I did some tests with a "best practices"-like pipeline to check if results were deterministic and found that they are not. Some posts already mention that UnifiedGenotyper is non-deterministic when using multi-threading as different seeds are used for downsampling. But I think I'm missing something if single-thread UnifiedGenotyper is deterministic, why would it chose exactly the same reads for downsampling? Wouldn't it always be non-deterministic when downsampling reads? Anyway, the difference was only of 31 variants for an exome sample.
About the VariantRecalibrator I guess the filtering is non-deterministic, but I did not found any reference to this in the forum. The difference between runs is greater in this case. After filtering I had 301 variants only non-filtered in the first run; and 1684 variants only non-filtered in the second run; the non-filtered variants in both runs were 11328.
Thanks in advance! Pablo.