Add AWS Credentials parsing from file

[Shershebnev]

Environment

Delta-rs version:

$ pip show deltalake
Name: deltalake
Version: 0.15.1
Summary: Native Delta Lake Python binding based on delta-rs with Pandas integration
Home-page: https://github.com/delta-io/delta-rs
Author: Qingping Hou <dave2008713@gmail.com>, Will Jones <willjones127@gmail.com>
Author-email: Qingping Hou <dave2008713@gmail.com>, Will Jones <willjones127@gmail.com>
License: Apache-2.0
Location: ...
Requires: pyarrow, pyarrow-hotfix
Required-by: 

Binding:
Python
Environment:

• Cloud provider: AWS (Ubuntu)
• OS: MacOS
• Other:

---

Bug

What happened:
It seems that credentials are not correctly obtained from ~/.aws/credentials and ~/.aws/config files. Just like here #1416 I'm getting OSError: Generic S3 error: Missing region when trying to read from S3

On MacOS locally:
Setting AWS_DEFAULT_REGION fixes this, but then it tries to retrieve instance metadata using http://169.254.169.254/latest/api/token which obviously fails when running not from AWS instance OSError: Generic S3 error: Error after 10 retries in 6.409805791s, max_retries:10, retry_timeout:180s, source:error sending request for url (http://169.254.169.254/latest/api/token): error trying to connect: tcp connect error: Host is down (os error 64)

On AWS instance:
Setting only AWS_DEFAULT_REGION results in OSError: Generic S3 error: Client error with status 403 Forbidden: <?xml version="1.0" encoding="UTF-8"?>

In both cases setting everything through env variables fixes the problem, e.g. AWS_DEFAULT_REGION=... AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... python. Other tools like boto3 don't have problems using credentials stored in default location:

$ python3.9
Python 3.9.16 (main, Aug  3 2023, 01:00:02) 
[GCC 9.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import polars as pl
>>> master_table_df = pl.scan_delta("s3://REDACTED.delta").select("audio_type", "parallel_id").collect()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/ubuntu/.local/lib/python3.9/site-packages/polars/io/delta.py", line 263, in scan_delta
    dl_tbl = _get_delta_lake_table(
  File "/home/ubuntu/.local/lib/python3.9/site-packages/polars/io/delta.py", line 306, in _get_delta_lake_table
    dl_tbl = deltalake.DeltaTable(
  File "/home/ubuntu/.local/lib/python3.9/site-packages/deltalake/table.py", line 396, in __init__
    self._table = RawDeltaTable(
OSError: Generic S3 error: Missing region
>>> import boto3
>>> ecr = boto3.client("ecr")
>>> 

What you expected to happen:
Credentials are properly read from default location ~/.aws/credentials and ~/.aws/config
How to reproduce it:
Install deltalake and try to read from S3 while having credentials set in default files. See example above with polars and deltalake

[r3stl355]

As far as I know, deltalake expects all the AWS parameters to be defined in the environment, exactly as you noted @Shershebnev.

~/.aws/* files on Linux/Mac are managed by AWS cli, and boto3 being an AWS SDK piggybacks on it to get the profile configs. If we implement the same/similar in deltalake then we'd have to either implement a dependency on AWS cli (or some sub-component of it) which, if possible, would just create another dependency we'd have to maintain, or use the "knowledge" of the possible locations of these files but this will have to be OS file system specific, and will also create a dependency on the corresponding file format.

To summarize, I don't see this as something that should be prioritized but If there is a strong support for implementing this I can have a stab

[ion-elgreco]

@r3stl355 Polars for example parses these files to grab the credentials, could likely take inspiration from that implementation

[r3stl355]

Yes @ion-elgreco, looks like Polars is using the second approach I mentioned - looking into specific files config and credentials files it "knows" may exist. However, it uses hard-code paths like "~/.aws/credentials" which, I believe, will break on Windows, hence there will be a need to handle OS specific file system as I mentioned

[rtyler]

Some of this will go away with apache/arrow-rs#1601 fwiw, right now there's kind of a hodge-podge of configuration possibilities between object_store and some of the rusoto crates we depend on

[mrocklin]

I don't see this as something that should be prioritized but If there is a strong support for implementing this I can have a stab

Speaking from a Dask perspective I'd certainly like to throw weight behind this. We certainly find that people commonly use .aws directories in their home directories. The ROI on looking for and parsing those files is hopefully fairly high. It's a common practice. With regards to Windows machines, I'm not sure what the convention there is, but I suspect that there is a fairly similar one. Hopefully addressing that convention as well is an easy switch.

In the meantime, can I ask what mechanisms are available to specify AWS credentials? Is it just environment variables? Is there something people can do to specify these programmatically in the meantime?

[wjones127]

In the meantime, can I ask what mechanisms are available to specify AWS credentials?

Environment variables or passing to storage_options parameter:

>>> storage_options = {"AWS_ACCESS_KEY_ID": "THE_AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY":"THE_AWS_SECRET_ACCESS_KEY"}
>>> dt = DeltaTable("../rust/tests/data/delta-0.2.0", storage_options=storage_options)

https://delta-io.github.io/delta-rs/usage/loading-table/

[jrbourbeau]

Thanks for the example showing AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY being used @wjones127. That was useful

The system I'm running on uses AWS_CONTAINER_CREDENTIALS_FULL_URI for managing AWS credentials (https://docs.aws.amazon.com/sdkref/latest/guide/feature-container-credentials.html). I tried passing AWS_CONTAINER_CREDENTIALS_FULL_URI via storage_options but unfortunately it didn't work (I access denied errors when trying to write a deltatable). It'd be great if other authentication options like AWS_CONTAINER_CREDENTIALS_FULL_URI were supported.

[mrocklin]

FWIW I think that the system referred to above is used by systems that are on AWS machines that have an IAM role attached and use that rather than AWS secret keys in environment variables. Systems know to go and read a local available endpoint to get access tokens.