Your Ray application may have dependencies that exist outside of your Ray script. For example:
* Your Ray script may import/depend on some Python packages.
* Your Ray script may be looking for some specific environment variables to be available.
* Your Ray script may import some files outside of the script.
One frequent problem when running on a cluster is that Ray expects these "dependencies" to exist on each Ray node. If these are not present, you may run into issues such as ``ModuleNotFoundError``, ``FileNotFoundError`` and so on.
To address this problem, you can (1) prepare your dependencies on the cluster in advance using the Ray :ref:`Cluster Launcher <ref-cluster-quick-start>`, or (2) use Ray's :ref:`runtime environments<runtime-environments>` to install them on the fly.
For production usage or non-changing environments, we recommend installing your dependencies into a container image and specifying the image using the Cluster Launcher.
-**Ray Application**. A program including a Ray script that calls ``ray.init()`` and uses Ray tasks or actors.
-**Dependencies**, or **Environment**. Anything outside of the Ray script that your application needs to run, including files, packages, and environment variables.
-**Files**: Code files, data files or other files that your Ray application needs to run.
-**Packages**: External libraries or executables required by your Ray application, often installed via ``pip`` or ``conda``.
-**Local machine** and **Cluster**. Usually, you may want to separate the Ray cluster compute machines/pods from the machine/pod that handles and submits the application. You can submit a Ray Job via :ref:`the Ray Job Submission mechanism <jobs-overview>`, or the :ref:`Ray Client<ray-client>` to connect to a cluster interactively. We call the machine submitting the job your *local machine*.
-**Job**. A period of execution between connecting to a cluster with ``ray.init()`` and disconnecting by calling ``ray.shutdown()`` or exiting the Ray script.
- You can build all your files and dependencies into a container image and specify this in your your :ref:`Cluster YAML Configuration<cluster-config>`.
- You can also install packages using ``setup_commands`` in the Ray Cluster configuration file (:ref:`reference<cluster-configuration-setup-commands>`); these commands will be run as each node joins the cluster.
This feature requires a full installation of Ray using ``pip install "ray[default]"``. This feature is available starting with Ray 1.4.0 and is currently supported on macOS and Linux. It is experimentally supported on Windows.
A **runtime environment** describes the dependencies your Ray application needs to run, including :ref:`files, packages, environment variables, and more <runtime-environments-api-ref>`. It is installed dynamically on the cluster at runtime.
You can specify a runtime environment for your whole job, whether running a script directly on the cluster, using :ref:`Ray Job submission <jobs-overview>`, or using :ref:`Ray Client<ray-client>`:
This will install the dependencies to the remote cluster. Any tasks and actors used in the job will use this runtime environment unless otherwise specified.
This allows you to have actors and tasks running in their own environments, independent of the surrounding environment. (The surrounding environment could be the job's runtime environment, or the system environment of the cluster.)
..warning::
Ray does not guarantee compatibility between tasks and actors with conflicting runtime environments.
For example, if an actor whose runtime environment contains a ``pip`` package tries to communicate with an actor with a different version of that package, it can lead to unexpected behavior such as unpickling errors.
Common Workflows
^^^^^^^^^^^^^^^^
This section describes some common use cases for runtime environments. These use cases are not mutually exclusive; all of the options described below can be combined in a single runtime environment.
For a development workflow, these might live on your local machine, but when it comes time to run things at scale, you will need to get them to your remote cluster.
The following simple example explains how to get your local files on the cluster.
..code-block:: python
# /path/to/files is a directory on the local machine.
# /path/to/files/hello.txt contains the string "Hello World!"
# Create a Ray task, which inherits the above runtime env.
@ray.remote
def f():
# The function will have its working directory changed to its node's
# local copy of /path/to/files.
return open("hello.txt").read()
print(ray.get(f.remote())) # Hello World!
..note::
The example above is written to run on a local machine, but as for all of these examples, it also works when specifying a Ray cluster to connect to
(e.g., using ``ray.init("ray://123.456.7.89:10001", runtime_env=...)`` or ``ray.init(address="auto", runtime_env=...)``).
The specified local directory will automatically be pushed to the cluster nodes when ``ray.init()`` is called.
You can also specify files via a remote cloud storage URI; see :ref:`remote-uris` for details.
Using ``conda`` or ``pip`` packages
"""""""""""""""""""""""""""""""""""
Your Ray application might depend on Python packages (for example, ``pendulum`` or ``requests``) via ``import`` statements.
Ray ordinarily expects all imported packages to be preinstalled on every node of the cluster; in particular, these packages are not automatically shipped from your local machine to the cluster or downloaded from any repository.
However, using runtime environments you can dynamically specify packages to be automatically downloaded and installed in an isolated virtual environment for your Ray job, or for specific Ray tasks or actors.
