schedulers/aws_batch: fix thread local sessions + raise error on missing memory resource

Author: d4l3k

torchx/components/serve.py

@@ -70,6 +70,7 @@ def torchserve(
                 entrypoint="python",
                 args=args,
                 port_map={"model-download": 8222},
+                resource=specs.Resource(cpu=1, gpu=0, memMB=1024),
             ),
         ],
     )

torchx/components/utils.py

@@ -39,6 +39,7 @@ def echo(
                 entrypoint="echo",
                 args=[msg],
                 num_replicas=num_replicas,
+                resource=specs.Resource(cpu=1, gpu=0, memMB=1024),
             )
         ],
     )
@@ -62,12 +63,21 @@ def touch(file: str, image: str = torchx.IMAGE) -> specs.AppDef:
                 entrypoint="touch",
                 args=[file],
                 num_replicas=1,
+                resource=specs.Resource(cpu=1, gpu=0, memMB=1024),
             )
         ],
     )
 
 
-def sh(*args: str, image: str = torchx.IMAGE, num_replicas: int = 1) -> specs.AppDef:
+def sh(
+    *args: str,
+    image: str = torchx.IMAGE,
+    num_replicas: int = 1,
+    cpu: int = 1,
+    gpu: int = 0,
+    memMB: int = 1024,
+    h: Optional[str] = None,
+) -> specs.AppDef:
     """
     Runs the provided command via sh. Currently sh does not support
     environment variable substitution.
@@ -76,7 +86,10 @@ def sh(*args: str, image: str = torchx.IMAGE, num_replicas: int = 1) -> specs.Ap
         args: bash arguments
         image: image to use
         num_replicas: number of replicas to run
-
+        cpu: number of cpus per replica
+        gpu: number of gpus per replica
+        memMB: cpu memory in MB per replica
+        h: a registered named resource (if specified takes precedence over cpu, gpu, memMB)
     """
 
     escaped_args = " ".join(shlex.quote(arg) for arg in args)
@@ -90,6 +103,7 @@ def sh(*args: str, image: str = torchx.IMAGE, num_replicas: int = 1) -> specs.Ap
                 entrypoint="sh",
                 args=["-c", escaped_args],
                 num_replicas=num_replicas,
+                resource=specs.resource(cpu=cpu, gpu=gpu, memMB=memMB, h=h),
             )
         ],
     )
@@ -102,7 +116,7 @@ def python(
     script: Optional[str] = None,
     image: str = torchx.IMAGE,
     name: str = "torchx_utils_python",
-    cpu: int = 2,
+    cpu: int = 1,
     gpu: int = 0,
     memMB: int = 1024,
     h: Optional[str] = None,
@@ -164,8 +178,12 @@ def python(
 def binary(
     *args: str,
     entrypoint: str,
-    name: str = "torchx_utils_python",
+    name: str = "torchx_utils_binary",
     num_replicas: int = 1,
+    cpu: int = 1,
+    gpu: int = 0,
+    memMB: int = 1024,
+    h: Optional[str] = None,
 ) -> specs.AppDef:
     """
     Test component
@@ -174,6 +192,10 @@ def binary(
         args: arguments passed to the program in sys.argv[1:] (ignored with `--c`)
         name: name of the job
         num_replicas: number of copies to run (each on its own container)
+        cpu: number of cpus per replica
+        gpu: number of gpus per replica
+        memMB: cpu memory in MB per replica
+        h: a registered named resource (if specified takes precedence over cpu, gpu, memMB)
     :return:
     """
     return specs.AppDef(
@@ -184,8 +206,8 @@ def binary(
                 image="<NONE>",
                 entrypoint=entrypoint,
                 num_replicas=num_replicas,
-                resource=specs.Resource(cpu=2, gpu=0, memMB=4096),
                 args=[*args],
+                resource=specs.resource(cpu=cpu, gpu=gpu, memMB=memMB, h=h),
             )
         ],
     )
@@ -219,6 +241,7 @@ def copy(src: str, dst: str, image: str = torchx.IMAGE) -> specs.AppDef:
                     "--dst",
                     dst,
                 ],
+                resource=specs.Resource(cpu=1, gpu=0, memMB=1024),
             ),
         ],
     )
@@ -261,6 +284,7 @@ def booth(
                     "--tracker_base",
                     tracker_base,
                 ],
+                resource=specs.Resource(cpu=1, gpu=0, memMB=1024),
             )
         ],
     )

