サンドボックスの設定

The examples on this page omit auth. To authenticate with your W&B API key, install cwsandbox[wandb] and pass auth=AuthStrategy.WANDB to Sandbox.run(), Sandbox(), or Sandbox.session(). See Get started.

This guide covers sandbox configuration options, including resources, mounted files, ports, annotations, and timeouts. Use it as a reference when you need to tailor a sandbox’s runtime environment to match the requirements of a workload, such as reserving GPU capacity, exposing a service, or setting operation timeouts. This page is for developers who use the cwsandbox Python SDK to launch and manage sandboxes.

Overview

You can set sandbox configuration in three places, listed from broadest to most specific scope:

SandboxDefaults: Shared defaults for all sandboxes in a session.
Sandbox.run() kwargs: Per-sandbox overrides.
@session.function() kwargs: Function-specific configuration.

from cwsandbox import ResourceOptions, Sandbox, SandboxDefaults, Session

# Via SandboxDefaults
defaults = SandboxDefaults(
    container_image="python:3.11",
    max_lifetime_seconds=3600,
    resources=ResourceOptions(
        requests={"cpu": "2", "memory": "2Gi"},
        limits={"cpu": "2", "memory": "2Gi"},
    ),
)

# Via Sandbox.run() kwargs
sandbox = Sandbox.run(
    defaults=defaults,
    resources=ResourceOptions(
        requests={"cpu": "4", "memory": "4Gi"},
        limits={"cpu": "4", "memory": "4Gi"},
    ),
)

# Via @session.function() kwargs
with Session(defaults) as session:
    @session.function(resources={"cpu": "1", "memory": "1Gi"})
    def compute(x: int) -> int:
        return x * 2

The following sections describe each configuration area in detail.

Resources

ResourceOptions データクラスまたはプレーンな dict を使用して、CPU、メモリ、GPU のリソースを設定します。サーバーレスの配置では、リクエストと制限に同じ値が使用されます。 The following examples use equal requests and limits:

from cwsandbox import ResourceOptions, Sandbox

# Using ResourceOptions
sandbox = Sandbox.run(
    resources=ResourceOptions(
        requests={"cpu": "2", "memory": "2Gi"},
        limits={"cpu": "2", "memory": "2Gi"},
    ),
)

# Using dict
sandbox = Sandbox.run(
    resources={
        "requests": {"cpu": "2", "memory": "2Gi"},
        "limits": {"cpu": "2", "memory": "2Gi"},
    },
)

Both forms are equivalent: the SDK automatically converts dicts to ResourceOptions internally. ResourceOptions separates resource requests from limits. Requests tell the scheduler what the sandbox needs. Limits set the maximum it can use. For background on how Kubernetes uses requests and limits to assign Quality of Service classes, see the Kubernetes documentation.

ResourceOptions fields

Field	Type	Description
`requests`	`dict[str, str] \| None`	CPU and memory requests for scheduling (for example, `{"cpu": "500m", "memory": "512Mi"}`).
`limits`	`dict[str, str] \| None`	CPU and memory limits the sandbox cannot exceed (for example, `{"cpu": "2", "memory": "2Gi"}`).
`gpu`	`dict[str, Any] \| None`	GPU configuration (for example, `{"count": 1, "type": "H100"}`).

All fields are optional in the client and default to None. The resolved allocation must satisfy the applicable policy. If you omit the resource allocation, complete, positive CPU and memory defaults from that policy supply both requests and limits.

Guaranteed QoS

When requests equal limits, the sandbox receives a Guaranteed Quality of Service class. The scheduler reserves the full allocation, and Kubernetes evicts Guaranteed Pods last under node resource pressure. A sandbox that tries to use more CPU than its limit is still throttled. Use this for latency-sensitive workloads.

sandbox = Sandbox.run(
    resources=ResourceOptions(
        requests={"cpu": "2", "memory": "4Gi"},
        limits={"cpu": "2", "memory": "4Gi"},
    ),
)

For Guaranteed QoS, a flat dict shorthand is available. The SDK normalizes flat dicts by setting both requests and limits to the same values.

# Flat dict shorthand: requests and limits are set to identical values
sandbox = Sandbox.run(
    resources={"cpu": "2", "memory": "4Gi"},
)

CPU values

Specify CPU in millicores or whole cores. For details, see Resource units in Kubernetes.

