Currently, the benchmark 120.uploader runs by pulling data from external servers: Wikipedia, GitHub, and PyTorch. Not only does this introduce the risk of failure (Wikipedia keeps changing its policies), but it also introduces an additional variability factor - we're not only measuring the I/O performance of Lambda and storage, but also of the external server.
We could introduce a new variant of the benchmark that serves the data from a self-deployed HTTP server on a VM with strong I/O performance or a scalable cloud storage.
Currently, the benchmark 120.uploader runs by pulling data from external servers: Wikipedia, GitHub, and PyTorch. Not only does this introduce the risk of failure (Wikipedia keeps changing its policies), but it also introduces an additional variability factor - we're not only measuring the I/O performance of Lambda and storage, but also of the external server.
We could introduce a new variant of the benchmark that serves the data from a self-deployed HTTP server on a VM with strong I/O performance or a scalable cloud storage.