|
| 1 | +## Summary |
| 2 | + |
| 3 | +### General Motivation |
| 4 | + |
| 5 | +Ray users can use the submit_job Api to send jobs to the ray cluster. |
| 6 | + |
| 7 | +At the same time, they can specify the location of the project's working directory in a remote file system, for example s3. |
| 8 | + |
| 9 | +This document suggests expanding the list of supported file systems and adding support for OBS (Object Storage Service). |
| 10 | + |
| 11 | +After the implementation of this improvement, the user will be able to run submit tasks specifying the location of the working directory in OBS. |
| 12 | + |
| 13 | + |
| 14 | +The UserCase of submiting ray job via obs is shown in Figure 1: |
| 15 | + |
| 16 | +- First, the user startups the remote ray cluster, and then upload the codes to the obs; |
| 17 | + |
| 18 | +- Second, the user submits ray job to OBS using the submit_job APIs, as well as setting the obs path in the submit_job API |
| 19 | + |
| 20 | +- Third, after submitting the ray job, the ray cluster automatically downloads, uncompresses, and executes the codes from the OBS |
| 21 | + |
| 22 | + |
| 23 | + |
| 24 | + |
| 25 | +### Should this change be within `ray` or outside? |
| 26 | + |
| 27 | +main `ray` project. Changes are made to Ray core components. |
| 28 | + |
| 29 | +## Stewardship |
| 30 | + |
| 31 | +### Required Reviewers |
| 32 | + |
| 33 | +- @jjyao |
| 34 | +- @ericl |
| 35 | + |
| 36 | + |
| 37 | +### Shepherd of the Proposal (should be a senior committer) |
| 38 | + |
| 39 | +- @ericl |
| 40 | + |
| 41 | + |
| 42 | +### Design and Architecture |
| 43 | + |
| 44 | +To submit a Ray job through OBS, perform the following steps: |
| 45 | + |
| 46 | +1. Compress the code into a zip or jar package and put it in OBS |
| 47 | + |
| 48 | +2. You can use environment variables or configuration files to access the AK, SK, and Endpoint of OBS |
| 49 | + |
| 50 | +3. When you submit a ray job, specify the path of the OBS service |
| 51 | + |
| 52 | +```python |
| 53 | +job_id = client.submit_job( |
| 54 | + entrypoint="python script.py", |
| 55 | + runtime_env={"working_dir": "obs://example_bucket/example_file.zip"} |
| 56 | +) |
| 57 | +``` |
| 58 | + |
| 59 | +4. The ray cluster automatically downloads the example_file.zip from the specified OBS bucket, decompresses it, and then runs the entry file script.py in the working path |
| 60 | + |
| 61 | + |
| 62 | +## Design Insights |
| 63 | + |
| 64 | +Modify the source code of Ray so that it can download and run the OBS code as follows: |
| 65 | + |
| 66 | +1. Parse the OBS path, for example, "obs://example_bucket/example_file.zip"; |
| 67 | + |
| 68 | +2. Read the AK, SK, and endpoint of OBS through environment variables and configuration files to access the remote OBS path. |
| 69 | + |
| 70 | +3. Download the example_file.zip from the specified OBS bucket, decompress it, and execute the user's code |
| 71 | + |
| 72 | +We can extend the ray project to support obs protocol, enabling the ray cluster to parse the obs URI and download the codes from obs: |
| 73 | + |
| 74 | +1. parse obs URI, such as "obs://example_bucket/example_file.zip"; |
| 75 | + |
| 76 | +2. config the obs AK, SK, and Endpoint via environment variables and config files; |
| 77 | + |
| 78 | +3. the ray cluster automatically downloads and execute the codes from obs. |
| 79 | + |
| 80 | +## Implementation Analysis |
| 81 | + |
| 82 | +To extend the ray project for obs, we first need to figure out the workflow of parsing and accessing remote URIs, which is shown in Figure 2. |
| 83 | + |
| 84 | + |
| 85 | + |
| 86 | +After the user submitting remote ray jobs, the ray cluster calls download_and_unpack_package function to download and uncrompress the remote files, as shown in Figure 2. To extend the ray project for OBS, we should extend the download_and_unpack_package function to support the OBS scheme, which is implemented via the following two steps. |
| 87 | + |
| 88 | +**Step 1**: Extend the **parse_uri** function to parse the OBS URIs, which is implemented in the file [ray/_private/runtime_env/packaging](https://github.com/ray-project/ray/blob/master/python/ray/_private/runtime_env/packaging.py). |
| 89 | + |
| 90 | +**Step 2**: Extend the third-party library [smart_open](https://github.com/piskvorky/smart_open) to read OBS objects, which is suggested to implement 3 interfaces, i.e., parse uri, open_uri, and open, as shown in Extending smart_open. |
| 91 | + |
| 92 | + |
| 93 | + |
| 94 | +The 3 interfaces in smart_open have their own intents: |
| 95 | + |
| 96 | +**parse_uri** : parse the remote URI "obs://bucketId/keyId" to obtain the following info: obs (scheme), bucketId, keyId. |
| 97 | + |
| 98 | +**open_uri** : using the parsed URI info to open the remote objects and call the open API to return an IO stream |
| 99 | + |
| 100 | +**open** : access the remote object and open it as an IO stream |
| 101 | + |
| 102 | + |
| 103 | +It is worth noting that we can extend the open API to access the OBS objects, which can be implemented to call the wrapper functions based on [obs SDK](https://pypi.org/project/esdk-obs-python/). |
0 commit comments