Nano Banana 2 is designed for steady-state throughput rather than absolute minimum latency. The internal buffering and batching mechanisms that make it efficient for high-volume workloads introduce a small amount of additional latency per record compared to a direct function call. In practice, the added latency is on the order of single-digit milliseconds for most configurations, which is acceptable for the batch and streaming use cases the library targets. If your application requires sub-millisecond response times for individual records—such as in a low-latency trading system—Nano Banana 2 is not the right fit for that workload.
Throughput is bounded by the slowest stage in the pipeline. If one stage takes significantly longer to process a record than others, it will become the bottleneck regardless of how much concurrency you configure for the rest of the pipeline. Nano Banana 2 exposes per-stage latency and throughput metrics that make it straightforward to identify where the bottleneck is. Common solutions include increasing concurrency for the slow stage, splitting it into parallel sub-stages, or optimizing the underlying transformation logic. The library handles all the coordination for parallel stage execution, so enabling additional concurrency is a configuration change rather than a code change.
The library does not currently support distributed execution across multiple hosts. All pipeline stages run within a single process on a single machine. If your workload exceeds what a single node can handle, you will need to run multiple instances of the pipeline in parallel and partition your input data across them externally. Coordination across instances—for example, ensuring that related records are processed by the same instance—is the application’s responsibility. A future version of the library may introduce native support for distributed coordination, but as of the current release, that capability is out of scope.