It seems that usage of ** in glob can lead to incorrect listings cache state so that subsequent commands like glob or ls produce incorrect results.

I have two examples.

Missing directories

import s3fs

print(s3fs.__version__)
s3 = s3fs.S3FileSystem()
s3.glob("s3://<redacted bucket>/<redacted directory>/*/<redacted>/**/*.txt")
print(s3.ls("s3://<redacted bucket>/<redacted directory>"))

Here s3.ls misses several directories that are contained in <redacted directory>. If I comment s3.glob out, or replace ** with * (the output of glob does not change, as there is only one subdirectory level in <redacted>), no missed folders in ls output.

Duplicated directories

I tried to reproduce the previous example with a simpler directory structure, and encountered different problem.

I have a very simple structure inside my bucket:

a/
b/somefile.pdf
c/

Then I do

import s3fs

print(s3fs.__version__)
s3 = s3fs.S3FileSystem()

s3.glob("s3://<test bucket>/**/*.pdf")
print(s3.ls("s3://<test bucket>/"))

The output:

2025.9.0
['<test bucket>/a/', '<test bucket>/b', '<test bucket>/b', '<test bucket>/b/', '<test bucket>/c/']

So directory b is present three times (one with a leading /, the other two without).

If I replace ** with *, I get different output:

2025.9.0
['<test bucket>/a', '<test bucket>/b', '<test bucket>/c']

In all cases, if I initialize the filesystem with s3 = s3fs.S3FileSystem(use_listings_cache=False), the problems disappear. So I believe the problem is the interaction between ** in glob and the cache.

I am using 2025.9.0 version of s3fs (also reproduces with 2025.9.0+3.g2ccadeb), Python 3.12.9 on Mac OS.