aipmodel 0.2.71__tar.gz → 0.2.73__tar.gz

{aipmodel-0.2.71 → aipmodel-0.2.73}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aipmodel
-Version: 0.2.71
+Version: 0.2.73
 Summary: SDK for model registration, versioning, and storage
 Home-page: https://github.com/AIP-MLOPS/model-registry
 Author: AIP MLOPS Team

{aipmodel-0.2.71 → aipmodel-0.2.73}/README.md

@@ -120,6 +120,71 @@ reset). The SDK now **retries these automatically**, so `add_model`,
   worker) is fully protected; for multiple workers/replicas add a distributed
   lock (Mlops-Core/Redis).
 
+## Registry ↔ Storage Consistency (Integrity, Repair & Disaster Recovery)
+
+A model lives in **two systems at once**: its metadata in the registry
+(Mlops-Core) and its files in S3/Ceph. If either side changes behind the SDK's
+back (someone deletes a folder in Ceph by hand, or the registry loses data),
+you get a **zombie** (registry entry pointing at missing files) or an
+**orphan** (files no registry entry references). The SDK detects, repairs, and
+recovers all of these:
+
+### Recovery sidecar: `_meta.json`
+
+Every `add_model` writes a small `_meta.json` next to each version's files
+(model name, version, number, category, tags, owner, size, created, public).
+Storage is therefore **self-describing** — a registry entry can be rebuilt from
+S3 alone.
+
+### Detect
+
+```python
+report = manager.verify_model_integrity(model_name="my_model")
+# {"status": "ok" | "missing_files" | "orphan_files",
+#  "versions": [{"version": "v1.0", "files_exist": True, ...}],
+#  "orphan_folders": [...]}
+
+manager.verify_all_models()   # one report per model in your project
+manager.find_orphan_models()  # S3 model folders unknown to the registry
+                              # (each flagged recoverable=True when _meta.json exists)
+```
+
+### Repair (each guarded against misuse)
+
+```python
+# Zombie version (files deleted in Ceph) -> remove it from the registry.
+# REFUSES if the files actually still exist (use delete_model for a real delete).
+manager.remove_missing_version("my_model", "v1.0")
+
+# Stray folders under a model that no version references -> delete them.
+manager.purge_orphan_folders(model_name="my_model")
+
+# S3 model folder the registry doesn't know -> delete it.
+# REFUSES if the id IS registered.
+manager.purge_orphan_model(model_id)
+```
+
+### Disaster recovery (registry data lost, S3 survives)
+
+```python
+for orphan in manager.find_orphan_models():
+    if orphan["recoverable"]:
+        manager.recover_model_from_storage(orphan["model_id"])
+# Rebuilds the registry entry (name, all versions, latest, sizes, category,
+# tags) from the _meta.json sidecars. Files stay where they are.
+```
+
+Consistency is also protected *proactively*: `add_model` rolls back files,
+registry entry and ACL record together on failure, and `delete_model` updates
+the registry **before** deleting S3 (a crash can leave a harmless orphan, never
+a zombie). Versions added before the sidecar feature (< 0.2.69) have no
+`_meta.json` and are reported as non-recoverable.
+
+> **S3-compat note:** the client is configured to stay compatible with
+> MinIO/Ceph RGW across boto3 versions (legacy `Content-MD5` checksums on
+> newer boto3, with a per-object delete fallback if a batch delete is
+> rejected) — no user action needed.
+
 ## Access Control (ACL)
 
