Path encoding bug

[pwinckles]

I recently discovered that the BagIt 1.0 specification requires that CR, LF, and % in file paths within manifest files are percent-encoded, and that there isn't a single BagIt implementation that does this correctly. Implementations either only encode CR and LF but not % or they encode nothing.

This implementation does not encode paths in the manifest, which means that it would fail to validate BagIt 1.0 bags that include file paths containing CR, LF, or %. Likewise, it would create bags that would fail BagIt 1.0 validation in the case that there are paths that naturally contain percent-encoded characters.

For example, let's say a bag contains the file data/file%0A1.txt. This file should be written to the manifest per the spec as data/file%250A1.txt. However, this implementation writes it as data/file%0A1.txt. This means, that when this implementation validates a properly constructed 1.0 bag it will look for the file data/file%250A1.txt which does not exist. Similarly, if another implementation that follows the spec attempts to validate a bag produced by this implementation, it would look for data/file\n1.txt, which does not exist.

It would seem desirable to me to move the ecosystem in the direction of properly implementing the 1.0 specification, while at the same acknowledging that there are a large number of 1.0 bags in existence that may then become invalid.

As such, it may be prudent to, when validating bags, fall back on a series of tests. You may want to first attempt to validate per the spec, and then, if a file cannot be found, attempt to locate it by either only decoding the CR and LF or leaving the path unchanged, ideally validating all of the files using the same method.

I have not examined fetch.txt implementations, but the same encoding requirements exist for paths in that file as well. This is potentially a thornier problem to address in a backward compatible way as it is unclear if the path data/file%250A1.txt is supposed to create data/file%250A1.txt (incorrect) or data/file%0A1.txt (correct).

Finally, I created a related ticket against the spec discussing this encoding problem, in particular how it breaks checksum utility compatibility.

[pwinckles]

It occurred to me that another approach to backwards compatible validation would be to only decode %0D, %0A, and %25 when decoding paths in manifests. Normally, when percent-decoding, you'd decode all encoded characters as described here. However, by only decoding these three a correct BagIt 1.0 implementation would still be able to validate most bags produced by existing implementations.

For example, if a bag contains the file test%201.txt, then an existing implementation would write it to the manifest as data/test%201.txt when it should actually be data/test%25201.txt. However, if you only decode %0D, %0A, and %25, then the paths are equivalent.

This approach does not work for files that naturally include these three strings. For example, if a bag has a file named test%251.txt. Existing implementations would write it to the manifest as data/test%251.txt when it should be data/test%25251.txt. These paths are not equivalent. The first decodes to data/test%1.txt and the second decodes to data/test%251.txt.

While not perfect, I think this approach would greatly improve validation compatibility.

[steffenfritz]

Thank you for the detailed and clear description.

I am working on this in the v0.5.0 branch.

[steffenfritz]

While not perfect, I think this approach would greatly improve validation compatibility.

I am not sure if implementations should be compatible to bags that are not valid. If there are a lot of bags out there not valid because the implementations were not following the spec, the bags should be re-created where necessary.