Value	Meaning
`"100m"`	100 millicores (0.1 CPU)
`"500m"`	500 millicores (0.5 CPU)
`"1000m"` or `"1"`	1 full CPU core
`"2000m"` or `"2"`	2 CPU cores

Memory values

Memory uses standard Kubernetes memory units:

Value	Meaning
`"128Mi"`	128 mebibytes
`"512Mi"`	512 mebibytes
`"1Gi"`	1 gibibyte
`"4Gi"`	4 gibibytes

GPU

Request GPU resources alongside CPU and memory. A GPU request reserves GPUs only, so set CPU and memory explicitly.

sandbox = Sandbox.run(
    resources=ResourceOptions(
        requests={"cpu": "4", "memory": "16Gi"},
        limits={"cpu": "4", "memory": "16Gi"},
        gpu={"count": 1},
    ),
)

GPU configuration keys:

キー	タイプ	説明
`count`	`int`	リクエストする GPU の数。
`type`	`str`	GPU タイプ。サーバーレスキャパシティでは GPU タイプが自動的に割り当てられるため、指定しないでください。詳しくは、GPU サンドボックスを実行するを参照してください。
`memory_gb`	`int`	GPU メモリの最小値 (GB)。ランナー上で、この値以上のメモリを搭載した GPU タイプが選択されます。条件を満たすタイプが複数ある場合、どのタイプが選択されるかは規定されていません。

type and memory_gb are alternatives. Set at most one of them.

Configuration library interop

All configuration types (NetworkOptions, Secret, ResourceOptions) accept either the dataclass or a plain dict. This means ML configuration libraries that resolve configs to dicts or dict-like objects can pass values directly to the SDK without manual conversion. SandboxDefaults.from_dict() accepts a plain dict or an OmegaConf DictConfig and coerces nested fields automatically. Dicts become NetworkOptions, Secret, or ResourceOptions as needed, and lists become tuples.

# sandbox.yaml
container_image: "pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime"
max_lifetime_seconds: 3600
tags:
  - training
  - experiment-42
resources:
  requests:
    cpu: "4"
    memory: "8Gi"
  limits:
    cpu: "4"
    memory: "8Gi"
  gpu:
    count: 1
network:
  egress_mode: "internet"
environment_variables:
  LOG_LEVEL: "info"

from omegaconf import OmegaConf
from cwsandbox import Sandbox, SandboxDefaults

cfg = OmegaConf.load("sandbox.yaml")
defaults = SandboxDefaults.from_dict(cfg)

with Sandbox.run(defaults=defaults) as sb:
    result = sb.exec(["python", "train.py"]).result()

Individual fields also accept dicts when you pass them directly to Sandbox.run() or session.sandbox(). The nested dict form for resources maps to ResourceOptions fields. The flat dict form ({"cpu": "1", "memory": "1Gi"}) is also accepted and treated as Guaranteed QoS.

Mounted files

Mounted files let you provide configuration files, scripts, or other read-only assets to the sandbox at startup, without baking them into a container image.

sandbox = Sandbox.run(
    mounted_files=[
        {
            "path": "/app/config.json",
            "content": '{"debug": true}',
        },
        {
            "path": "/app/script.py",
            "content": "print('hello')",
        },
    ],
)

# Files are available immediately
result = sandbox.exec(["python", "/app/script.py"]).result()

Mount options

Field	Type	Description
`path`	`str`	Absolute path in sandbox.
`content`	`str`	File content (text).

Mounted files are read-only. Use write_file() for files that require modification.

Ports

Expose ports so processes inside the sandbox can serve traffic to outside clients:

sandbox = Sandbox.run(
    "python", "-m", "http.server", "8080",
    ports=[
        {"container_port": 8080},
    ],
)

Port configuration

Field	Type	Description
`container_port`	`int`	Port inside the sandbox.

To give a port an internet-reachable address with platform TLS or TLS passthrough, see Public endpoints.

Network

Configure network options using the NetworkOptions dataclass or a plain dict:

from cwsandbox import NetworkOptions, Sandbox

# Using NetworkOptions
sandbox = Sandbox.run(
    network=NetworkOptions(
        ingress_mode="public",
        exposed_ports=(8080,),
    ),
)

# Using dict
sandbox = Sandbox.run(
    network={"ingress_mode": "public", "exposed_ports": [8080]},
)

Both forms are equivalent: the SDK automatically converts dicts to NetworkOptions internally.

