To support appending, the Parquet data source tries to find out the max part number of part-files in the destination directory (the <id> in output file name "part-r-<id>.gz.parquet") at the beginning of the write job. In 1.3.0, this step happens on driver side before any files are written. However, in 1.4.0, this is moved to task side. Thus, for tasks scheduled later, they may see wrong max part number generated by newly written files by other finished tasks within the same job. This actually causes a race condition. In most cases, this only causes nonconsecutive IDs in output file names. But when the DataFrame contains thousands of RDD partitions, it's likely that two tasks may choose the same part number, thus one of them gets overwritten by the other.
The following Spark shell snippet can reproduce nonconsecutive part numbers:
"16" can be replaced with any integer that is greater than the default parallelism on your machine (usually it means core number, on my machine it's 8).
And here is another Spark shell snippet for reproducing overwriting:
Expected answer should be 10000, but you may see a number like 9960 due to overwriting. The actual number varies for different runs and different nodes.
Notice that the newly added ORC data source is less likely to hit this issue because it uses task ID and System.currentTimeMills() to generate the output file name. Thus, the ORC data source may hit this issue only when two tasks with the same task ID (which means they are in two concurrent jobs) are writing to the same location within the same millisecond.