[Notes] [Git][BuildStream/buildstream][valentindavid/cache_server_fill

Valentin David pushed to branch valentindavid/cache_server_fill_up at BuildStream / buildstream

Commits:

ed6647b0

by Valentin David at 2018-11-02T15:04:39Z

Move cas server from ref-based to object-based garbage collection.

Also add locking to not collect in parallel. Collect a bit more than
needed. Make the cache sizes configurable.

2 changed files:

buildstream/_artifactcache/cascache.py
buildstream/_artifactcache/casserver.py

Changes:

buildstream/_artifactcache/cascache.py

@@ -593,6 +593,41 @@ class CASCache(ArtifactCache):
          # first element of this list will be the file modified earliest.
          return [ref for _, ref in sorted(zip(mtimes, refs))]
 +    # list_objects():
 +    #
 +    # List cached objects in Least Recently Modified (LRM) order.
 +    #
 +    # Returns:
 +    #     (list) - A list of objects and timestamps in LRM order
 +    #
 +    def list_objects(self):
 +        objs = []
 +        mtimes = []
++
 +        for root, _, files in os.walk(os.path.join(self.casdir, 'objects')):
 +            for filename in files:
 +                obj_path = os.path.join(root, filename)
 +                try:
 +                    mtimes.append(os.path.getmtime(obj_path))
 +                except FileNotFoundError:
 +                    pass
 +                else:
 +                    objs.append(obj_path)
++
 +        # NOTE: Sorted will sort from earliest to latest, thus the
 +        # first element of this list will be the file modified earliest.
 +        return sorted(zip(mtimes, objs))
++
 +    def clean_up_refs_until(self, time):
 +        ref_heads = os.path.join(self.casdir, 'refs', 'heads')
++
 +        for root, _, files in os.walk(ref_heads):
 +            for filename in files:
 +                ref_path = os.path.join(root, filename)
 +                # Obtain the mtime (the time a file was last modified)
 +                if os.path.getmtime(ref_path) < time:
 +                    os.unlink(ref_path)
++
      # remove():
+     #
      # Removes the given symbolic ref from the repo.

