Check filesystem for existing corpora on creation, import and deletion

[matthias-stemmler]

I found an issue with the check whether a corpus with a given name exists when trying to create an empty corpus, import a corpus or delete a corpus: The current implementation only checks if there is a cache entry for the given name, but not whether a corpus actually exists on disk. This can be a problem in two cases:

• When a corpus has never been added to the cache
• When a corpus has been evicted from the cache because the cache grew too large

We originally stumbled upon this in matthias-stemmler/annimate#471, where it caused an interesting case of data corruption. The sequence of events was presumably as follows:

• The user imported all corpora of ReM 1.0.
• They loaded all corpora, but due to limited memory, some got evicted from the cache.
• They tried to import all corpora of ReM 2.x.
  • Since the names are identical in both versions and I set overwrite_existing to false, this should have failed. However, since existence is checked against the cache, the import went through, at least for those corpora that had been evicted.
• Additionally, the old data weren't deleted, but instead the new data were imported on top of them.
  • This means that components that exist in ReM 1.0, but not in ReM 2.x, remained on disk. In this case it was Coverage/default_ns/, Coverage/VIRTUAL/ and Ordering/default_ns/.
• Now node ids from the ReM 1.0 components were reinterpreted in the context of ReM 2.x, causing nonsensical edges such as coverage between nodes of different documents.
• This made the subgraph method return wrong results, causing a broken export.

We could of course fix it by deleting and reimporting the corpora.

When looking through the code, I found that the same issue as for the import also applies to corpus creation and deletion. All other operations seem fine to me.

For the fix, I'm now checking whether a corpus exists on disk in addition to the cache check. I kept the latter to avoid changing the behavior in case a corpus exists only in the cache and not on disk (which can happen, for instance, when the info method is called on a nonexistent corpus).
I also noticed that the deletion logic acquires an exclusive lock on the cache entry before deleting from disk, so I figured the import should do the same in case the corpus already exists.

Thanks to @S-Oppermann for reporting the original issue and providing the data for analysis.

[codecov]

Codecov Report

Attention: Patch coverage is 60.71429% with 11 lines in your changes missing coverage. Please review.

Project coverage is 67.86%. Comparing base (c368b15) to head (67e17c9).
Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
graphannis/src/annis/db/corpusstorage.rs	60.71%	5 Missing and 6 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #313      +/-   ##
==========================================
+ Coverage   67.70%   67.86%   +0.15%     
==========================================
  Files          76       76              
  Lines       18958    18968      +10     
  Branches    18958    18968      +10     
==========================================
+ Hits        12835    12872      +37     
+ Misses       4662     4633      -29     
- Partials     1461     1463       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[thomaskrause]

Thanks for the fix and especially the thorough test cases. The change looks good to me, I would add a changelog entry later to the main branch.