ocrd workspace: file paths should be made relative to self.directory
On the CLI, the user expects that paths are resolved once (on the input side) and from then onwards everything is automatically done relative to the workspace (whenever possible). So if a file path is absolute, then its workspace prefix gets removed regarding METS FLocat. Or if it is relative, then it is relative to the CWD, but gets resolved relative to the workspace.
Unfortunately, that's not what we have implemented yet, esp. if you combine with an overall -d path/to/workspace (so CWD is not equal workspace directory).
For example, in add_file, we currently have:
https://github.com/OCR-D/core/blob/299301003688c14a2824d865804f3a69480cdbbd/src/ocrd/cli/workspace.py#L221-L222
Meaning, the workspace directory gets added, with no regard to the caller's CWD.
Another example is in bulk_add, where we do:
https://github.com/OCR-D/core/blob/299301003688c14a2824d865804f3a69480cdbbd/src/ocrd/cli/workspace.py#L375-L381
Here, again, no difference is made between CWD and self.directory.
In the latter case (bulk-add), what makes things worse is that the glob pattern gets treated verbatim if empty: https://github.com/OCR-D/core/blob/299301003688c14a2824d865804f3a69480cdbbd/src/ocrd/cli/workspace.py#L333-L335
So if one tries to compensate for the above path resolving issue by passing a glob relative to the workspace, then that itself will break the METS with a verbatim asterisk path entry:
ocrd workspace -d nd1967-12-14_var1/ bulk-add -r "TEXTRACT_PAGE/(?P<page>.*)[.]xml" -G TEXTRACT_PAGE -g '{{ page }}' "TEXTRACT_PAGE/*.xml"
INFO ocrd.cli.workspace.bulk-add - [ 1/1] TEXTRACT_PAGE/*.xml
Same is true for the --mets-server-url argument: relative paths don't work (not relative to the workspace and not relative to CWD).
(Again, the user expectation is that whatever resolves at startup time from the user's CWD should be kept.)
Same is true for the
--mets-server-urlargument: relative paths don't work (not relative to the workspace and not relative to CWD).(Again, the user expectation is that whatever resolves at startup time from the user's CWD should be kept.)
This also just bit me in the Processing Server context: passing a relative path to the --mets argument of a /processor/run/... post yields a HTTP 500, caused by
File "/build/core/src/ocrd_network/processing_server.py", line 474, in validate_and_forward_job_to_network_agent
mets_server_url = self.deployer.start_uds_mets_server(ws_dir_path=ws_dir_path)
File "/build/core/src/ocrd_network/runtime_data/deployer.py", line 157, in start_uds_mets_server
pid = OcrdMetsServer.create_process(mets_server_url=str(mets_server_url), ws_dir_path=str(ws_dir_path), log_file=str(log_file))
File "/build/core/src/ocrd/mets_server.py", line 427, in create_process
raise RuntimeError(f"Mets server starting failed. See {log_file} for errors")
RuntimeError: Mets server starting failed. See /data/logs/mets_servers/Sounding_Spirit_sample_pages.log for errors
And that log in turns says:
File "/build/core/src/ocrd/cli/workspace.py", line 924, in workspace_serve_start
workspace=Workspace(ctx.resolver, directory=ctx.directory, mets_basename=ctx.mets_basename),
File "/build/core/src/ocrd/workspace.py", line 102, in __init__
mets = OcrdMets(filename=self.mets_target)
File "/build/core/src/ocrd_models/ocrd_mets.py", line 78, in __init__
super().__init__(**kwargs)
File "/build/core/src/ocrd_models/ocrd_xml_base.py", line 35, in __init__
raise Exception('File does not exist: %s' % filename)
Exception: File does not exist: /data/Sounding-Spirit-sample-pages/Sounding-Spirit-sample-pages/mets.xml
So clearly, the directory name was added twice.
The call chain was: https://github.com/OCR-D/core/blob/8ac528f948d0888653117c07d2f5fdfb29769cad/src/ocrd_network/processing_server.py#L473-L474
https://github.com/OCR-D/core/blob/8ac528f948d0888653117c07d2f5fdfb29769cad/src/ocrd_network/runtime_data/deployer.py#L145-L157
https://github.com/OCR-D/core/blob/8ac528f948d0888653117c07d2f5fdfb29769cad/src/ocrd/mets_server.py#L418-L420
Pinging @kba @MehmedGIT because this severely affects deployments. (The only workaround would be using absolute path names.)
Same is true for the
--mets-server-urlargument: relative paths don't work (not relative to the workspace and not relative to CWD).
Ideally, the internal low-level methods should still be absolute path names to avoid passing the workspace context around and keep testing of separate methods simple. I am unsure how difficult it will be to include both relative and absolute paths on the CLI level for the Processing Server and/or the Workspace methods. However, I am open to discussions to try to find a solution.