buildstream/_artifactcache/casserver.py

@@ -24,9 +24,9 @@ import signal
  import sys
  import tempfile
  import uuid
 -import time
  import errno
  import ctypes
 +import threading
  import click
  import grpc
@@ -62,7 +62,9 @@ class ArtifactTooLargeException(Exception):
  #     repo (str): Path to CAS repository
  #     enable_push (bool): Whether to allow blob uploads and artifact updates
+ #
 -def create_server(repo, *, enable_push):
 +def create_server(repo, *, enable_push,
 +                  max_head_size=int(10e9),
 +                  min_head_size=int(2e9)):
      context = Context()
      context.artifactdir = os.path.abspath(repo)
@@ -72,11 +74,13 @@ def create_server(repo, *, enable_push):
      max_workers = (os.cpu_count() or 1) * 5
      server = grpc.server(futures.ThreadPoolExecutor(max_workers))
 +    cache_cleaner = _CacheCleaner(artifactcache, max_head_size, min_head_size)
++
      bytestream_pb2_grpc.add_ByteStreamServicer_to_server(
 -        _ByteStreamServicer(artifactcache, enable_push=enable_push), server)
 +        _ByteStreamServicer(artifactcache, cache_cleaner, enable_push=enable_push), server)
      remote_execution_pb2_grpc.add_ContentAddressableStorageServicer_to_server(
 -        _ContentAddressableStorageServicer(artifactcache, enable_push=enable_push), server)
 +        _ContentAddressableStorageServicer(artifactcache, cache_cleaner, enable_push=enable_push), server)
      remote_execution_pb2_grpc.add_CapabilitiesServicer_to_server(
          _CapabilitiesServicer(), server)
@@ -94,9 +98,19 @@ def create_server(repo, *, enable_push):
  @click.option('--client-certs', help="Public client certificates for TLS (PEM-encoded)")
  @click.option('--enable-push', default=False, is_flag=True,
                help="Allow clients to upload blobs and update artifact cache")
 +@click.option('--head-room-min', type=click.INT,
 +              help="Disk head room minimum in bytes",
 +              default=2e9)
 +@click.option('--head-room-max', type=click.INT,
 +              help="Disk head room maximum in bytes",
 +              default=10e9)
  @click.argument('repo')
 -def server_main(repo, port, server_key, server_cert, client_certs, enable_push):
 -    server = create_server(repo, enable_push=enable_push)
 +def server_main(repo, port, server_key, server_cert, client_certs, enable_push,
 +                head_room_min, head_room_max):
 +    server = create_server(repo,
 +                           max_head_size=head_room_max,
 +                           min_head_size=head_room_min,
 +                           enable_push=enable_push)
      use_tls = bool(server_key)
@@ -168,11 +182,12 @@ class _FallocateCall:
  class _ByteStreamServicer(bytestream_pb2_grpc.ByteStreamServicer):
 -    def __init__(self, cas, *, enable_push):
 +    def __init__(self, cas, cache_cleaner, *, enable_push):
          super().__init__()
          self.cas = cas
          self.enable_push = enable_push
          self.fallocate = _FallocateCall()
 +        self.cache_cleaner = cache_cleaner
      def Read(self, request, context):
          resource_name = request.resource_name
@@ -234,7 +249,7 @@ class _ByteStreamServicer(bytestream_pb2_grpc.ByteStreamServicer):
                          if client_digest.size_bytes == 0:
                              break
                          try:
 -                            _clean_up_cache(self.cas, client_digest.size_bytes)
 +                            self.cache_cleaner.clean_up(client_digest.size_bytes)
                          except ArtifactTooLargeException as e:
                              context.set_code(grpc.StatusCode.RESOURCE_EXHAUSTED)
                              context.set_details(str(e))
@@ -280,10 +295,11 @@ class _ByteStreamServicer(bytestream_pb2_grpc.ByteStreamServicer):
  class _ContentAddressableStorageServicer(remote_execution_pb2_grpc.ContentAddressableStorageServicer):
 -    def __init__(self, cas, *, enable_push):
 +    def __init__(self, cas, cache_cleaner, *, enable_push):
          super().__init__()
          self.cas = cas
          self.enable_push = enable_push
 +        self.cache_cleaner = cache_cleaner
      def FindMissingBlobs(self, request, context):
          response = remote_execution_pb2.FindMissingBlobsResponse()
@@ -352,7 +368,7 @@ class _ContentAddressableStorageServicer(remote_execution_pb2_grpc.ContentAddres
                  continue
              try:
 -                _clean_up_cache(self.cas, digest.size_bytes)
 +                self.cache_cleaner.clean_up(digest.size_bytes)
                  with tempfile.NamedTemporaryFile(dir=self.cas.tmpdir) as out:
                      out.write(blob_request.data)
@@ -473,63 +489,80 @@ def _digest_from_upload_resource_name(resource_name):
          return None
 -# _clean_up_cache()
 -#
 -# Keep removing Least Recently Pushed (LRP) artifacts in a cache until there
 -# is enough space for the incoming artifact
 -#
 -# Args:
 -#   cas: CASCache object
 -#   object_size: The size of the object being received in bytes
 -#
 -# Returns:
 -#   int: The total bytes removed on the filesystem
 -#
 -def _clean_up_cache(cas, object_size):
 -    # Determine the available disk space, in bytes, of the file system
 -    # which mounts the repo
 -    stats = os.statvfs(cas.casdir)
 -    buffer_ = int(2e9)                # Add a 2 GB buffer
