Having studied hundreds of Holocaust testimonies (though not only them of course), I can say that they’re both extremely valuable (the detailed Holocaust history could not have been written without them, not even with the surviving German documents) and extremely fragile. Witnesses (regardless of whether victims, perpetrators or bystanders) would contradict themselves on numerous details from testimony to testimony, and the contradictions between different witnesses are sometimes barely reconcilable. Witnesses would incorporate various details that they learned only later, they would embellish, exaggerate, repeat rumors and their own interpretations as facts. Dr. Deborah Lipstadt described the problem as follows:

Lots of survivors who arrived at Auschwitz will tell you they were examined by Mengele. Then you ask them the date of their arrival, and you say, ‘Well, Mengele wasn’t in Auschwitz yet at that point.’ There were lots of doctors… they all become Mengele.”

Indeed, the identification errors were especially notorious, as the Israeli Demjanjuk trial fiasco has shown (where Treblinka survivors were absolutely sure that he was Ivan the Terrible of Treblinka, though later he turned out to be Ivan the Less Terrible of Sobibor).

My conclusion from the studies is that the general narrative is usually relatively reliable, but when it comes to the details (dates, numbers, sizes, …), the more corroboration there is (from other witnesses, from documents, etc.), the better. And of course the earlier a testimony is, the more reliable it usually is. And: almost any testimony will contain errors, so finding errors doesn’t by itself necessarily impeach a testimony.