There are some strange differences in the test results without changing any code, e.g. org.apache.pdfbox.pdmodel.font.PDFontTest.testPDFox5048()
It looks like we can't rely on InputStream.available when reading a pdf from a stream (based on an URL) using RandomAccessReadBuffer. According to the javadoc ofInputStream.available the behaviour can be different for different implementations. In our case it returns 0 and PDFBox doesn't read any more data which leads to an incomplete file.