Add --no-cache option to build.py when building containers (#13182)

Add functionality to build.py to disable caching

ci/README.md

@@ -5,6 +5,8 @@ Docker containers
 
 You need docker and nvidia docker if you have a GPU.
 
+Also you need to run `pip3 install docker` as it uses the [docker python module](https://docker-py.readthedocs.io/en/stable/containers.html#)
+
 If you are in ubuntu an easy way to install Docker CE is executing the
 following script:
 

ci/build.py

@@ -101,6 +101,8 @@ def get_platforms(path: str = get_dockerfiles_path()) -> List[str]:
 
 def get_docker_tag(platform: str, registry: str) -> str:
     """:return: docker tag to be used for the container"""
+    if not registry:
+        registry = "mxnet_local"
     return "{0}/build.{1}".format(registry, platform)
 
 
@@ -112,14 +114,14 @@ def get_docker_binary(use_nvidia_docker: bool) -> str:
     return "nvidia-docker" if use_nvidia_docker else "docker"
 
 
-def build_docker(platform: str, docker_binary: str, registry: str, num_retries: int, use_cache: bool) -> str:
+def build_docker(platform: str, docker_binary: str, registry: str, num_retries: int, no_cache: bool) -> str:
     """
     Build a container for the given platform
     :param platform: Platform
     :param docker_binary: docker binary to use (docker/nvidia-docker)
     :param registry: Dockerhub registry name
     :param num_retries: Number of retries to build the docker image
-    :param use_cache: will pass cache_from to docker to use the previously pulled tag
+    :param no_cache: pass no-cache to docker to rebuild the images
     :return: Id of the top level image
     """
     tag = get_docker_tag(platform=platform, registry=registry)
@@ -144,7 +146,9 @@ def build_docker(platform: str, docker_binary: str, registry: str, num_retries:
            "-f", get_dockerfile(platform),
            "--build-arg", "USER_ID={}".format(os.getuid()),
            "--build-arg", "GROUP_ID={}".format(os.getgid())]
-    if use_cache:
+    if no_cache:
+        cmd.append("--no-cache")
+    elif registry:
         cmd.extend(["--cache-from", tag])
     cmd.extend(["-t", tag, get_dockerfiles_path()])
 
@@ -422,7 +426,7 @@ def main() -> int:
                         action='store_true')
 
     parser.add_argument("-d", "--docker-registry",
-                        help="Dockerhub registry name to retrieve cache from. Default is 'mxnetci'",
+                        help="Dockerhub registry name to retrieve cache from.",
                         default='mxnetci',
                         type=str)
 
@@ -431,10 +435,8 @@ def main() -> int:
                         default=1,
                         type=int)
 
-    parser.add_argument("-c", "--no-dockerhub-cache", action="store_true",
-                        help="Disables use of --cache-from option on docker build, allowing docker"
-                        " to use local layers for caching. If absent, we use the cache from dockerhub"
-                        " which is the default.")
+    parser.add_argument("--no-cache", action="store_true",
+                        help="passes --no-cache to docker build")
 
     parser.add_argument("command",
                         help="command to run in the container",
@@ -447,9 +449,6 @@ def main() -> int:
 
     args = parser.parse_args()
 
-    def use_cache():
-        return not args.no_dockerhub_cache or under_ci()
-
     command = list(chain(*args.command))
     docker_binary = get_docker_binary(args.nvidiadocker)
 
@@ -472,10 +471,10 @@ def signal_handler(signum, _):
     elif args.platform:
         platform = args.platform
         tag = get_docker_tag(platform=platform, registry=args.docker_registry)
-        if use_cache():
+        if args.docker_registry:
             load_docker_cache(tag=tag, docker_registry=args.docker_registry)
         build_docker(platform=platform, docker_binary=docker_binary, registry=args.docker_registry,
-                     num_retries=args.docker_build_retries, use_cache=use_cache())
+                     num_retries=args.docker_build_retries, no_cache=args.no_cache)
         if args.build_only:
             logging.warning("Container was just built. Exiting due to build-only.")
             return 0
@@ -512,10 +511,9 @@ def signal_handler(signum, _):
         logging.info("Artifacts will be produced in the build/ directory.")
         for platform in platforms:
             tag = get_docker_tag(platform=platform, registry=args.docker_registry)
-            if use_cache():
-                load_docker_cache(tag=tag, docker_registry=args.docker_registry)
+            load_docker_cache(tag=tag, docker_registry=args.docker_registry)
             build_docker(platform, docker_binary=docker_binary, registry=args.docker_registry,
-                         num_retries=args.docker_build_retries, use_cache=use_cache())
+                         num_retries=args.docker_build_retries, no_cache=args.no_cache)
             if args.build_only:
                 continue
             shutil.rmtree(buildir(), ignore_errors=True)

ci/docker_cache.py

@@ -125,6 +125,10 @@ def load_docker_cache(registry, docker_tag) -> None:
     :return: None
     """
     # We don't have to retag the image since it's already in the right format
+    if not registry:
+        return
+    assert docker_tag
+
     logging.info('Loading Docker cache for %s from %s', docker_tag, registry)
     pull_cmd = ['docker', 'pull', docker_tag]
     subprocess.call(pull_cmd)  # Don't throw an error if the image does not exist

ci/requirements.txt

@@ -0,0 +1 @@
+docker==3.5.0