ob-metaflow-extensions 1.1.146__tar.gz → 1.1.148__tar.gz

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ob-metaflow-extensions
-Version: 1.1.146
+Version: 1.1.148
 Summary: Outerbounds Platform Extensions for Metaflow
 Author: Outerbounds, Inc.
 License: Commercial

ob-metaflow-extensions-1.1.148/metaflow_extensions/outerbounds/plugins/checkpoint_datastores/__init__.py

@@ -0,0 +1,2 @@
+from .nebius import nebius_checkpoints
+from .coreweave import coreweave_checkpoints

ob-metaflow-extensions-1.1.148/metaflow_extensions/outerbounds/plugins/checkpoint_datastores/coreweave.py

@@ -0,0 +1,139 @@
+from metaflow.user_configs.config_decorators import (
+    MutableFlow,
+    MutableStep,
+    CustomFlowDecorator,
+)
+import os
+
+
+class coreweave_checkpoints(CustomFlowDecorator):
+
+    """
+
+    This decorator is used for setting the coreweave object store as the artifact store for checkpoints/models created by the flow.
+
+    Parameters
+    ----------
+    secrets: list
+        A list of secrets to be added to the step. These secrets should contain any secrets that are required globally and the secret
+        for the coreweave object store. The secret should contain the following keys:
+        - COREWEAVE_ACCESS_KEY
+        - COREWEAVE_SECRET_KEY
+
+    bucket_path: str
+        The path to the bucket to store the checkpoints/models.
+
+    Usage
+    -----
+    ```python
+    from metaflow import checkpoint, step, FlowSpec, coreweave_checkpoints
+
+    @coreweave_checkpoints(secrets=[], bucket_path=None)
+    class MyFlow(FlowSpec):
+        @checkpoint
+        @step
+        def start(self):
+            # Saves the checkpoint in the coreweave object store
+            current.checkpoint.save("./foo.txt")
+
+        @step
+        def end(self):
+            pass
+    ```
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def init(self, *args, **kwargs):
+        self.bucket_path = kwargs.get("bucket_path", None)
+
+        self.secrets = kwargs.get("secrets", [])
+        if self.bucket_path is None:
+            raise ValueError(
+                "`bucket_path` keyword argument is required for the coreweave_datastore"
+            )
+        if not self.bucket_path.startswith("s3://"):
+            raise ValueError(
+                "`bucket_path` must start with `s3://` for the coreweave_datastore"
+            )
+
+        self.coreweave_endpoint_url = f"https://cwobject.com"
+        if self.secrets is None:
+            raise ValueError(
+                "`secrets` keyword argument is required for the coreweave_datastore"
+            )
+
+    def evaluate(self, mutable_flow: MutableFlow) -> None:
+        from metaflow import (
+            checkpoint,
+            model,
+            huggingface_hub,
+            secrets,
+            with_artifact_store,
+        )
+
+        def _add_secrets(step: MutableStep) -> None:
+            decos_to_add = []
+            swapping_decos = {
+                "huggingface_hub": huggingface_hub,
+                "model": model,
+                "checkpoint": checkpoint,
+            }
+            already_has_secrets = False
+            secrets_present_in_deco = []
+            for d in step.decorators:
+                if d.name in swapping_decos:
+                    decos_to_add.append((d.name, d.attributes))
+                elif d.name == "secrets":
+                    already_has_secrets = True
+                    secrets_present_in_deco.extend(d.attributes["sources"])
+
+            # If the step aleady has secrets then take all the sources in
+            # the secrets and add the addtional secrets to the existing secrets
+            secrets_to_add = self.secrets
+            if already_has_secrets:
+                secrets_to_add.extend(secrets_present_in_deco)
+
+            secrets_to_add = list(set(secrets_to_add))
+
+            if len(decos_to_add) == 0:
+                if already_has_secrets:
+                    step.remove_decorator("secrets")
+
+                step.add_decorator(
+                    secrets,
+                    sources=secrets_to_add,
+                )
+                return
+
+            for d, _ in decos_to_add:
+                step.remove_decorator(d)
+
+            step.add_decorator(
+                secrets,
+                sources=secrets_to_add,
+            )
+            for d, attrs in decos_to_add:
+                _deco_to_add = swapping_decos[d]
+                step.add_decorator(_deco_to_add, **attrs)
+
+        def _coreweave_config():
+            return {
+                "root": self.bucket_path,
+                "client_params": {
+                    "aws_access_key_id": os.environ.get("COREWEAVE_ACCESS_KEY"),
+                    "aws_secret_access_key": os.environ.get("COREWEAVE_SECRET_KEY"),
+                    "endpoint_url": self.coreweave_endpoint_url,
+                    "config": dict(s3={"addressing_style": "virtual"}),
+                },
+            }
+
+        mutable_flow.add_decorator(
+            with_artifact_store,
+            type="coreweave",
+            config=_coreweave_config,
+        )
+
+        for step_name, step in mutable_flow.steps:
+            _add_secrets(step)

ob-metaflow-extensions-1.1.148/metaflow_extensions/outerbounds/plugins/checkpoint_datastores/nebius.py

@@ -0,0 +1,144 @@
+from metaflow.user_configs.config_decorators import (
+    MutableFlow,
+    MutableStep,
+    CustomFlowDecorator,
+)
+import os
+
+NEBIUS_ENDPOINT_URL = "https://storage.eu-north1.nebius.cloud:443"
+
+
+class nebius_checkpoints(CustomFlowDecorator):
+
+    """
+
+    This decorator is used for setting the nebius's S3 compatible object store as the artifact store for
+    checkpoints/models created by the flow.
+
+    Parameters
+    ----------
+    secrets: list
+        A list of secrets to be added to the step. These secrets should contain any secrets that are required globally and the secret
+        for the nebius object store. The secret should contain the following keys:
+        - NEBIUS_ACCESS_KEY
+        - NEBIUS_SECRET_KEY
+
+    bucket_path: str
+        The path to the bucket to store the checkpoints/models.
+
+    endpoint_url: str
+        The endpoint url for the nebius object store. Defaults to `https://storage.eu-north1.nebius.cloud:443`
+
+    Usage
+    -----
+    ```python
+    from metaflow import checkpoint, step, FlowSpec, nebius_checkpoints
+
+    @nebius_checkpoints(secrets=[], bucket_path=None)
+    class MyFlow(FlowSpec):
+        @checkpoint
+        @step
+        def start(self):
+            # Saves the checkpoint in the nebius object store
+            current.checkpoint.save("./foo.txt")
+
+        @step
+        def end(self):
+            pass
+    ```
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def init(self, *args, **kwargs):
+        self.bucket_path = kwargs.get("bucket_path", None)
+
+        self.secrets = kwargs.get("secrets", [])
+        if self.bucket_path is None:
+            raise ValueError(
+                "`bucket_path` keyword argument is required for the coreweave_datastore"
+            )
+        if not self.bucket_path.startswith("s3://"):
+            raise ValueError(
+                "`bucket_path` must start with `s3://` for the coreweave_datastore"
+            )
+
+        self.nebius_endpoint_url = kwargs.get("endpoint_url", NEBIUS_ENDPOINT_URL)
+        if self.secrets is None:
+            raise ValueError(
+                "`secrets` keyword argument is required for the coreweave_datastore"
+            )
+
+    def evaluate(self, mutable_flow: MutableFlow) -> None:
+        from metaflow import (
+            checkpoint,
+            model,
+            huggingface_hub,
+            secrets,
+            with_artifact_store,
+        )
+
+        def _add_secrets(step: MutableStep) -> None:
+            decos_to_add = []
+            swapping_decos = {
+                "huggingface_hub": huggingface_hub,
+                "model": model,
+                "checkpoint": checkpoint,
+            }
+            already_has_secrets = False
+            secrets_present_in_deco = []
+            for d in step.decorators:
+                if d.name in swapping_decos:
+                    decos_to_add.append((d.name, d.attributes))
+                elif d.name == "secrets":
+                    already_has_secrets = True
+                    secrets_present_in_deco.extend(d.attributes["sources"])
+
+            # If the step aleady has secrets then take all the sources in
+            # the secrets and add the addtional secrets to the existing secrets
+            secrets_to_add = self.secrets
+            if already_has_secrets:
+                secrets_to_add.extend(secrets_present_in_deco)
+
+            secrets_to_add = list(set(secrets_to_add))
+
+            if len(decos_to_add) == 0:
+                if already_has_secrets:
+                    step.remove_decorator("secrets")
+
+                step.add_decorator(
+                    secrets,
+                    sources=secrets_to_add,
+                )
+                return
+
+            for d, _ in decos_to_add:
+                step.remove_decorator(d)
+
+            step.add_decorator(
+                secrets,
+                sources=secrets_to_add,
+            )
+            for d, attrs in decos_to_add:
+                _deco_to_add = swapping_decos[d]
+                step.add_decorator(_deco_to_add, **attrs)
+
+        def _nebius_config():
+            return {
+                "root": self.bucket_path,
+                "client_params": {
+                    "aws_access_key_id": os.environ.get("NEBIUS_ACCESS_KEY"),
+                    "aws_secret_access_key": os.environ.get("NEBIUS_SECRET_KEY"),
+                    "endpoint_url": self.nebius_endpoint_url,
+                },
+            }
+
+        mutable_flow.add_decorator(
+            with_artifact_store,
+            type="s3",
+            config=_nebius_config,
+        )
+
+        for step_name, step in mutable_flow.steps:
+            _add_secrets(step)

