When trying to merge some documents (using the PDFMergerUtility class) I got a NullPointerException and the merge failed. I traced through to eventually discover that some objects were being overwritten when the PDFParser called document.dereferenceObjectStreams(); (line 207 of PDFParser.java).
Having multiple objects with the same object ID is a violation of the PDF specification, so how this should be dealt with is undefined. The "use the first object" mentality enabled my file to be processed and it is consistent with the other code in PDFBox. For another example of where PDFBox deals with reading in an object which already exists, you can see PDFParser (on line 541) checks to see if the object has already been read and put in the pool. If not, it adds it to the list of conflicts. Later, when resolveConflicts() is called, it overwrites the object only if it's specifically referenced in the xref table. This is a reasonable way to resolve conflicts because if the object isn't in the xref table, it is likely the wrong one.
Since we're reading from a stream of compressed data, we can not give a particular byte offset. This means we can't add these conflicts to the conflict list and try to determine if this object is legitimate or not. It's best to use the data we've already read, as using the one from the stream has been confirmed to cause problems. I've done regression testing with other files which have this problem, including the file from
PDFBOX-720 and have not seen any issues.
Unfortunately I can not provide the PDF which demonstrates this problem and solution as it contains information I'm not authorized to release.