Catalog entries can get skipped by catalog2registration when there are transient failures

[joelverhagen]

From PR NuGet/NuGet.Services.Metadata#150:

Suppose you have a page with:

ID   Version   Commit Timestamp
--   -------   ----------------
B    1.0.0     10:00 PM
A    1.0.0     10:00 PM
A    1.0.1     10:05 PM

Also suppose you have the cursor set to 9:55 PM. This means all of the items in the page need to be processed. You will have two batches processed by RegistrationCollector:

Batch 1: B 1.0.0
 - Commit Timestamp: 10:00 PM

Batch 2: A 1.0.0, A 1.0.1
 - Commit Timestamp: 10:00 PM

Suppose batch 1 is processed successfully. The CommitCollector will update the cursor to 10:00 PM (the commit timestamp of batch 1) then start processing batch 2.

Now suppose the processing of batch 2 fails such that neither A 1.0.0 nor A 1.0.1 are completed.

The job will crash due to the failure then start up again. Since the cursor is now at 10:00 PM, the set of items in the page considered will be:

ID   Version   Commit Timestamp
--   -------   ----------------
A    1.0.1     10:05 PM

This means A 1.0.0 is skipped. Remember that the cursor means "everything strictly after this timestamp needs to be processed".

[skofman1]

reopen for tracking