Skip to content

Commit

Permalink
Storage mounting (#658)
Browse files Browse the repository at this point in the history 
* squash

* fix

* yapf workaround
  • Loading branch information
@romilbhardwaj
romilbhardwaj authored Apr 3, 2022
1 parent 795b872 commit de5ac9b
Show file tree
Hide file tree
Showing 24 changed files with 1,704 additions and 732 deletions.
266 changes: 181 additions & 85 deletions docs/source/reference/storage.rst
View file
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
.. _sky-storage:

Sky Storage
=======
===========

A Sky Storage object represents an abstract data store containing large data
files required by the task. Compared to file_mounts, storage is faster and
Expand All @@ -9,13 +10,20 @@ Behind the scenes, storage automatically uploads all data in the source
to a backing object store in a particular cloud (S3/GCS/Azure Blob).

A storage object is used by "mounting" it to a task. On mounting, the data
specified in the source becomes available at the destination mount_path.
Please note that sky.Storage does not guarantee preservation of file
permissions - you may need to set file permissions during task execution.
specified in the source becomes available at the destination mount path.

A storage object can used in either :code:`MOUNT` mode or :code:`COPY` mode.

* In :code:`MOUNT` mode, the backing store is directly "mounted" to the remote VM.
I.e., files are fetched when accessed by the task and files written to the
mount path are also written to the remote store.

* In :code:`COPY` mode, the files are pre-fetched and cached on the local disk.
Writes are not replicated on the remote store.

.. note::
Sky file mounting currently does not support syncing writes.
Any writes made at a mounted folder will not reflect at the mounting source.
sky.Storage does not guarantee preservation of file
permissions - you may need to set file permissions during task execution.

Using Sky Storage
-----------------
Expand All @@ -27,141 +35,229 @@ the files to a cloud store (e.g. S3, GCS) and have them persist there by
specifying the :code:`name`, :code:`source` and :code:`persistent` fields. By
enabling persistence, file_mount sync can be made significantly faster.

.. note::
Symbolic links are handled differently in :code:`file_mounts` depending on whether Sky Storage is used. For mounts backed by Sky Storage, referenced data for all symbolic links is copied to remote. For mounts not using Sky Storage (e.g., those using rsync) the symbolic links are directly copied. Their targets must be separately mounted or else the symlinks may break.
Your usage of sky storage can fall under four broad use cases:

1. **You want to upload your local data to remote VM -** specify the name and
source fields. Name sets the bucket name that will be used, and source
specifies the local path to be uploaded.

2. **You want to mount an existing S3/GCS bucket to your remote VM -** specify
just the source field (e.g., s3://my-bucket/)

3. **You want to have a write-able path to directly write files to S3 buckets
-** specify a name (to create a bucket if it doesn't exist) and set the mode
to MOUNT. This is useful for writing code outputs, such as checkpoints or
logs directly to a S3 bucket.

4. **You want to have a shared file-system across workers running on different
nodes -** specify a name (to create a bucket if it doesn't exist) and set
the mode to MOUNT. This will create an empty scratch space that workers
can write to. Any writes will show up on all worker's mount points.

When specifying a storage object, you can specify either of two modes:

- :code:`mode: MOUNT` (default)
This mode directly mounts the bucket at the specified path on the VM.
In effect, files are streamed from the backing source bucket as and when
they are accessed by applications. This mode also allows applications to
write to the mount path. All writes are replicated to remote bucket (and
any other VMs mounting the same bucket). Please note that this mode
uses a close-to-open consistency model, which means a file write is
committed to the backing store only after :code:`close()` is called on it.

- :code:`mode: COPY`
This mode pre-fetches your files from remote storage and caches them on the
local disk. Note that in this mode, any writes to the mount path are not
replicated to the source bucket.

Here are a few examples covering a range of use cases for sky file_mounts
and storage mounting:

