Answers are grounded in this paper and cite their sources.
What does a managed distributed framework cost, and what does it buy, at the scale a small team actually runs? This study pre-registered that question against an existing record: the CPU training baselines of an earlier study by the same author (pytorch-depth-lab), which measured a single process and hand-rolled PyTorch DDP on the same workload family. The design was frozen in a pre-registration committed before any run (ResNet-18 on a stratified 4,000-image EuroSAT subset, 3 epochs, effective batch 128, seeds 42, 43, 44, loader workers 0 in every arm), with three arms (single process, hand-rolled DDP at world size 2, Ray Train at world size 2), a serving comparison (a plain in-process loop against Ray Serve at an equal total thread budget of 4), and three hypotheses with frozen verdict rules. Before the first measured run, the frozen machine of record failed completely: gRPC, which Ray's control plane requires, does not function in that sandbox under any poll strategy, a failure established with a minimal grpcio probe and preserved in the repository. A dated amendment moved every arm to a successor machine (8 cores, 14 GB RAM) before any measured run, so every within-study comparison stands; the baseline study's absolute numbers became context only. Results: H1 CONFIRMED, Ray Train matches hand-rolled DDP on scaling (median throughput 50.7 samples/s single process, 57.8 DDP, 55.1 Ray Train; scaling efficiencies 0.570 and 0.543, an absolute difference of 0.027 inside the frozen plus or minus 0.10 band; the framework's training tax is about 5 percent of throughput, and both distributed arms sit far from linear scaling, 0.54 to 0.57). H2 CONFIRMED, accuracy parity holds (mean final test accuracy 0.7180 single, 0.6923 DDP at minus 2.57 percentage points, 0.7077 Ray Train at minus 1.04 points, both inside the frozen plus or minus 5.0 point band that was set knowing the baseline study's DDP arm once diverged by 9.59 points). H3 INCONCLUSIVE by design: at batch 128 Ray Serve served 218.5 images/s against the plain loop's 200.4 (ratio 1.090, p50 latency 1.067 s against 1.135 s), ahead but below the registered 1.25 confirmation line and above the 1.00 refutation line. The claims are deliberately narrow: one machine, one small model, world size 2, CPU only. Within those bounds, at this scale the framework is chosen for its ergonomics and headroom, not for speed, because the hand-rolled version is not meaningfully faster.
No takes yet. Share an insight, caveat, or question.
Jesun Ahmad Ushno (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: