CoverageData with no_disk does not update correctly

[lordmauve]

Describe the bug

A CoverageData constructed with no_disk=True cannot be updated with another CoverageData. The operation succeeds but records nothing:

>>> from coverage import CoverageData
>>> a = CoverageData('.foo')         
>>> a.add_arcs({'foo.py': {}})       
>>> a.measured_files()               
{'foo.py'}                           
                                     
>>> b = CoverageData(no_disk=True)   
>>> b.update(a)                      
>>> b.measured_files()               
set()                                

(This applies also for arcs etc., b remains empty).

Expected behavior

The expected behaviour for no_disk=True is the same as when no_disk is left unspecified; after an update() call we should see data in the updated CoverageData:

>>> c = CoverageData()
>>> c.update(a)       
>>> c.measured_files()
{'foo.py'}

To Reproduce

1. What version of Python are you using? 3.9.7
2. What version of coverage.py shows the problem? 6.3.1
3. The output of coverage debug sys is helpful.

It's very long due to a tremendously long PYTHONPATH, but perhaps this part is most relevant:

                sqlite3_version: 2.6.0                                                 
         sqlite3_sqlite_version: 3.36.0                                                
             sqlite3_temp_store: 0                                                     
        sqlite3_compile_options: COMPILER=gcc-8.3.0; ENABLE_FTS4; ENABLE_FTS5          
                                 ENABLE_GEOPOLY; ENABLE_JSON1; ENABLE_MATH_FUNCTIONS   
                                 ENABLE_RTREE; MAX_MMAP_SIZE=17179869184; SYSTEM_MALLOC
                                 THREADSAFE=1                                          

[nedbat]

Thanks for the interesting bug report. Just for my own curiosity, can you say more about how you are using these APIs?

[lordmauve]

We do a huge test run of our monorepo with tests parallelised across a test farm. The coverage files get rsynced back to one host where we:

1. Merge them
2. Remap paths captured in the coverage files to paths that exists on that host at that time
3. Build HTML reports
4. Build a table of the coverage statistics per component

I think if you're asking whether this could just use the CLI, I think the command line doesn't let me supply the regex-based path mapper class we use for step 2.

[lordmauve]

Oh, the reason I was reluctant to use a disk file was that a) we don't keep that file, no need for it to be on disk and b) for some reason this is happening on a read-only filesystem

[nedbat]

Hmm, I can confirm this happens, but it's a mystery (so far) to me why...

[nedbat]

This is now fixed in 84f70f6.

[nedbat]

This is now released as part of coverage 6.4.