manta-node 0.6b0.dev607__py3-none-any.whl → 0.6b0.dev615__py3-none-any.whl

manta_node/infrastructure/container/__init__.py

@@ -1,6 +1,6 @@
 """Container infrastructure - Docker runtime management."""
 
-from .docker_adapter import DockerRaiser
+from .docker_adapter import DockerRaiser, ImageFreshness
 from .image_executor import ImageExecutor
 from .manager import ContainerManager
 
@@ -8,4 +8,5 @@ __all__ = [
     "ContainerManager",
     "DockerRaiser",
     "ImageExecutor",
+    "ImageFreshness",
 ]

manta_node/infrastructure/container/docker_adapter.py

@@ -1,6 +1,7 @@
 import asyncio
 import concurrent.futures
 from datetime import datetime
+from enum import Enum
 
 from aiodocker import Docker
 from aiodocker.docker import DockerContainer
@@ -10,6 +11,30 @@ from manta_common.errors import MantaDockerError
 from manta_common.traces import Tracer
 
 
+class ImageFreshness(Enum):
+    """How the local copy of an image compares with the registry's.
+
+    THREE outcomes, and ``UNKNOWN`` must never be collapsed into ``CURRENT``.
+    A caller that cannot tell *these are the registry's bytes* from *nobody
+    could ask* runs stale bytes while reporting an ordinary start -- silent in
+    both directions, which is the defect in #457.
+
+    What separates ``UNKNOWN`` from ``CURRENT`` is WHY no comparison happened,
+    never that none happened. An image with no ``RepoDigests`` was built
+    locally and never pushed: there is nothing it could be stale against, so it
+    is ``CURRENT``. That is not a degraded reading and must not become one --
+    ``install_manta_platform.sh`` builds ``manta-light:pytorch`` under exactly
+    that unprefixed name, and only the name distinguishes it from the registry
+    image, so reporting it ``UNKNOWN`` would put a permanent warning on a
+    correct state and spend the third outcome's credibility on the case where
+    it means nothing.
+    """
+
+    CURRENT = "current"
+    STALE = "stale"
+    UNKNOWN = "unknown"
+
+
 class DockerRaiser:
     async_client: Docker
     tracer: Tracer
@@ -21,9 +46,9 @@ class DockerRaiser:
         max_workers=4, thread_name_prefix="docker-pull"
     )
 
-    async def image_needs_pull(self, image: str) -> bool:
+    async def image_freshness(self, image: str) -> ImageFreshness:
         """
-        Check whether ``image`` is missing locally or stale relative to the registry.
+        Compare the local copy of ``image`` with the registry's.
 
         Compares the local image's ``RepoDigests`` against the remote manifest
         digest, obtained via the Docker daemon's own distribution API
@@ -39,21 +64,23 @@ class DockerRaiser:
 
         Returns
         -------
-        bool
-            True if the image should be (re)pulled.
+        ImageFreshness
+            ``STALE`` if the image must be (re)pulled, ``CURRENT`` if the local
+            bytes are the right ones, ``UNKNOWN`` if that question could not be
+            answered.
 
         Notes
         -----
-        Graceful degradation: if the registry check itself fails (unreachable,
-        unauthenticated, rate-limited, ...), a warning is logged and this falls
-        back to "pull only if absent locally" — it never raises out of task
-        launch.
+        Never raises out of task launch. A registry that cannot be reached
+        (unreachable, unauthenticated, rate-limited, ...) yields ``UNKNOWN``
+        and a warning, NOT ``CURRENT``: the caller decides what to do about an
+        unanswered question, and it must be able to see that it was unanswered.
         """
         try:
             local = await self.async_client.images.inspect(image)
         except Exception:
             # Image not present locally: must pull regardless of the registry.
-            return True
+            return ImageFreshness.STALE
 
         local_digests = {
             digest.split("@", 1)[1]
@@ -63,9 +90,9 @@ class DockerRaiser:
         if not local_digests:
             # Locally-built/tagged image (never pushed/pulled), or only
             # malformed RepoDigests entries: nothing to compare against the
-            # registry. Treat as up-to-date rather than overwriting a
-            # local-only image on every launch.
-            return False
+            # registry. CURRENT, not UNKNOWN — there is no question left open
+            # here, so the registry is never consulted and no warning is due.
+            return ImageFreshness.CURRENT
 
         try:
             # Default executor, NOT _PULL_EXECUTOR: this quick probe must not
@@ -75,13 +102,74 @@ class DockerRaiser:
             )
             remote_digest = distribution["Descriptor"]["digest"]
         except Exception as exc:
+            # Wording pinned by test_image_freshness_keeps_the_operator_grep_target.
             self.tracer.warning(
                 f"Cannot verify registry digest for {image}, skipping "
                 f"auto-pull staleness check: {exc!r}"
             )
-            return False
+            return ImageFreshness.UNKNOWN
 