NetworkOptions fields

Field	Type	Description
`ingress_mode`	`str \| None`	Controls inbound traffic, such as `"public"` (internet accessible) or `"internal"` (cluster only).
`exposed_ports`	`tuple[int, ...] \| None`	Ports to expose (required with `ingress_mode`). Pass as tuple `(8080,)` or list `[8080]`.
`egress_mode`	`str \| None`	Controls outbound traffic, such as `"internet"` (full access), `"isolated"` (no external), or `"org"` (org-internal only).

All fields are optional and default to None, which uses backend defaults.

Set network in SandboxDefaults

Set a default network configuration for all sandboxes:

from cwsandbox import NetworkOptions, SandboxDefaults, Session

defaults = SandboxDefaults(
    network=NetworkOptions(egress_mode="internet"),
)

with Session(defaults) as session:
    # All sandboxes inherit the network config
    sb1 = session.sandbox()  # Uses egress_mode="internet"
    sb2 = session.sandbox()  # Uses egress_mode="internet"

    # Override for specific sandbox
    sb3 = session.sandbox(network=NetworkOptions(egress_mode="user"))

Annotations

Add Kubernetes pod annotations to sandboxes. Annotations are key-value string pairs attached to the underlying pod, useful for integrations that read pod metadata, such as schedulers, cost-allocation tools, or external automation.

sandbox = Sandbox.run(
    annotations={
        "team": "ml-infra",
        "experiment": "training-run-42",
    },
)

Session-level defaults

Set default annotations for all sandboxes in a session:

defaults = SandboxDefaults(
    annotations={
        "team": "ml-infra",
        "managed-by": "sandbox-sdk",
    },
)

with Session(defaults) as session:
    # Inherits default annotations
    sb1 = session.sandbox()

    # Merge: explicit annotations override defaults on key collision
    sb2 = session.sandbox(
        annotations={"experiment": "run-42", "team": "ml-research"},
    )
    # sb2 receives: team=ml-research, managed-by=sandbox-sdk, experiment=run-42

Merge behavior

When you provide both SandboxDefaults.annotations and per-sandbox annotations, the SDK merges them. Explicit per-sandbox values win on key collision, matching the same semantics as environment_variables.

Validation

The SDK does not validate annotation keys or values. The server handles validation of reserved key prefixes, value format, and maximum entry count.

シークレット

W&B チームのシークレット注入、認証の要件、Python および TypeScript のサンプルについては、W&B シークレットを使用するを参照してください。

Timeouts

Four settings control how long a sandbox and its operations run. The names are similar, so the following table maps each one to what it bounds.

Setting	Bounds	Enforced by	Default
`max_lifetime_seconds`	Total sandbox lifetime before automatic termination	Server	Runner-policy default, or 10 minutes if none is set
`timeout_seconds`	A single `exec()`, `read_file()`, or `write_file()` call	Client	Falls back to `request_timeout_seconds`
`max_timeout_seconds`	How long the server waits to place a sandbox before the attempt fails	Server	30 seconds
`request_timeout_seconds`	Client-side HTTP timeout for most remote procedure calls (RPCs)	Client	300 seconds

Only max_lifetime_seconds terminates a sandbox. The other three bound individual requests and leave the sandbox running.

max_lifetime_seconds

max_lifetime_seconds sets the total wall-clock lifetime of a sandbox. When a sandbox reaches this age, the platform terminates it even if work is still in progress. If you don’t set max_lifetime_seconds, the runner-policy default applies. If no positive policy default supplies a lifetime, the platform uses 10 minutes. Set the value explicitly for long-running work:

from cwsandbox import Sandbox

with Sandbox.run(max_lifetime_seconds=7200) as sandbox:  # 2 hours
    sandbox.exec(["python", "train.py"], timeout_seconds=7000).result()

To apply one lifetime to every sandbox a session creates, set it in SandboxDefaults:

from cwsandbox import Sandbox, SandboxDefaults

defaults = SandboxDefaults(
    container_image="python:3.11",
    max_lifetime_seconds=3600,  # 1 hour
)

with Sandbox.run(defaults=defaults) as sandbox:
    sandbox.exec(["python", "-c", "print('working')"]).result()

An explicit max_lifetime_seconds argument on Sandbox.run() takes precedence over the value in defaults. Because the argument’s default is None, you can’t use it to remove a lifetime that defaults already sets. リクエストできる有効期間の上限は 30 日 (2,592,000 秒) です。これを超える値をリクエストすると、プラットフォームによって拒否されます。

You can’t change max_lifetime_seconds after a sandbox starts. No API extends the lifetime of a running sandbox, so choose the value when you create it. To keep working past the limit, capture what you need and start a new sandbox with a lifetime of up to 30 days.

