S3FileSystem.checksum always returns cached value

Hi,

Encountered this unexpected behavior when using the checksum method to monitor changes in a file in a long-running app.

Description

S3FileSystem.checksum returns the same value for an object, even after the object has changed

Reproduce steps

1. Create object s3://MY_BUCKET/MY_OBJECT on s3
2. Run

import s3fs
fs=s3fs.S3FileSystem()
fs.checksum("s3://MY_BUCKET/foo")

3. Overwrite the object with new contents (not using fs, but using e.g. the aws cli)
4. In the same Python process as above, run fs.checksum("s3://MY_BUCKET/MY_OBJECT")

The results from 2 and 4 are identical, even though the file has changed.

Cause

It seems that checksum uses the default implementation of fsspec.AbstractFileSystem, which in turn creates the checksum from the output of fs.ls(detail=True). Since in the S3FileSystem implementation this gets passed a default argument refresh=False, the checksum is always computed on cached results.

Proposed solution

checksum should be implemented in S3FileSystem, in such a way that it always checks the actual state of the object, or at least provides the option to do so.

Start at S3FileSystem.checksum and compare it with fsspec.AbstractFileSystem's default checksum implementation, then trace fs.ls(detail=True) and its refresh=False default. Reproduce the external overwrite in one Python process; done means a later checksum can reflect the changed object, or an explicit refresh option is documented and tested.