-        return remote_digest not in local_digests
+        if remote_digest in local_digests:
+            return ImageFreshness.CURRENT
+        return ImageFreshness.STALE
+
+    async def pull_if_stale(
+        self, image: str, force_pull: bool = False
+    ) -> ImageFreshness:
+        """
+        The single site that decides whether a container start pulls first.
+
+        BOTH container-start paths call this — ``ContainerManager.run_container``
+        (zipapp) and ``ImageExecutor.run_image`` (native entrypoint). That is the
+        invariant, and it is the whole point: a start path that decides for
+        itself can silently stop deciding, because nothing about one path's
+        behaviour is asserted by the other's tests (#457, #306).
+
+        ``STALE`` and ``UNKNOWN`` are deliberately NOT treated alike. Known-wrong
+        bytes are refused loudly: the pull runs and ``pull_image_or_raise``
+        propagates its failure, so the launch dies rather than running an image
+        we have established is out of date. An unanswered question proceeds on
+        the local copy and says so, because refusing there would turn any
+        registry outage into a fleet-wide stoppage — and ``image_freshness``
+        commits to never raising out of task launch.
+
+        Parameters
+        ----------
+        image : str
+            Image reference the container will run
+        force_pull : bool
+            Pull unconditionally, skipping the comparison entirely — including
+            the registry round trip that comparison costs
+
+        Returns
+        -------
+        ImageFreshness
+            What is known about the bytes the container will run: ``CURRENT``
+            when they are the registry's (compared equal, just pulled, or a
+            local-only image with nothing to compare), ``UNKNOWN`` when the
+            registry could not be asked. ``STALE`` is never returned — a stale
+            image is pulled, and a pull that fails raises instead of returning.
+        """
+        if force_pull:
+            # Ahead of the comparison, not merged into it: the comparison is a
+            # registry round trip, and evaluating it here would make the
+            # documented bypass the slowest path in this function.
+            self.tracer.info(f"Pulling image {image}...")
+            await self.pull_image_or_raise(image)
+            return ImageFreshness.CURRENT
+
+        freshness = await self.image_freshness(image)
+        if freshness is ImageFreshness.STALE:
+            self.tracer.info(f"Pulling image {image}...")
+            await self.pull_image_or_raise(image)
+            return ImageFreshness.CURRENT
+        if freshness is ImageFreshness.UNKNOWN:
+            self.tracer.warning(
+                f"Starting {image} from the local copy with its freshness "
+                f"unverified: the registry digest could not be read, so these "
+                f"bytes are unconfirmed rather than known-current."
+            )
+        return freshness
 
     async def run_or_raise(self, config: dict) -> DockerContainer:
         """
@@ -147,7 +235,7 @@ class DockerRaiser:
         Pulls via the sync docker-py client rather than the async aiodocker
         client. docker-py auto-resolves registry credentials from
         ``~/.docker/config.json`` (including credHelpers/credsStore) — the
-        same client ``image_needs_pull`` already uses for
+        same client ``image_freshness`` already uses for
         ``inspect_distribution`` — whereas aiodocker pulls anonymously and
         404s on private images with "pull access denied ... may require
         'docker login'" (refs #348).

manta_node/infrastructure/container/image_executor.py

@@ -166,9 +166,7 @@ class ImageExecutor(DockerRaiser):
                 metadata=self.tracer.metadata,
             )
 
-        if force_pull or await self.image_needs_pull(image):
-            self.tracer.info(f"Pulling image {image}...")
-            await self.pull_image_or_raise(image)
+        await self.pull_if_stale(image, force_pull=force_pull)
 
         # Copy (not alias) gpu_config's HostConfig — it is shared, long-lived
         # state on self; mutating it in place would leak NetworkMode/ExtraHosts

manta_node/infrastructure/container/manager.py

@@ -239,9 +239,8 @@ class ContainerManager(DockerRaiser):
         for key, value in config.items():
             self.tracer.debug(f"      {key}: {value}")
 
-        if force_pull:
-            self.tracer.info(f"Pulling image {image}...")
-            await self.pull_image_or_raise(image)
+        # Must resolve the image before the container is created, never after.
+        await self.pull_if_stale(image, force_pull=force_pull)
 
         self.containers[(swarm_id, task_id)] = await self.run_or_raise(config)
 

{manta_node-0.6b0.dev607.dist-info → manta_node-0.6b0.dev615.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: manta_node
-Version: 0.6b0.dev607
+Version: 0.6b0.dev615
 Summary: Node side software of the Manta platform. It is responsible for the communication with the Manta Core and the execution of the tasks.
 Author-email: Hugo Miralles <hugo.miralles@manta-tech.io>, Matthew Thompson <matthew.thompson@manta-tech.io>
 License-Expression: AGPL-3.0-or-later

{manta_node-0.6b0.dev607.dist-info → manta_node-0.6b0.dev615.dist-info}/RECORD

@@ -24,11 +24,11 @@ manta_node/infrastructure/__init__.py,sha256=yM60jKhdaD18BOsLb-0aR9D2XxGvrTnV92o
 manta_node/infrastructure/executor.py,sha256=wbH5-G4Y3iNTPr8ROWRDT-bWD-CYl1YFlZQ8dd9tpW4,2636
 manta_node/infrastructure/config/__init__.py,sha256=VnKidPAVfzlZMb5WM-MggtsDLM8tdAFFWlB7WSiawzc,617
 manta_node/infrastructure/config/node_config_manager.py,sha256=Js6PISjR_JRXMhgFC7lzI6StjW5paFhdhmtteao9j5A,35228
-manta_node/infrastructure/container/__init__.py,sha256=EO4SndUwmXrHczM4qkleUbxC7lTGz3c3Hd3iGrPg6D8,262
-manta_node/infrastructure/container/docker_adapter.py,sha256=DCCcKN7ahQK7rEX1gjzRH3lR2mi9YeEi7dFfNvSaF28,11990
+manta_node/infrastructure/container/__init__.py,sha256=ACDyzGvtM7nfTsBXDAYFjkKgU2yu4Umj3CQ_rlRfw1I,300
+manta_node/infrastructure/container/docker_adapter.py,sha256=Uxd3cNZMPyu7p68fq5QWvkF_vju2SjEBCqsvcnWYI-8,16368
 manta_node/infrastructure/container/executor.py,sha256=0aTFXDv-oXz81qwGi6zl_4qtkFriPWzn7iiiTRg7KJM,7274
-manta_node/infrastructure/container/image_executor.py,sha256=TnukiJWEtEcutqo45rYWkLH4fPXBzPnHMxcwCP7TOBs,10421
-manta_node/infrastructure/container/manager.py,sha256=VA-j6jO74Y53uu7Gvy8ibWSI-8GclwZ3dRCKTEA-v7A,15779
+manta_node/infrastructure/container/image_executor.py,sha256=noNwde03hw9k3EnK1JuQGtByUf-XDHqx623X0DKgHZ0,10315
+manta_node/infrastructure/container/manager.py,sha256=kylDl-5VZk01tzSXVrWI9VPn8mqi6tV-7Rf1tkMgee0,15790
 manta_node/infrastructure/filesystem/__init__.py,sha256=nf6Fakp3wuFKjSAm_oihI6WE_yyUY6Uelr7A1GwSvXs,145
 manta_node/infrastructure/filesystem/dataset_manager.py,sha256=sW7vPKYBLjwbtju1yzFEob9GlY1v0VI-xu6gxF9KM90,15075
 manta_node/infrastructure/grpc/__init__.py,sha256=6I-yzQM42gdBfcZ1FPVEKF2-pgJMdW-UssDHgz2EFcc,254
@@ -50,9 +50,9 @@ manta_node/infrastructure/process/executor.py,sha256=Ab1XvNuelEGCD5KdZLgYYnpjYZ5
 manta_node/infrastructure/security/__init__.py,sha256=lI6vkbUmfmt0m363iO4HzuzVNoQAc7Clp2R5vAhZbgE,64
 manta_node/infrastructure/security/auth_agent.py,sha256=WkPTg_g0gix7xomb_on_Ft3TNX0FpeqS7gwgJW9mTvk,11037
 manta_node/infrastructure/security/token_provider.py,sha256=u3bcbQYNBskFzunO-wOSZpGkDN_CU1BCB5HUbZTC4Vg,1832
-manta_node-0.6b0.dev607.dist-info/licenses/LICENSE,sha256=9mBgLhiB9goQPQjg9XQJ6FjgVxYw9Q8lyOZYkPGwF_c,34484
-manta_node-0.6b0.dev607.dist-info/METADATA,sha256=z0twRxUnPrpGmoilw-uIUvYmgNc3UxRKAxZ33Dxq6TA,4213
-manta_node-0.6b0.dev607.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
-manta_node-0.6b0.dev607.dist-info/entry_points.txt,sha256=C_zcw6mdmztNxVUZGLZ83UH5TwcSuqb2C-rB4e2CQoc,56
-manta_node-0.6b0.dev607.dist-info/top_level.txt,sha256=nbT5J5WETuXM9UPoQckl29uUOq8wMYdRb9QDx5kAFeo,11
-manta_node-0.6b0.dev607.dist-info/RECORD,,
+manta_node-0.6b0.dev615.dist-info/licenses/LICENSE,sha256=9mBgLhiB9goQPQjg9XQJ6FjgVxYw9Q8lyOZYkPGwF_c,34484
+manta_node-0.6b0.dev615.dist-info/METADATA,sha256=5GbkeVrFKSatf7YuxFnn-QqLT2KEE3htAxlFv5jx0Hc,4213
+manta_node-0.6b0.dev615.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
+manta_node-0.6b0.dev615.dist-info/entry_points.txt,sha256=C_zcw6mdmztNxVUZGLZ83UH5TwcSuqb2C-rB4e2CQoc,56
+manta_node-0.6b0.dev615.dist-info/top_level.txt,sha256=nbT5J5WETuXM9UPoQckl29uUOq8wMYdRb9QDx5kAFeo,11
+manta_node-0.6b0.dev615.dist-info/RECORD,,