..code-block:: python
import ray
import requests
# This example runs on a local machine, but you can also do
# ray.init(address=..., runtime_env=...) to connect to a cluster.
You may also specify your ``pip`` dependencies either via a Python list or a ``requirements.txt`` file.
Alternatively, you can specify a ``conda`` environment, either as a Python dictionary or via a ``environment.yml`` file. This conda environment can include ``pip`` packages.
For details, head to the :ref:`API Reference<runtime-environments-api-ref>`.
..note::
The ``ray[default]`` package itself will automatically be installed in the isolated environment. However, if you are using any Ray libraries (for example, Ray Serve), then you will need to specify the library in the runtime environment (e.g. ``runtime_env = {"pip": ["requests", "ray[serve]"}]}``.)
..warning::
Since the packages in the ``runtime_env`` are installed at runtime, be cautious when specifying ``conda`` or ``pip`` packages whose installations involve building from source, as this can be slow.
Library Development
"""""""""""""""""""
Suppose you are developing a library ``my_module`` on Ray.
A typical iteration cycle will involve
1. Making some changes to the source code of ``my_module``
2. Running a Ray script to test the changes, perhaps on a distributed cluster.
To ensure your local changes show up across all Ray workers and can be imported properly, use the ``py_modules`` field.
Note: This feature is currently limited to modules that are packages with a single directory containing an ``__init__.py`` file. For single-file modules, you may use ``working_dir``.
The ``runtime_env`` is a Python dictionary or a python class :class:`ray.runtime_env.RuntimeEnv <ray.runtime_env.RuntimeEnv>` including one or more of the following fields:
-``working_dir`` (str): Specifies the working directory for the Ray workers. This must either be (1) an local existing directory with total size at most 100 MiB, (2) a local existing zipped file with total unzipped size at most 100 MiB (Note: ``excludes`` has no effect), or (3) a URI to a remotely-stored zip file containing the working directory for your job. See :ref:`remote-uris` for details.
-``py_modules`` (List[str|module]): Specifies Python modules to be available for import in the Ray workers. (For more ways to specify packages, see also the ``pip`` and ``conda`` fields below.)
Each entry must be either (1) a path to a local directory, (2) a URI to a remote zip file (see :ref:`remote-uris` for details), (3) a Python module object, or (4) a path to a local `.whl` file.
The modules will be downloaded to each node on the cluster.
Note: Setting options (1) and (3) per-task or per-actor is currently unsupported, it can only be set per-job (i.e., in ``ray.init()``).
Note: For option (1), if your local directory contains a ``.gitignore`` file, the files and paths specified therein will not be uploaded to the cluster.
Note: This feature is currently limited to modules that are packages with a single directory containing an ``__init__.py`` file. For single-file modules, you may use ``working_dir``.
-``excludes`` (List[str]): When used with ``working_dir`` or ``py_modules``, specifies a list of files or paths to exclude from being uploaded to the cluster.
-``pip`` (dict | List[str] | str): Either (1) a list of pip `requirements specifiers <https://pip.pypa.io/en/stable/cli/pip_install/#requirement-specifiers>`_, (2) a string containing the path to a pip
`“requirements.txt” <https://pip.pypa.io/en/stable/user_guide/#requirements-files>`_ file, or (3) a python dictionary that has three fields: (a) ``packages`` (required, List[str]): a list of pip packages,
(b) ``pip_check`` (optional, bool): whether to enable `pip check <https://pip.pypa.io/en/stable/cli/pip_check/>`_ at the end of pip install, defaults to ``False``.
(c)``pip_version`` (optional, str): the version of pip; Ray will spell the package name "pip" in front of the ``pip_version`` to form the final requirement string.
The syntax of a requirement specifier is defined in full in `PEP 508 <https://www.python.org/dev/peps/pep-0508/>`_.
When specifying a ``requirements.txt`` file, referencing local files `within` that file is not supported (e.g. ``-r ./my-laptop/more-requirements.txt``, ``./my-pkg.whl``).
or (3) the name of a local conda environment already installed on each node in your cluster (e.g., ``"pytorch_p36"``).
In the first two cases, the Ray and Python dependencies will be automatically injected into the environment to ensure compatibility, so there is no need to manually include them.
Note that the ``conda`` and ``pip`` keys of ``runtime_env`` cannot both be specified at the same time---to use them together, please use ``conda`` and add your pip dependencies in the ``"pip"`` field in your conda ``environment.yaml``.
-``env_vars`` (Dict[str, str]): Environment variables to set. Environment variables already set on the cluster will still be visible to the Ray workers; so there is
no need to include ``os.environ`` or similar in the ``env_vars`` field.
Note: ``container`` is experimental now. If you have some requirements or run into any problems, raise issues in `github <https://github.com/ray-project/ray/issues>`__.
-``eager_install`` (bool): Indicates whether to install the runtime environment on the cluster at ``ray.init()`` time, before the workers are leased. This flag is set to ``True`` by default.
