Different authors with the same fake email address are linked to the same author ID

[ssegura]

Hi,

Congratulations for the tool. I have come across a rather problematic (and common) case: repositories where (several) developers use the same invalid email address such as "devnull@localhost" or "none@none". In some repositories, this causes multiple users (up to 41!) to be identified with the same author_id. Any suggestions for handling such cases? Thanks.

[wschuell]

Hola Sergio;
Do they use distinct names or do you see them as different thanks to displayed login on github? Either one could rewrite the emails of each commit as the names before running gambit; or (when even names are invalid) you could use the github logins attributed to the commits via API queries (provided it s github, gitlab also has options). I implemented the github option but also running it just once per email, but if you process the emails in each commit with e.g. the commit number, you ll have the query done for each commit: github.com/wschuell/repodepo ; https://github.com/wschuell/repodepo/blob/master/repo_tools_example.py the 'GHLogins' part does the processing in this case (there is also a graphQL version, not only REST: line 1808 https://github.com/wschuell/repodepo/blob/master/repodepo/fillers/github_gql.py ). In case you would need this extra step tell me (I worked with @gotec s structure in parallel for another project so they are compatible)

[ssegura]

Hi,

Thank you for the quick response. I am cloning the repositories and extracting information about their commits. Some repositories contain too many commits for the GitHub REST APIs standard quote. The commit names are valid, but the emails are often not. I was considering to search for commits with non-valid e-mails (e.g., "devnull@localhost" is a common one) and then rewriting them before running gambit into something like name@name.com. For example:

"name": "yuriy.brun",
"email": "devnull@localhost"

would be changed to:

"name": "yuriy.brun",
"email": "yuriy.brun@yuriy.brun.com"

Would this make sense? Thanks.

[wschuell]

This could easily be done with an SQL update command (ie not yet as a feature of gambit); i ll send a working example tomorrow. how many such emails would there be? (distinct).

[ssegura]

Thanks. I will try that then.

Not sure how may email addresses yet. For now I have found many occurrences of "devnull@localhost" and "none@none"

[gotec]

Hi Sergio

Great to log in and see this active discussion. Thanks @wschuell for the quick response! I ran into the same issue in the past and "resolved" the cases manually (see option 2 below). However, I also see alternatives:

1. We provide the generic_usernames option to provide that are not matched. See also the discussion here: Added support for generic_usernames #2. I understand this is not exactly the same issue but maybe it can be adapted?

2. As you also mention, GitHub's REST API was infeasible due to the large number of commits processed. However, what worked well was to (i) apply gambit (ii) identify clusters by checking for specifically short names/emails and author_ids with many matching identities. These I then double checked manually. Typically this was significantly less work than I had feared.

3. With @wschuell code you can also run repodepo just for the clusters of identities where you'd be uncertain instead of manually checking them.

4. As you suggested, you could also change the emails and then rerun gambit a second time. This would have the benefit of matching the identities with the invalid emails could be matched with other identities if they exist in the repository.

5. If you have a suggestion on how to include that case in the set of rules used by gambit you could also clone the gambit repository and add them. As gambit is rule based the code is very linear and it's rather simple to add custom rules if required.

Without a real ground truth which we never have in these cases, it's hard to suggest which of the options would work best. Hence, depending on your use case it might be sensible to at least have a look manually to see how these identities behave and if/how they should be merged (but I understand you already did that). Based on that information you can probably identify a good option from above.

Hope that helps.
Cheers!

[ssegura]

Hi,

Thanks a lot for the detailed suggestions. I intend to analyses +1K repositories so I would prefer not to rely on manual analysis. I think I will try option 4: changing invalid emails (e.g. devnull@localhost) to [author_name]@[author_name].com before running gambit and see how it works. I will also identify clusters with too many matching identities as a sanity check.

Thanks again for the assistance and congrats for the tool.

[wschuell]

Here is the promised query that would modify the emails (you just have to fill in the problematic emails - twice here):

UPDATE commits SET committer_email=committer_name || committer_email
WHERE committer_email in ('noreply@github.com','devnull@localhost');

UPDATE commits SET author_email=author_name || author_email
WHERE author_email in ('noreply@github.com','devnull@localhost');

Maybe another way of identifying the candidates to filter, emails with different names (>1):

WITH dev_info as (select DISTINCT committer_name as name,committer_email as email from commits
UNION 
select DISTINCT author_name as name,author_email as email from commits
order by email)
SELECT email,COUNT(*) AS cnt,(GROUP_CONCAT(name,' | ')) FROM dev_info
GROUP BY email
HAVING COUNT(*) > 1
ORDER BY cnt DESC

If you need more specific help for your project don't hesitate to contact me (and/or Christoph) by email