.. code-block:: yaml
name: storage-demo
resources:
cloud: aws
instance_type: m5.2xlarge
file_mounts:
# This uses rsync to directly copy files from your machine to the remote
# VM at /datasets. Since this uses rsync, the ~/datasets folder is
# uploaded on each execution.
# *** Copying files from local ***
#
# This uses rsync to directly copy files from your machine to the remote VM at
# /datasets.
/datasets: ~/datasets
# *** Copying files from S3 ***
#
# This re-uses a predefined bucket (public bucket used here, but can be
# private) and copies it's contents directly to /datasets-s3.
/datasets-s3: s3://enriched-topical-chat
# *** Copying files from GCS ***
#
# This copies a single object (train-00001-of-01024) from a remote cloud
# storage to local disk.
/train-00001-of-01024: gs://cloud-tpu-test-datasets/fake_imagenet/train-00001-of-01024
# *** Persistent Data Storage by copying from S3 ***
#
# This uses sky Storage to first create a S3 bucket named sky-dataset,
# copies the contents of ~/datasets to the remote bucket and makes the
# bucket persistent (i.e., the bucket is not deleted after the completion of
# this sky task, and future invocations of this bucket will be much faster).
# The bucket is mounted at /datasets-storage.
# If the bucket already exists, it is fetched and re-used.
# When the VM is initialized, the contents of the bucket are copied to
# /datasets-storage. If the bucket already exists, it is fetched and re-used.
/datasets-storage:
name: sky-dataset
name: sky-dataset-romil # Make sure this name is unique or you own this bucket
source: ~/datasets
force_stores: [s3] # Could be [s3, gcs], [gcs] default: None
store: s3 # Could be either of [s3, gcs]; default: None
persistent: True # Defaults to True, can be set to false.
mode: COPY # Defaults to MOUNT if not specified
# *** Persistent Data Storage by MOUNTING S3 ***
#
# This uses the exact same storage object defined above, but uses the MOUNT
# mode. This means instead of copying contents of the remote bucket to the VM,
# sky "mounts" the bucket at /dataset-storage-mount. Files are streamed from
# S3 as they are read by the task. Any writes made at /dataset-storage-mount
# are also replicated on the remote S3 bucket and any other storage mounts
# using the same bucket with MOUNT mode. Note that the source is synced with
# the remote bucket everytime this task is run.
/dataset-storage-mount:
name: sky-dataset-romil
source: ~/datasets
mode: MOUNT
# *** Mounting very large public buckets ***
#
# This uses the MOUNT mode to mount a mount at 3.5 TB public bucket at the
# specified path. Since MOUNT mode is used, the bucket is not copied at init,
# instead contents are streamed from S3 as they are requested. This saves disk
# space on the remote VM.
# Since this is a public bucket, any writes to the path will fail.
/huge-dataset-mount:
source: s3://digitalcorpora
mode: MOUNT
# *** Collecting outputs of tasks on S3 ***
#
# This uses the MOUNT mode to create an output mount path. This creates an
# empty bucket with the specified name and mounts it at the path.
# Any files written to /outputs-mount will also be synced to my-output-bucket.
# This is useful when you want to collect outputs of your task directly in a
# S3 bucket and browse it from your laptop later.
#
# Since writes are synced across workers mounting the same bucket,
# this approach can also be used to create a shared filesystem across workers.
# See examples/storage/pingpong.yaml for an example.
/outputs-mount:
name: romil-output-bucket
mode: MOUNT
run: |
pwd
ls -la /
If you have files that already exist on remote cloud stores, you can also
directly mount s3/gcs buckets and objects in your remote VM by providing the
path to the s3/gcs bucket or object.
# Remember to run `sky storage ls` and `sky storage delete` to delete the
# created storage objects!
.. code-block:: yaml
.. note::
Stopping a running cluster will cause any Storage mounted with :code:`MOUNT`
mode to be unmounted. These mounts will not be re-mounted on running
:code:`sky start`, or even :code:`sky exec`. Please run :code:`sky launch`
again on the same cluster to ensure :code:`MOUNT` mode Storages are mounted
again.

name: storage-demo
.. note::
Symbolic links are handled differently in :code:`file_mounts` depending on whether Sky Storage is used.
For mounts backed by Sky Storage, referenced data for all symbolic links is copied to remote.
For mounts not using Sky Storage (e.g., those using rsync) the symbolic links are directly copied.
Their targets must be separately mounted or else the symlinks may break.

resources:
cloud: aws
instance_type: m5.2xlarge
Creating a shared file system
-----------------------------

Sky Storage can also be used to create a shared file-system that multiple tasks
on different nodes can read and write to. This allows developers to pass files
between workers and even use files as a medium for inter-process communication (IPC).

To create a shared filesystem, simply create a Storage object without a source
and use mount mode when attaching it to your tasks like so:

.. code-block:: yaml
file_mounts:
# This re-uses a predefined bucket (sky-dataset, defined above) and
# mounts it directly at /datasets-s3.
/datasets-s3: s3://sky-dataset
/sharedfs:
name: my-sky-sharedfs
mode: MOUNT
# This copies a single object (train-00001-of-01024) from a remote cloud
# storage to local disk.
/train-00001-of-01024: gs://cloud-tpu-test-datasets/fake_imagenet/train-00001-of-01024
run: |
pwd
ls -la /
Here is a `simple example <https://github.com/sky-proj/sky/blob/master/examples/storage/pingpong.yaml>`_
using sky storage to perform communication between processes using files.

Alternate Usage - Declarative Storage API
------------------------------------------
.. warning::
The declarative storage YAML API has been deprecated.
If you need to create Storage objects but not mount them, use the storage
CLI once it is supported.

