Is there any performance difference between ThreadLocal<SplittableRandom> and ThreadLocalRandom?
The place where we rely on randomness end up being fundamental to the correctness of the algorithms. Allowing the seed to be set, at least by end users, is actually a risk to correctness in production. We had someone trying to create deterministic output from a quantiles sketch, for instance, which ends up breaking any sort of guarantees we can make about the error properties! Aside from debugging, where I'm willing to say it's ok to hack in a fixed-seed RNG, I'd generally vote for making it harder to break sketches results in production systems (and to make us be less lazy with tests of stochastic data structures). Which means I'd find it harder to support ThreadLocal<SplittableRandom> if there's even a fairly modest performance impact. At the same time, my intensity of support in this case isn't very high. jon On Thu, Aug 22, 2019, 3:24 PM Roman Leventov <[email protected]> wrote: > ThreadLocal<SplittableRandom> may be used to avoid locks and retain > determinism in tests. > > On Fri, 23 Aug 2019, 00:49 Jon Malkin, <[email protected]> wrote: > > > We got a PR to change reservoir sampling from using java.util.Random to > > java.util.ThreadLocalRandom, saying it sped up a multithreaded app > > significantly. In a scenario with multiple threads writing to independent > > reservoirs, ThreadLocalRandom really should avoid a lot of locks, so > that's > > fine. > > > > The only drawback I see is that it makes deterministic, reproducible > > testing impossible since you can't set the seed for ThreadLocalRanom. > > That's ultimately a good thing -- it'd be a terrible design to have > > separate reservoirs initialized with the same seed! And our tests already > > don't pick a seed. > > > > I'm inclined to approve the PR, and even to go farther and recommend that > > we (eventually) move to TLRandom more generally. But I first wanted to > see > > if there's some unknown drawback of which I am not currently aware? > > > > jon > > >