 ACL is **off by default** and fully optional. Toggle it with one env var:
@@ -314,7 +379,32 @@ manager.delete_model(model_name="demo_model_hf")
 | **`set_latest_version`** | `model_name`, `version` | Updates the 'latest' pointer to a specific version without modifying files. |
 | **`set_model_public`** | `model_name`, `public` | Moves a model's files between the private and public buckets (transactional, parallel server-side copy). |
 | **`share_model`** | `model_name`, `target_user`, `access_level` | Grants another user access to a model. Requires `ACL_ENABLED=true`. |
-| **`delete_model`** | `model_name`, `version` | If `version` provided: deletes that version (and rolls "latest" back to the next one if it was the latest). If that was the *last* version, the model container is removed too. If `version` is omitted: deletes the entire model in one call — this already is "delete all versions", no need to delete one by one. |
+| **`delete_model`** | `model_name`, `version` | If `version` provided: deletes that version (and rolls "latest" back to the next one if it was the latest). If that was the *last* version, the model container is removed too. If `version` is omitted: deletes the entire model in one call — this already is "delete all versions", no need to delete one by one. Registry metadata is updated **before** S3 files are removed (crash-safe ordering). |
+| **`verify_model_integrity`** | `model_name` or `model_id` | Cross-checks one model between registry and S3. Returns `{status: ok\|missing_files\|orphan_files, versions: [...], orphan_folders: [...]}`. |
+| **`verify_all_models`** | — | Runs `verify_model_integrity` over every model in your project; returns a list of reports. |
+| **`find_orphan_models`** | — | Lists S3 model folders the registry doesn't know about, each flagged `recoverable` when a `_meta.json` sidecar exists. |
+| **`remove_missing_version`** | `model_name`, `version` | Sync-repair for a zombie version whose S3 files were deleted: removes it from the registry. Refuses if the files still exist. |
+| **`purge_orphan_folders`** | `model_name` or `model_id` | Deletes S3 folders under a model that no registry version references. Returns the purged folder names. |
+| **`purge_orphan_model`** | `model_id` | Deletes the S3 storage of a model unknown to the registry. Refuses for registered ids. |
+| **`recover_model_from_storage`** | `model_id` | Disaster recovery: rebuilds the registry entry (all versions, latest, category, tags) from the `_meta.json` sidecars in S3. Returns the new registry id. |
+
+### Backward compatibility
+
+The ClearML → Mlops-Core rename kept the old public names working, so code
+written against the pre-0.2.65 SDK does not break:
+
+- **Health-check methods** — `check_clearml_service()` and `check_clearml_auth()`
+  still exist as thin aliases of `check_mlops_core_service()` /
+  `check_mlops_core_auth()`.
+- **Instance attributes** — `mgr.CLEARML_API_HOST`, `CLEARML_WEB_HOST`,
+  `CLEARML_USERNAME`, `CLEARML_ACCESS_KEY`, `CLEARML_SECRET_KEY` are kept as
+  aliases of the canonical `MLOPS_CORE_*` attributes.
+- **Constructor / env** — the `CLEARML_API_HOST` / `CLEARML_WEB_HOST` kwargs and
+  env vars are accepted as deprecated fallbacks (the `MLOPS_CORE_*` names win if
+  both are set).
+- **Method signatures are additive only** — every new parameter (`owner`,
+  `model_folder_name`, the `list_models` filters) is optional with a default
+  that reproduces the old behavior; nothing was removed or reordered.
 
 ---
 

{aipmodel-0.2.71 → aipmodel-0.2.73}/aipmodel/__init__.py

@@ -11,5 +11,5 @@ if _os.environ.get("AIPMODEL_UPDATE_CHECK", "").lower() in ("1", "true", "yes"):
 
 __all__ = ["MLOpsManager", "CephS3Manager"]
 
-__version__ = "0.2.71"
+__version__ = "0.2.73"
 __description__ = "SDK for model registration, versioning, and storage"

{aipmodel-0.2.71 → aipmodel-0.2.73}/aipmodel/model_registry.py

@@ -434,7 +434,17 @@ class MLOpsManager:
         # from the user-management wire format — do not rename them.
         self.MLOPS_CORE_ACCESS_KEY = user_info["clearml_access_key"]
         self.MLOPS_CORE_SECRET_KEY = user_info["clearml_secret_key"]
-        
+
+        # Backward-compatibility attribute aliases (pre-0.2.65 names). Callers
+        # that read mgr.CLEARML_API_HOST / CLEARML_WEB_HOST / CLEARML_USERNAME /
+        # CLEARML_ACCESS_KEY / CLEARML_SECRET_KEY keep working after the
+        # Mlops-Core rename — the new MLOPS_CORE_* names are canonical.
+        self.CLEARML_API_HOST = self.MLOPS_CORE_API_HOST
+        self.CLEARML_WEB_HOST = self.MLOPS_CORE_WEB_HOST
+        self.CLEARML_USERNAME = self.MLOPS_CORE_USERNAME
+        self.CLEARML_ACCESS_KEY = self.MLOPS_CORE_ACCESS_KEY
+        self.CLEARML_SECRET_KEY = self.MLOPS_CORE_SECRET_KEY
+
         if not skip_connection_check:
             if self.verbose:
                 print("[SDK_INFO] Performing health checks...")
@@ -608,6 +618,16 @@ class MLOpsManager:
                 print(f"[SDK_FAIL] Mlops-Core Auth Check Failed: {e}")
             return False
 
