Concurrent programs are notoriously hard to write correctly, as scheduling nondeterminism introduces subtle errors that are both hard to detect and to reproduce. The most common concurrency errors are \emph{(data) races}, which occur when memory-conflicting actions are executed concurrently. Consequently, considerable efforts are made towards efficient techniques for race detection. The most usual approach is \emph{dynamic race prediction}: given an observed, race-free trace $\sigma$ of a concurrent program, the task is to decide whether events of $\sigma$ can be correctly reordered to a trace $\sigma^*$ that witnesses a race hidden in $\sigma$.

In this work we introduce the notion of \emph{sync(hronization)-preserving races}. A sync-preserving race occurs in $\sigma$ when there is a witness $\sigma^*$ in which synchronization operations (e.g., acquisition and release of locks) appear in the same order as in $\sigma$. This is a broad definition that \emph{strictly subsumes} the famous notion of happens-before races. Our main results are as follows. First, we develop a sound and complete algorithm for predicting sync-preserving races. For moderate values of parameters like the number of threads, the algorithm runs in $\widetilde{O}(\mathcal{N})$ time and space, where $\mathcal{N}$ is the length of the trace $\sigma$. Second, we show that the problem has a $\Omega(\mathcal{N}/\log^2 \mathcal{N})$ space lower bound, and thus our algorithm is essentially \emph{time and space optimal}. Third, we show that predicting races with \emph{even just a single} reversal of two sync operations is $\mathsf{NP}$-complete and even $\mathsf{W}[1]$-hard when parameterized by the number of threads. Thus, sync-preservation characterizes \emph{exactly} the tractability boundary of race prediction, and our algorithm is nearly \emph{optimal} for the tractable side. Our experiments show that our algorithm is fast in practice, while sync-preservation characterizes races often missed by state-of-the-art methods.