torchx/schedulers/aws_batch_scheduler.py

@@ -38,7 +38,17 @@
 import threading
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Dict, Iterable, Mapping, Optional, Any, TYPE_CHECKING, Tuple
+from typing import (
+    Dict,
+    Iterable,
+    Mapping,
+    Optional,
+    Any,
+    TYPE_CHECKING,
+    Tuple,
+    TypeVar,
+    Callable,
+)
 
 import torchx
 import yaml
@@ -86,8 +96,10 @@ def _role_to_node_properties(idx: int, role: Role) -> Dict[str, object]:
     reqs.append({"type": "VCPU", "value": str(cpu)})
 
     memMB = resource.memMB
-    if memMB <= 0:
-        memMB = 1000
+    if memMB < 0:
+        raise ValueError(
+            f"AWSBatchScheduler requires memMB to be set to a positive value, got {memMB}"
+        )
     reqs.append({"type": "MEMORY", "value": str(memMB)})
 
     if resource.gpu > 0:
@@ -162,17 +174,29 @@ def __repr__(self) -> str:
         return str(self)
 
 
-def _thread_local_session() -> "boto3.session.Session":
-    KEY = "torchx_aws_batch_session"
-    local = threading.local()
-    if hasattr(local, KEY):
-        # pyre-ignore[16]
-        return getattr(local, KEY)
+T = TypeVar("T")
+
+
+def _thread_local_cache(f: Callable[[], T]) -> Callable[[], T]:
+    local: threading.local = threading.local()
+    key: str = "value"
+
+    def wrapper() -> T:
+        if key in local.__dict__:
+            return local.__dict__[key]
+
+        v = f()
+        local.__dict__[key] = v
+        return v
+
+    return wrapper
+
+
+@_thread_local_cache
+def _local_session() -> "boto3.session.Session":
     import boto3.session
 
-    session = boto3.session.Session()
-    setattr(local, KEY, session)
-    return session
+    return boto3.session.Session()
 
 
 class AWSBatchScheduler(Scheduler, DockerWorkspace):
@@ -244,14 +268,14 @@ def __init__(
     def _client(self) -> Any:
         if self.__client:
             return self.__client
-        return _thread_local_session().client("batch")
+        return _local_session().client("batch")
 
     @property
     # pyre-fixme[3]: Return annotation cannot be `Any`.
     def _log_client(self) -> Any:
         if self.__log_client:
             return self.__log_client
-        return _thread_local_session().client("logs")
+        return _local_session().client("logs")
 
     def schedule(self, dryrun_info: AppDryRunInfo[BatchJob]) -> str:
         cfg = dryrun_info._cfg

torchx/schedulers/test/aws_batch_scheduler_test.py

@@ -4,6 +4,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import threading
 import unittest
 from contextlib import contextmanager
 from typing import Generator
@@ -15,6 +16,7 @@
     create_scheduler,
     AWSBatchScheduler,
     _role_to_node_properties,
+    _local_session,
 )
 
 
@@ -198,6 +200,11 @@ def test_volume_mounts(self) -> None:
             mounts=[
                 specs.VolumeMount(src="efsid", dst_path="/dst", read_only=True),
             ],
+            resource=specs.Resource(
+                cpu=1,
+                memMB=1000,
+                gpu=0,
+            ),
         )
         props = _role_to_node_properties(0, role)
         self.assertEqual(
@@ -402,3 +409,16 @@ def test_log_iter(self) -> None:
                 "foobar",
             ],
         )
+
+    def test_local_session(self) -> None:
+        a: object = _local_session()
+        self.assertIs(a, _local_session())
+
+        def worker() -> None:
+            b = _local_session()
+            self.assertIs(b, _local_session())
+            self.assertIsNot(a, b)
+
+        t = threading.Thread(target=worker)
+        t.start()
+        t.join()