How Shipping JPEGs Instead of Tensors Cut ML Payload Size ~17× and Network Latency by 60%

Moving image pre/post-processing into Triton reduced cross-cluster payloads from 5.17 MB float tensors to ~304 KB JPEGs and cut indexing network latency from 400–600ms to 150–200ms — without changing inference. Gzipping tensors did not help; it only burned CPU. When Trendyol’s Image Search platform migrated indexing inference to cloud GPUs, the architecture looked sensible on paper: preprocess images in the API layer and send ready-to-run tensors to NVIDIA Triton on GKE.