作成のサンプルについては、長時間実行のサンドボックスを実行するを参照してください。

Behavior when the lifetime expires

The platform enforces the lifetime by terminating the sandbox’s underlying Kubernetes Job, so termination is immediate rather than graceful. Two consequences matter in practice:

Processes inside the sandbox don’t receive the grace period that stop() provides. Anything not already written out is lost.
The client doesn’t receive a distinct error or reason for the termination. The SDK reports no termination reason on a Sandbox, so you can’t distinguish lifetime expiry from other failures programmatically. Track the age of long-lived sandboxes yourself.

Set a lifetime that comfortably exceeds the work you expect, and stop sandboxes explicitly when you finish rather than relying on expiry.

timeout_seconds

timeout_seconds bounds a single operation on a running sandbox:

result = sandbox.exec(
    ["python", "long_script.py"],
    timeout_seconds=300,  # 5 minutes
).result()

This controls how long the client waits for a response. If the wait exceeds the timeout, the SDK raises SandboxTimeoutError. The sandbox keeps running. The clock starts only after the sandbox reaches RUNNING, so startup time doesn’t count against it. Keep timeout_seconds below the sandbox’s max_lifetime_seconds. A command whose timeout exceeds the sandbox’s remaining lifetime stops when the sandbox terminates.

max_timeout_seconds

max_timeout_seconds bounds how long the server waits to place a sandbox on a runner before the attempt fails. It’s a server-side placement budget, not a client read timeout, and it doesn’t affect how long the sandbox runs afterward:

sandbox = Sandbox.run(max_timeout_seconds=120)

The server accepts values from 1 to 300 seconds and defaults to 30. The server clamps values above 300 to 300. If sandboxes fail to start under heavy cluster load, raise the value. max_timeout_seconds is available on Sandbox.run() and session.sandbox(), but not on SandboxDefaults.

request_timeout_seconds

request_timeout_seconds is the client-side HTTP timeout applied to most RPCs, and it’s the fallback when a call doesn’t pass its own timeout_seconds. Set it in SandboxDefaults:

defaults = SandboxDefaults(request_timeout_seconds=600.0)

Status polling uses poll_rpc_timeout_seconds instead, so a stalled poll fails fast rather than blocking for the full request timeout. For the full set of client-side tuning options, see the SandboxDefaults reference.

Complete example

The following example combines defaults, per-sandbox overrides, mounted files, and ports into a single configuration that mirrors how a production workload might be launched:

from cwsandbox import NetworkOptions, ResourceOptions, Sandbox, SandboxDefaults

defaults = SandboxDefaults(
    container_image="python:3.11",
    max_lifetime_seconds=1800,  # 30 minutes
    tags=("production", "ml-pipeline"),
    resources=ResourceOptions(
        requests={"cpu": "4", "memory": "8Gi"},
        limits={"cpu": "4", "memory": "8Gi"},
    ),
    network=NetworkOptions(egress_mode="internet"),
)

with Sandbox.run(
    defaults=defaults,
    mounted_files=[
        {
            "path": "/app/config.yaml",
            "content": "model: gpt-4\nmax_tokens: 1000",
        },
    ],
    ports=[
        {"container_port": 8000},
    ],
) as sandbox:
    # Install dependencies
    sandbox.exec(["pip", "install", "fastapi", "uvicorn"]).result()

    # Run application
    result = sandbox.exec(
        ["python", "-c", "print('Server started')"],
        timeout_seconds=60,
    ).result()
    print(result.stdout)

​Overview

​Resources

​ResourceOptions fields

​Guaranteed QoS

​CPU values

​Memory values

​GPU

​Configuration library interop

​Mounted files

​Mount options

​Ports

​Port configuration

​Network

​NetworkOptions fields

​Set network in SandboxDefaults

​Annotations

​Session-level defaults

​Merge behavior

​Validation

​シークレット

​Timeouts

​max_lifetime_seconds

​Behavior when the lifetime expires

​timeout_seconds

​max_timeout_seconds

​request_timeout_seconds

​Complete example

Overview

Resources

ResourceOptions fields

Guaranteed QoS

CPU values

Memory values

GPU

Configuration library interop

Mounted files

Mount options

Ports

Port configuration

Network

NetworkOptions fields

Set network in SandboxDefaults

Annotations

Session-level defaults

Merge behavior

Validation

シークレット

Timeouts

max_lifetime_seconds

Behavior when the lifetime expires

timeout_seconds

max_timeout_seconds

request_timeout_seconds

Complete example