We gave five metrics backends the same budget, the same week of data and the same ten queries, then ran those queries at rising concurrency. At one query at a time, the open source engines looked strong. Under concurrency, two of them stopped answering, and the rest slowed down by an order of magnitude.