If set to ``False``, the runtime environment will be only installed when the first task is invoked or when the first actor is created.
Currently, specifying this option per-actor or per-task is not supported.
-``config`` (dict | :class:`ray.runtime_env.RuntimeEnvConfig <ray.runtime_env.RuntimeEnvConfig>`): config for runtime environment. Either a dict or a RuntimeEnvConfig. Field: (1) setup_timeout_seconds, the timeout of runtime environment creation, timeout is in seconds.
Runtime environment resources on each node (such as conda environments, pip packages, or downloaded ``working_dir`` or ``py_modules`` directories) will be cached on the cluster to enable quick reuse across different runtime environments within a job. Each field (``working_dir``, ``py_modules``, etc.) has its own cache whose size defaults to 10 GB. To change this default, you may set the environment variable ``RAY_RUNTIME_ENV_<field>_CACHE_SIZE_GB`` on each node in your cluster before starting Ray e.g. ``export RAY_RUNTIME_ENV_WORKING_DIR_CACHE_SIZE_GB=1.5``.
When the cache size limit is exceeded, resources not currently used by any actor, task or job will be deleted.
To disable all deletion behavior (for example, for debugging purposes) you may set the environment variable ``RAY_RUNTIME_ENV_SKIP_LOCAL_GC`` to ``1`` on each node in your cluster before starting Ray.
The runtime environment is inheritable, so it will apply to all tasks/actors within a job and all child tasks/actors of a task or actor once set, unless it is overridden by explicitly specifying a runtime environment for the child task/actor.
3. If you'd like to still use current runtime env, you can use the API :ref:`ray.get_runtime_context().runtime_env <runtime-context-apis>` to get the current runtime env and modify it by yourself.
If a runtime environment is specified in ``ray.init(runtime_env=...)``, then the environment will be installed on every node. See :ref:`Per-Job <rte-per-job>` for more details.
(Note, by default the runtime environment will be installed eagerly on every node in the cluster. If you want to lazily install the runtime environment on demand, set the ``eager_install`` option to false: ``ray.init(runtime_env={..., "eager_install": False}``.)
When specified per-task or per-actor, the environment is installed when the task is invoked or the actor is instantiated (i.e. when you call ``my_task.remote()`` or ``my_actor.remote()``.)
See :ref:`Per-Job <rte-per-job>`:ref:`Per-Task/Actor, within a job <rte-per-task-actor>` for more details.
Where are the environments cached?
""""""""""""""""""""""""""""""""""
Any local files downloaded by the environments are cached at ``/tmp/ray/session_latest/runtime_resources``.
How long does it take to install or to load from cache?
The install time usually mostly consists of the time it takes to run ``pip install`` or ``conda create`` / ``conda activate``, or to upload/download a ``working_dir``, depending on which ``runtime_env`` options you're using.
This could take seconds or minutes.
On the other hand, loading a runtime environment from the cache should be nearly as fast as the ordinary Ray worker startup time, which is on the order of a few seconds. A new Ray worker is started for every Ray actor or task that requires a new runtime environment.
(Note that loading a cached ``conda`` environment could still be slow, since the ``conda activate`` command sometimes takes a few seconds.)
You can set ``setup_timeout_seconds`` config to avoid the installation hanging for a long time. If the installation is not finished within this time, your tasks or actors will fail to start.
Check for hidden files and metadata directories (e.g. ``__MACOSX/``) in zipped dependencies.
You can inspect a zip file's contents by running the ``zipinfo -1 zip_file_name.zip`` command in the Terminal.
Some zipping methods can cause hidden files or metadata directories to appear in the zip file at the top level.
This will cause Ray to throw an error because the structure of the zip file is invalid since there is more than a single directory at the top level.
You can avoid this by using the ``zip -r`` command directly on the directory you want to compress.
Currently, three types of remote URIs are supported for hosting ``working_dir`` and ``py_modules`` packages:
-``HTTPS``: ``HTTPS`` refers to URLs that start with ``https``.
These are particularly useful because remote Git providers (e.g. GitHub, Bitbucket, GitLab, etc.) use ``https`` URLs as download links for repository archives.
This allows you to host your dependencies on remote Git providers, push updates to them, and specify which dependency versions (i.e. commits) your jobs should use.
To use packages via ``HTTPS`` URIs, you must have the ``smart_open`` library (you can install it using ``pip install smart_open``).
-``S3``: ``S3`` refers to URIs starting with ``s3://`` that point to compressed packages stored in `AWS S3 <https://aws.amazon.com/s3/>`_.
To use packages via ``S3`` URIs, you must have the ``smart_open`` and ``boto3`` libraries (you can install them using ``pip install smart_open`` and ``pip install boto3``).
Ray does not explicitly pass in any credentials to ``boto3`` for authentication.