Some power users may want to only upload their files to an object store
without mounting it, while others may want to re-use pre-existing storage
objects. They can do so using the storage and storage_mount fields, which are
at 1:1 parity with the sky.Storage python API.
Using Sky Storage CLI tools
---------------------------

Here's an example using the declarative API.
To manage persistent Storage objects, the sky CLI provides two useful commands -
:code:`sky storage ls` and :code:`sky storage delete`.

.. code-block:: yaml
1. :code:`sky storage ls` shows the currently provisioned Storage objects.

name: storage-demo
.. code-block:: console
resources:
cloud: aws
instance_type: m5.2xlarge
$ sky storage ls
NAME CREATED STORE COMMAND STATUS
sky-dataset-romil 3 mins ago S3 sky launch -c demo examples/storage_demo.yaml READY
storage:
- name: sky-dataset-decl
source: ~/datasets
#force_stores: [s3] # Could be [s3, gcs], [gcs] default: None
persistent: True
2. :code:`sky storage delete` allows you to delete any Storage objects managed
by sky.

storage_mounts:
- storage: sky-dataset-decl # Name of the storage defined above
mount_path: /datasets-decl # Path to mount the storage at
.. code-block:: console
run: |
pwd
ls -la /
$ sky storage delete sky-dataset-romil
Deleting storage object sky-dataset-romil...
I 04-02 19:42:24 storage.py:336] Detected existing storage object, loading Storage: sky-dataset-romil
I 04-02 19:42:26 storage.py:683] Deleting S3 Bucket sky-dataset-romil
.. note::
:code:`sky storage ls` only shows Storage objects whose buckets were created
by sky. Storage objects using externally managed buckets or public buckets
are not listed in :code:`sky storage ls` and cannot be managed through sky.

Storage YAML field reference:
Storage YAML reference
----------------------

::

storage: List[sky.Storage]
sky.Storage

Fields:
sky.Storage.name: str
Identifier for the storage object, used as reference in storage_mount
Identifier for the storage object.

sky.Storage.source: str
The source attribute specifies the local path that must be made available
in the storage object. It can either be a local path, in which case data
is uploaded to the cloud to an appropriate object store (s3 or gcs), or it
can be a remote path (s3://, gs://), in which case it is mounted directly.
can be a remote path (s3://, gs://), in which case it is copied or mounted
directly (see mode flag below).

sky.Storage.force_stores: List[str]
If you wish to force sky.Storage to be backed by specific cloud object
stores, you can specify them here. If the Storage object does not already
exist there, it will be replicated onto those clouds.
sky.Storage.store: str; either of 's3' or 'gcs'
If you wish to force sky.Storage to be backed by a specific cloud object
store, you can specify it here.

sky.Storage.persistent: str
sky.Storage.persistent: bool
Whether the remote backing stores in the cloud should be deleted after
execution of this task or not. Set to True to avoid uploading files again
in subsequent runs (at the cost of storing your data in the cloud). If
files change between runs, new files are synced to the bucket.


Storage Mounts YAML field reference:

::

storage_mounts: List[sky.storage_mounts]

Storage mounts specify where the storage objects defined above should be
mounted when the task is run.

Fields:
sky.StorageMount.storage: str
Name reference to the storage object being mounted

sky.StorageMount.mount_path: str
Path where the storage object is to be mounted
sky.Storage.mode: str; either of MOUNT or COPY, defaults to MOUNT
Whether to mount the storage object by copying files, or actually
mounting the remote storage object. With MOUNT mode, files are streamed
from the remote object store and writes are replicated to the object
store (and consequently, to other workers mounting the same Storage).
With mount mode, files are copied at VM initialization and any writes to
the mount path will not be replicated on the object store.
9 changes: 5 additions & 4 deletions docs/source/reference/yaml-spec.rst
View file
Original file line number Diff line number Diff line change
Expand Up @@ -67,10 +67,11 @@ describe all fields available.
#
# Mounts the bucket at /datasets-storage on every node of the cluster.
/datasets-storage:
name: sky-dataset
source: /local/path/datasets
force_stores: [s3] # Could be [s3, gcs], [gcs]; default: None
persistent: True # Defaults to True; can be set to false
name: sky-dataset # Name of storage, optional when source is bucket URI
source: /local/path/datasets # Source path, can be local or s3/gcs URL. Optional, do not specify to create an empty bucket.
store: s3 # Could be either 's3' or 'gcs'; default: None. Optional.
persistent: True # Defaults to True; can be set to false. Optional.
mode: MOUNT # Either MOUNT or COPY. Optional.
# Copies a cloud object store URI to the cluster. Can be private buckets.
/datasets-s3: s3://my-awesome-dataset
Expand Down
Loading

0 comments on commit de5ac9b

Please sign in to comment.