Fix issue with cache exceeding set size

There were two main issues with the current implementation:
  - the size of files was considered to be the number of bytes which is
    not the best for large number of small files which can happen when the
    cache is dealing with artifacts (lots of small ones). The fix is to reason
    based on the block size.
  - there was a possibility where a file could be marked as deletable and would
    then be re-created only to be then deleted again. This fix addresses this
    situation.

services/ui_backend_service/data/cache/client/cache_server.py

@@ -294,6 +294,9 @@ def _error_callback(self, worker, res):
 def cli(root=None,
  max_actions=None,
  max_size=None):
+ # NOTE: The store will only be accessed by this process. The processes
+ # in the pool never touch the store itself. This is done in the __init__ and
+ # terminate methods in Worker which all happen in this process.
  store = CacheStore(root, max_size, echo)
  Scheduler(store, max_actions).loop()
 

services/ui_backend_service/data/cache/client/cache_store.py

@@ -1,3 +1,4 @@
+import math
 import os
 import time
 import shutil
@@ -28,7 +29,7 @@ def makedirs(path):
 
 
 def key_filename(key):
- return hashlib.sha1(key.encode('utf-8')).hexdigest()
+ return hashlib.sha1(key.encode("utf-8")).hexdigest()
 
 
 def object_path(root, key):
@@ -38,7 +39,7 @@ def object_path(root, key):
 
 
 def stream_path(root, key):
- return object_path(root, key) + '.stream'
+ return object_path(root, key) + ".stream"
 
 
 def is_safely_readable(path):
@@ -50,11 +51,17 @@ def is_safely_readable(path):
 
 
 def filesize(path):
- return os.stat(path).st_size
+ try:
+ blk_sz = os.statvfs(path).f_bsize
+ sz = os.stat(path).st_size
+ if sz == 0:
+ return blk_sz
+ return blk_sz * math.ceil(sz / blk_sz)
+ except Exception:
+ return None
 
 
 class CacheStore(object):
-
  def __init__(self, root, max_size, echo, fill_factor=0.8):
  self.root = os.path.abspath(root)
  self.tmproot = self._init_temp(self.root)
@@ -79,7 +86,7 @@ def warn(self, ex, msg):
  self.echo("IO ERROR: (%s) %s" % (ex, msg))
 
  def _init_temp(self, root):
- tmproot = os.path.join(root, 'tmp')
+ tmproot = os.path.join(root, "tmp")
  if os.path.exists(tmproot):
  self.safe_fileop(shutil.rmtree, tmproot)
  makedirs(tmproot)
@@ -94,18 +101,23 @@ def _init_gc(self, root):
  self.safe_fileop(os.unlink, path)
  else:
  stat_res = self.safe_fileop(os.stat, path)
- if stat_res:
+ sz = filesize(path)
+ if stat_res and sz is not None:
  _, info = stat_res
  if info.st_mtime == TIMESTAMP_FOR_DELETABLE:
- self.safe_fileop(os.unlink, path)
+ # If we can't unlink it, we add it to the gc_queue
+ # to try at least another time
+ if not self.safe_fileop(os.unlink, path):
+ self.gc_queue[path] = (time.time(), sz)
  elif info.st_mtime == TIMESTAMP_FOR_DISPOSABLE:
- self.disposables_queue[path] = info.st_size
+ self.disposables_queue[path] = sz
  else:
- objects.append((info.st_mtime, (path, info.st_size)))
+ objects.append((info.st_mtime, (path, sz)))
 
  self.objects_queue.update(x for _, x in sorted(objects))
- self.total_size = sum(self.disposables_queue.values()) +\
- sum(self.objects_queue.values())
+ self.total_size = sum(self.disposables_queue.values()) + sum(
+ self.objects_queue.values()
+ )
 
  # It is possible that the datastore contains more than gc_watermark
  # bytes. To ensure that we start below the gc_watermark, we run the GC:
@@ -115,17 +127,18 @@ def _init_gc(self, root):
  # yet, so we can safely delete the marked objects without a quarantine:
  self._gc_objects(quarantine=-1)
 