+    # Backward-compatibility aliases. These methods were renamed
+    # check_clearml_* -> check_mlops_core_* as part of the Mlops-Core rename;
+    # the old names are kept as thin forwarders so pre-0.2.65 callers that do
+    # mgr.check_clearml_service() / mgr.check_clearml_auth() keep working.
+    def check_clearml_service(self):
+        return self.check_mlops_core_service()
+
+    def check_clearml_auth(self):
+        return self.check_mlops_core_auth()
+
     def get_model_id_by_name(self, name, owner=None):
         """
         Resolves a model name to its Mlops-Core ID.
@@ -1657,7 +1677,8 @@ class MLOpsManager:
 
         Returns:
             {
-              "model_id", "model_name", "status": "ok"|"missing_files"|"orphan_files",
+              "model_id", "model_name",
+              "status": "ok"|"missing_files"|"orphan_files"|"empty",
               "versions": [{"version","path","files_exist","status"}...],
               "orphan_folders": [<S3 folders under the model not referenced by any version>],
             }
@@ -1713,6 +1734,15 @@ class MLOpsManager:
             report["status"] = "missing_files"
         elif report["orphan_folders"]:
             report["status"] = "orphan_files"
+        elif not versions_map:
+            # Registry holds the model container but NO versions and NO files.
+            # This is the shell an interrupted add leaves behind when the process
+            # is hard-killed mid-upload (pod rollout, OOM, node eviction) — the
+            # in-process rollback in add_model never runs in that case. Surface it
+            # (it was previously reported "ok") so it can be cleaned up or the add
+            # retried. Note: a version add still in flight shows its temp folder as
+            # an orphan_folder above, so it reports "orphan_files", not "empty".
+            report["status"] = "empty"
         return report
 
     def verify_all_models(self):
@@ -1793,6 +1823,23 @@ class MLOpsManager:
         self.delete_model(model_name=model_name, version=version)
         print(f"[SDK_SUCCESS] Synced: removed zombie version '{version}' of '{model_name}' from registry.")
 
+    def remove_empty_model(self, model_name=None, model_id=None):
+        """
+        Sync-repair for an EMPTY model container: the registry holds the model
+        but it has no versions and no stored files — the shell an interrupted /
+        hard-killed add leaves behind. Deletes the container so the registry is
+        clean. Refuses to run if the model actually has versions or files (so a
+        real model, or an add still in flight, is never removed by mistake).
+        """
+        report = self.verify_model_integrity(model_name=model_name, model_id=model_id)
+        if report["status"] != "empty":
+            raise ValueError(
+                f"Model '{report['model_name']}' is not empty (status={report['status']}); "
+                f"refusing to remove it. Use delete_model for a real delete."
+            )
+        self.delete_model(model_id=report["model_id"])
+        print(f"[SDK_SUCCESS] Removed empty model shell '{report['model_name']}'.")
+
     def purge_orphan_folders(self, model_name=None, model_id=None):
         """
         Sync-repair for ORPHAN folders inside a registered model: S3 folders

{aipmodel-0.2.71 → aipmodel-0.2.73}/aipmodel.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aipmodel
-Version: 0.2.71
+Version: 0.2.73
 Summary: SDK for model registration, versioning, and storage
 Home-page: https://github.com/AIP-MLOPS/model-registry
 Author: AIP MLOPS Team

{aipmodel-0.2.71 → aipmodel-0.2.73}/setup.py

@@ -31,7 +31,7 @@ version, description = read_meta()
 
 setup(
     name="aipmodel",
-    version="0.2.71",
+    version="0.2.73",
     description=description,
     author="AIP MLOPS Team",
     author_email="mohmmadweb@gmail.com",