During initial sync, there is a race between when we create our indexes and when we insert documents with multiple arrays. Consider the following order of events (possible in find_dedup.js):

1. Secondary starts fetching operations from primary for its oplog buffer
2. Primary creates a compound index on two fields a and b
3. Primary drops the collection
4. Primary inserts document

{a: [1,2,3], b: [4,5,6]}

5. Secondary copies the collection, including document

{a: [1,2,3], b: [4,5,6]}

, but not including the compound index on fields a and b.
6. Secondary starts to apply its oplog
7. Secondary creates a compound index on fields a and b
8. Secondary gets a "cannot index parallel arrays" error.

Later on the secondary will drop this collection and reinsert

{a: [1,2,3], b: [4,5,6]}

and get to a consistent state if it is allowed to continue.