- self.echo("Cache initialized with %d permanents objects, "
-  "%d disposable objects, totaling %d bytes."
-  % (len(self.objects_queue),
-  len(self.disposables_queue),
-  self.total_size))
+ self.echo(
+ "Cache initialized with %d permanents objects, "
+ "%d disposable objects, totaling %d bytes."
+ % (len(self.objects_queue), len(self.disposables_queue), self.total_size)
+ )
 
  def _gc_objects(self, quarantine=GC_MARKER_QUARANTINE):
  def mark_for_deletion(path, size):
- self.safe_fileop(os.utime, path, (TIMESTAMP_FOR_DELETABLE,
- TIMESTAMP_FOR_DELETABLE))
- self.gc_queue[path] = (time.time(), size)
+ if self.safe_fileop(
+ os.utime, path, (TIMESTAMP_FOR_DELETABLE, TIMESTAMP_FOR_DELETABLE)
+ ):
+ self.gc_queue[path] = (time.time(), size)
 
  # 1) delete marked objects that are past their quarantine period
  limit = time.time() - quarantine
@@ -135,6 +148,13 @@ def mark_for_deletion(path, size):
  if self.safe_fileop(os.unlink, path):
  del self.gc_queue[path]
  self.total_size -= size
+ else:
+ self.warn(
+ "Could not remove file at '%s' -- removing from GC" % path
+ )
+ # We still remove to prevent the garbage collector from
+ # being stuck a few lines below.
+ del self.gc_queue[path]
  else:
  break
 
@@ -169,14 +189,15 @@ def ensure_path(self, path):
 
  def open_tempdir(self, token, action_name, stream_key):
  self._gc_objects()
-
  if self.total_size > self.max_size:
- self.warn(None, "Cache soft limit reached! Used %d bytes, max %s bytes"
- % (self.total_size, self.max_size))
+ self.warn(
+ None,
+ "Cache soft limit reached! Used %d bytes, max %s bytes"
+ % (self.total_size, self.max_size),
+ )
 
  try:
- tmp = tempfile.mkdtemp(prefix='cache_action_%s.' % token,
- dir=self.tmproot)
+ tmp = tempfile.mkdtemp(prefix="cache_action_%s." % token, dir=self.tmproot)
  except Exception as ex:
  msg = "Could not create a temp directory for request %s" % token
  self.warn(ex, msg)
@@ -187,7 +208,7 @@ def open_tempdir(self, token, action_name, stream_key):
  # make sure that the new symlink points at a valid (empty!)
  # file by creating a dummy destination file
  self.ensure_path(src)
- open_res = self.safe_fileop(open, dst, 'w')
+ open_res = self.safe_fileop(open, dst, "w")
  if open_res:
  _, f = open_res
  f.close()
@@ -198,8 +219,7 @@ def open_tempdir(self, token, action_name, stream_key):
  # simultaneously. We don't consider an existing
  # symlink (errno 17) to be an error.
  if ex.errno != 17:
- err = "Could not create a symlink %s->%s"\
- % (src, dst)
+ err = "Could not create a symlink %s->%s" % (src, dst)
  self.warn(ex, err)
  except Exception as ex:
  self.warn(ex, "Unknown error")
@@ -227,6 +247,12 @@ def _insert(queue, key, value):
  # previous entry first
  queue.pop(key, None)
  queue[key] = value
+ # If we are inserting something in disposables_queue or objects_queue,
+ # we make sure it is no longer in the gc_queue. This can happen if, for
+ # example, an object is marked as deletable, is therefore not "readable"
+ # and is therefore re-created.
+ if key in self.gc_queue:
+ del self.gc_queue[key]
 
  disposables = frozenset(disposable_keys)
  missing = []
@@ -243,21 +269,19 @@ def _insert(queue, key, value):
  if os.path.exists(src):
  dst = object_path(self.root, key)
  self.ensure_path(dst)
- stat_res = self.safe_fileop(os.stat, src)
- if stat_res and self.safe_fileop(os.rename, src, dst):
- _, info = stat_res
- size = info.st_size
+ sz = filesize(src)
+ if sz is not None and self.safe_fileop(os.rename, src, dst):
  if key in disposables:
  # we proceed even if we fail to mark the object as
  # disposable. It just means that during a possible
  # restart the object is treated as a non-disposable
  # object
  tstamp = TIMESTAMP_FOR_DISPOSABLE
  self.safe_fileop(os.utime, dst, (tstamp, tstamp))
- _insert(self.disposables_queue, dst, size)
+ _insert(self.disposables_queue, dst, sz)
  else:
- _insert(self.objects_queue, dst, size)
- self.total_size += size
+ _insert(self.objects_queue, dst, sz)
+ self.total_size += sz
  else:
  missing.append(key)