, if multiple threads are downloading
+      # the same filename at the same time, they won't interfere with each
+      # other (they will both download the file, and one will "win" in the end)
+      temp_filename = '%s-%d' % (local_filepath,
+                                 threading.current_thread().ident)
+      with open(temp_filename, 'wb') as file_handle:
         shutil.copyfileobj(fsrc=url_handle, fdst=file_handle)
+
+      # Keep count of how many colliding downloads we encounter;
+      # if it's a large number, we may want to change our download strategy
+      # to minimize repeated downloads.
+      if os.path.exists(local_filepath):
+        global_file_collisions += 1
+      else:
+        os.rename(temp_filename, local_filepath)
+
   return _open_image(local_filepath)
 
 
 def _open_image(filepath):
   """Wrapper for Image.open(filepath) that yields more useful error messages.
 
   Args:
     filepath: path on local disk to load image from
 
   Returns: a PIL image object
@@ skipping 20 matching lines @@
   _mkdir_unless_exists(os.path.dirname(filepath))
   image.save(filepath, format)
 
 
 def _mkdir_unless_exists(path):
   """Unless path refers to an already-existing directory, create it.
 
   Args:
     path: path on local disk
   """
-  if not os.path.isdir(path):
+  try:
     os.makedirs(path)
+  except OSError as e:
+    if e.errno == errno.EEXIST:
+      pass
 
 
 def _sanitize_locator(locator):
   """Returns a sanitized version of a locator (one in which we know none of the
   characters will have special meaning in filenames).
 
   Args:
     locator: string, or something that can be represented as a string
   """
   return DISALLOWED_FILEPATH_CHAR_REGEX.sub('_', str(locator))
 
 
+def _generate_key(expected_image_locator, actual_image_locator):
+  """Returns a key suitable for looking up this image pair.
+
+  Args:
+    expected_image_locator: a unique ID string under which we will store the
+        expected image within storage_root (probably including a checksum to
+        guarantee uniqueness)
+    actual_image_locator: a unique ID string under which we will store the
+        actual image within storage_root (probably including a checksum to
+        guarantee uniqueness)
+  """
+  return (_sanitize_locator(expected_image_locator),
+          _sanitize_locator(actual_image_locator))
+
+
 def _get_difference_locator(expected_image_locator, actual_image_locator):
   """Returns the locator string used to look up the diffs between expected_image
   and actual_image.
 
   We must keep this function in sync with getImageDiffRelativeUrl() in
   static/loader.js
 
   Args:
     expected_image_locator: locator string pointing at expected image
     actual_image_locator: locator string pointing at actual image
 
   Returns: already-sanitized locator where the diffs between expected and
       actual images can be found
   """
   return "%s-vs-%s" % (_sanitize_locator(expected_image_locator),
                        _sanitize_locator(actual_image_locator))