 -    free_disk_space = (stats.f_bavail * stats.f_bsize) - buffer_
 -    total_disk_space = (stats.f_blocks * stats.f_bsize) - buffer_
+-
 -    if object_size > total_disk_space:
 -        raise ArtifactTooLargeException("Artifact of size: {} is too large for "
 -                                        "the filesystem which mounts the remote "
 -                                        "cache".format(object_size))
+-
 -    if object_size <= free_disk_space:
 -        # No need to clean up
 -        return 0
+-
 -    # obtain a list of LRP artifacts
 -    LRP_artifacts = cas.list_artifacts()
+-
 -    keep_after = time.time() - _OBJECT_MIN_AGE
+-
 -    removed_size = 0  # in bytes
 -    if object_size - removed_size > free_disk_space:
 -        # First we try to see if some unreferenced objects became old
 -        # enough to be removed.
 -        removed_size += cas.prune(keep_after=keep_after)
+-
 -    while object_size - removed_size > free_disk_space:
 -        try:
 -            to_remove = LRP_artifacts.pop(0)  # The first element in the list is the LRP artifact
 -        except IndexError:
 -            # This exception is caught if there are no more artifacts in the list
 -            # LRP_artifacts. This means the the artifact is too large for the filesystem
 -            # so we abort the process
 -            raise ArtifactTooLargeException("Artifact of size {} is too large for "
 -                                            "the filesystem which mounts the remote "
 -                                            "cache".format(object_size))
 +class _CacheCleaner:
 -        cas.remove(to_remove, defer_prune=True)
 -        removed_size += cas.prune(keep_after=keep_after)
 +    __cleanup_cache_lock = threading.Lock()
 -    if removed_size > 0:
 -        logging.info("Successfully removed {} bytes from the cache".format(removed_size))
 -    else:
 -        logging.info("No artifacts were removed from the cache.")
 +    def __init__(self, cas, max_head_size, min_head_size=int(2e9)):
 +        self.__cas = cas
 +        self.__max_head_size = max_head_size
 +        self.__min_head_size = min_head_size
++
 +    def __has_space(self, object_size):
 +        stats = os.statvfs(self.__cas.casdir)
 +        free_disk_space = (stats.f_bavail * stats.f_bsize) - self.__min_head_size
 +        total_disk_space = (stats.f_blocks * stats.f_bsize) - self.__min_head_size
++
 +        if object_size > total_disk_space:
 +            raise ArtifactTooLargeException("Artifact of size: {} is too large for "
 +                                            "the filesystem which mounts the remote "
 +                                            "cache".format(object_size))
 -    return removed_size
 +        return object_size <= free_disk_space
++
 +    # _clean_up_cache()
 +    #
 +    # Keep removing Least Recently Pushed (LRP) artifacts in a cache until there
 +    # is enough space for the incoming artifact
 +    #
 +    # Args:
 +    #   object_size: The size of the object being received in bytes
 +    #
 +    # Returns:
 +    #   int: The total bytes removed on the filesystem
 +    #
 +    def clean_up(self, object_size):
 +        if self.__has_space(object_size):
 +            return 0
++
 +        with _CacheCleaner.__cleanup_cache_lock:
 +            if self.__has_space(object_size):
 +                # Another thread has done the cleanup for us
 +                return 0
++
 +            stats = os.statvfs(self.__cas.casdir)
 +            target_disk_space = (stats.f_bavail * stats.f_bsize) - self.__max_head_size
++
 +            # obtain a list of LRP artifacts
 +            LRP_objects = self.__cas.list_objects()
++
 +            removed_size = 0  # in bytes
++
 +            last_mtime = 0
++
 +            while object_size - removed_size > target_disk_space:
 +                try:
 +                    last_mtime, to_remove = LRP_objects.pop(0)  # The first element in the list is the LRP artifact
 +                except IndexError:
 +                    # This exception is caught if there are no more artifacts in the list
 +                    # LRP_artifacts. This means the the artifact is too large for the filesystem
 +                    # so we abort the process
 +                    raise ArtifactTooLargeException("Artifact of size {} is too large for "
 +                                                    "the filesystem which mounts the remote "
 +                                                    "cache".format(object_size))
++
 +                try:
 +                    size = os.stat(to_remove).st_size
 +                    os.unlink(to_remove)
 +                    removed_size += size
 +                except FileNotFoundError:
 +                    pass
++
 +            self.__cas.clean_up_refs_until(last_mtime)
++
 +            if removed_size > 0:
 +                logging.info("Successfully removed {} bytes from the cache".format(removed_size))
 +            else:
 +                logging.info("No artifacts were removed from the cache.")
++
 +            return removed_size

[Notes] [Git][BuildStream/buildstream][valentindavid/cache_server_fill_up] Move cas server from ref-based to object-based garbage collection.

Valentin David pushed to branch valentindavid/cache_server_fill_up at BuildStream / buildstream

Commits:

2 changed files:

Changes: