Weakly consistent data stores have become popular because they enable highly available, scalable distributed applications. However, some data needs strong consistency. For applications that mix accesses to strongly and weakly consistent data, programmers must currently choose between bad performance and possible data corruption. We instead introduce a safe mixed-consistency programming model in which programmers choose the consistency level on a per-object basis. Further, they can use atomic, isolated transactions to access both strongly consistent (e.g., linearizable) data and weakly consistent (e.g., causally consistent) data within the same transaction. Compile-time checking ensures that mixing consistency levels is safe: the guarantees of each object's consistency level are enforced. Programmers avoid being locked into one consistency level; they can make an application-specific tradeoff between performance and consistency. We have implemented this programming model as part of a new system called MyriaStore. MyriaStore demonstrates that safe mixed consistency can be implemented on top of off-the-shelf data stores with their own native, distinct consistency guarantees. Our performance measurements demonstrate that significant performance improvements can be obtained for geodistributed applications that need strong consistency for some critical operations but that also need the high performance and low latency possible with causally consistent data.