``boto3`` will use your environment variables, shared credentials file, and/or AWS config file to authenticate access.
See the `AWS boto3 documentation <https://boto3.amazonaws.com/v1/documentation/api/latest/guide/credentials.html>`_ to learn how to configure these.
-``GS``: ``GS`` refers to URIs starting with ``gs://`` that point to compressed packages stored in `Google Cloud Storage <https://cloud.google.com/storage>`_.
To use packages via ``GS`` URIs, you must have the ``smart_open`` and ``google-cloud-storage`` libraries (you can install them using ``pip install smart_open`` and ``pip install google-cloud-storage``).
Ray does not explicitly pass in any credentials to the ``google-cloud-storage``'s ``Client`` object.
``google-cloud-storage`` will use your local service account key(s) and environment variables by default.
Follow the steps on Google Cloud Storage's `Getting started with authentication <https://cloud.google.com/docs/authentication/getting-started>`_ guide to set up your credentials, which allow Ray to access your remote package.
You can store your dependencies in repositories on a remote Git provider (e.g. GitHub, Bitbucket, GitLab, etc.), and you can periodically push changes to keep them updated.
In this section, you will learn how to store a dependency on GitHub and use it in your runtime environment.
..note::
These steps will also be useful if you use another large, remote Git provider (e.g. BitBucket, GitLab, etc.).
For simplicity, this section refers to GitHub alone, but you can follow along on your provider.
First, create a repository on GitHub to store your ``working_dir`` contents or your ``py_module`` dependency.
By default, when you download a zip file of your repository, the zip file will already contain a single top-level directory that holds the repository contents,
so you can directly upload your ``working_dir`` contents or your ``py_module`` dependency to the GitHub repository.
Once you have uploaded your ``working_dir`` contents or your ``py_module`` dependency, you need the HTTPS URL of the repository zip file, so you can specify it in your ``runtime_env`` dictionary.
You have two options to get the HTTPS URL.
Option 1: Download Zip (quicker to implement, but not recommended for production environments)
The second option is to manually create this URL by pattern-matching your specific use case with one of the following examples.
**This is recommended** because it provides finer-grained control over which repository branch and commit to use when generating your dependency zip file.
These options prevent consistency issues on Ray Clusters (see the warning above for more info).
To create the URL, pick a URL template below that fits your use case, and fill in all parameters in brackets (e.g. [username], [repository], etc.) with the specific values from your repository.
For instance, suppose your GitHub username is ``example_user``, the repository's name is ``example_repository``, and the desired commit hash is ``abcdefg``.
If ``example_repository`` is public and you want to retrieve the ``abcdefg`` commit (which matches the first example use case), the URL would be:
You can also enable ``runtime_env`` debugging log streaming by setting an environment variable ``RAY_RUNTIME_ENV_LOG_TO_DRIVER_ENABLED=1`` on each node before starting Ray, for example using ``setup_commands`` in the Ray Cluster configuration file (:ref:`reference<cluster-configuration-setup-commands>`).
This will print the full ``runtime_env`` setup log messages to the driver (the script that calls ``ray.init()``).
Example log output:
..code-block:: text
>>> ray.init(runtime_env={"pip" ["requests"]})
(pid=runtime_env) 2022-02-28 14:12:33,653 INFO pip.py:188 -- Creating virtualenv at /tmp/ray/session_2022-02-28_14-12-29_909064_87908/runtime_resources/pip/0cc818a054853c3841171109300436cad4dcf594/virtualenv, current python dir /Users/user/anaconda3/envs/ray-py38
(pid=runtime_env) 2022-02-28 14:12:33,653 INFO utils.py:76 -- Run cmd[1] ['/Users/user/anaconda3/envs/ray-py38/bin/python', '-m', 'virtualenv', '--app-data', '/tmp/ray/session_2022-02-28_14-12-29_909064_87908/runtime_resources/pip/0cc818a054853c3841171109300436cad4dcf594/virtualenv_app_data', '--reset-app-data', '--no-periodic-update', '--system-site-packages', '--no-download', '/tmp/ray/session_2022-02-28_14-12-29_909064_87908/runtime_resources/pip/0cc818a054853c3841171109300436cad4dcf594/virtualenv']
(pid=runtime_env) 2022-02-28 14:12:34,267 INFO utils.py:97 -- Output of cmd[1]: created virtual environment CPython3.8.11.final.0-64 in 473ms
Regardless of the value of ``RAY_RUNTIME_ENV_LOG_TO_DRIVER_ENABLED``, these logs can always be found in the file ``runtime_env_setup-[job_id].log`` for per-actor, per-task and per-job environments, or in
``runtime_env_setup-ray_client_server_[port].log`` for per-job environments when using Ray Client.
See :ref:`Logging Directory Structure <logging-directory-structure>` for more details.