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/metaflow_extensions/outerbounds/plugins/nvcf/nvcf_decorator.py

@@ -198,7 +198,12 @@ class NvcfDecorator(StepDecorator):
             meta["nvcf-nspectid"] = os.environ.get("NVCF_NSPECTID")
 
             entries = [
-                MetaDatum(field=k, value=v, type=k, tags=[])
+                MetaDatum(
+                    field=k,
+                    value=v,
+                    type=k,
+                    tags=["attempt_id:{0}".format(retry_count)],
+                )
                 for k, v in meta.items()
                 if v is not None
             ]

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/metaflow_extensions/outerbounds/plugins/snowflake/snowflake.py

@@ -7,9 +7,10 @@ from metaflow.metaflow_config import SERVICE_URL
 from metaflow.metaflow_config_funcs import init_config
 from typing import Dict
 from os import environ
-
+import sys
 import json
 import requests
+import random
 import time
 
 
@@ -75,7 +76,9 @@ def get_snowflake_token(user: str = "", role: str = "", integration: str = "") -
     }
     json_payload = json.dumps(payload)
     headers = provisioner.get_service_auth_header()
-    response = requests.get(snowflake_token_url, data=json_payload, headers=headers)
+    response = _api_server_get(
+        snowflake_token_url, data=json_payload, headers=headers, conn_error_retries=5
+    )
     response.raise_for_status()
     return response.json()["token"]
 
@@ -150,6 +153,39 @@ def connect(user: str = "", role: str = "", integration: str = "", **kwargs):
         )
 
 
+def _api_server_get(*args, conn_error_retries=2, **kwargs):
+    """
+    There are two categories of errors that we need to handle when dealing with any API server.
+    1. HTTP errors. These are are errors that are returned from the API server.
+        - How to handle retries for this case will be application specific.
+    2. Errors when the API server may not be reachable (DNS resolution / network issues)
+        - In this scenario, we know that something external to the API server is going wrong causing the issue.
+        - Failing pre-maturely in the case might not be the best course of action since critical user jobs might crash on intermittent issues.
+        - So in this case, we can just planely retry the request.
+
+    This function handles the second case. It's a simple wrapper to handle the retry logic for connection errors.
+    If this function is provided a `conn_error_retries` of 5, then the last retry will have waited 32 seconds.
+    Generally this is a safe enough number of retries after which we can assume that something is really broken. Until then,
+    there can be intermittent issues that would resolve themselves if we retry gracefully.
+    """
+    _num_retries = 0
+    noise = random.uniform(-0.5, 0.5)
+    while _num_retries < conn_error_retries:
+        try:
+            return requests.get(*args, **kwargs)
+        except requests.exceptions.ConnectionError:
+            if _num_retries <= conn_error_retries - 1:
+                # Exponential backoff with 2^(_num_retries+1) seconds
+                time.sleep((2 ** (_num_retries + 1)) + noise)
+                _num_retries += 1
+            else:
+                print(
+                    "[@snowflake] Failed to connect to the API server. ",
+                    file=sys.stderr,
+                )
+                raise
+
+
 class Snowflake:
     def __init__(
         self, user: str = "", role: str = "", integration: str = "", **kwargs
@@ -273,7 +309,9 @@ class SnowflakeIntegrationProvisioner:
         retryable_status_codes = [409]
         json_payload = json.dumps(payload)
         for attempt in range(2):  # 0 = initial attempt, 1-2 = retries
-            response = requests.get(url, data=json_payload, headers=request_headers)
+            response = _api_server_get(
+                url, data=json_payload, headers=request_headers, conn_error_retries=5
+            )
             if response.status_code not in retryable_status_codes:
                 break
 
@@ -281,7 +319,9 @@ class SnowflakeIntegrationProvisioner:
                 sleep_time = 0.5 * (attempt + 1)
                 time.sleep(sleep_time)
 
-        response = requests.get(url, data=json_payload, headers=request_headers)
+        response = _api_server_get(
+            url, data=json_payload, headers=request_headers, conn_error_retries=5
+        )
         self._handle_error_response(response)
         return response.json()
 

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/metaflow_extensions/outerbounds/toplevel/global_aliases_for_metaflow_package.py

@@ -52,3 +52,4 @@ def S3(*args, **kwargs):
 
 from .. import profilers
 from ..plugins.snowflake import Snowflake
+from ..plugins.checkpoint_datastores import nebius_checkpoints, coreweave_checkpoints

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/ob_metaflow_extensions.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ob-metaflow-extensions
-Version: 1.1.146
+Version: 1.1.148
 Summary: Outerbounds Platform Extensions for Metaflow
 Author: Outerbounds, Inc.
 License: Commercial

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/ob_metaflow_extensions.egg-info/SOURCES.txt

@@ -15,6 +15,9 @@ metaflow_extensions/outerbounds/plugins/card_utilities/__init__.py
 metaflow_extensions/outerbounds/plugins/card_utilities/async_cards.py
 metaflow_extensions/outerbounds/plugins/card_utilities/extra_components.py
 metaflow_extensions/outerbounds/plugins/card_utilities/injector.py
+metaflow_extensions/outerbounds/plugins/checkpoint_datastores/__init__.py
+metaflow_extensions/outerbounds/plugins/checkpoint_datastores/coreweave.py
+metaflow_extensions/outerbounds/plugins/checkpoint_datastores/nebius.py
 metaflow_extensions/outerbounds/plugins/fast_bakery/__init__.py
 metaflow_extensions/outerbounds/plugins/fast_bakery/docker_environment.py
 metaflow_extensions/outerbounds/plugins/fast_bakery/fast_bakery.py

ob-metaflow-extensions-1.1.148/ob_metaflow_extensions.egg-info/requires.txt

@@ -0,0 +1,3 @@
+boto3
+kubernetes
+ob-metaflow==2.15.10.1

{ob-metaflow-extensions-1.1.146 → ob-metaflow-extensions-1.1.148}/setup.py

@@ -2,7 +2,7 @@ from setuptools import setup, find_namespace_packages
 from pathlib import Path
 
 
-version = "1.1.146"
+version = "1.1.148"
 this_directory = Path(__file__).parent
 long_description = (this_directory / "README.md").read_text()
 
@@ -18,6 +18,6 @@ setup(
     install_requires=[
         "boto3",
         "kubernetes",
-        "ob-metaflow == 2.15.7.2",
+        "ob-metaflow == 2.15.10.1",
     ],
 )

ob-metaflow-extensions-1.1.146/ob_metaflow_extensions.egg-info/requires.txt

@@ -1,3 +0,0 @@
-boto3
-kubernetes
-ob-metaflow==2.15.7.2