Measure bytes fetched per useful training byte

Record remote bytes transferred, useful in-scope payload bytes, request count, and the observation window. Keep byte representations consistent: comparing compressed network bytes with uncompressed tensor bytes can produce a ratio that measures compression as well as unwanted reads. Metadata, retries, duplicated requests, and speculative prefetch may belong in the transferred total depending on the question. State that choice. An average over one epoch can hide short bursts where each random sample opens a different object. Collect enough identity and range information to attribute the extra traffic without exposing training data contents in diagnostics.

A range-capable object API can retrieve a specified byte interval, but the file and compression format must support meaningful selective decoding at that interval. Amazon S 3 GetObject documents byte ranges and version selection; it does not automatically build a record index inside an arbitrary compressed archive. Whole-stream compression may require reading earlier data before a desired record can be decoded. Indexed blocks or columnar layouts can reduce unnecessary reads under compatible access patterns. Verify what the reader actually requests and decodes, rather than infer one-record access cost from an API's support for ranges.

Apache Parquet's format separates file metadata and column chunks, with row groups defining groups of rows. Readers can use metadata to locate relevant chunks rather than fetch every column, but actual pruning depends on the query, indexes, compression, and implementation. A training pipeline that needs only text and labels should not assume a reader automatically avoids unrelated image bytes. Likewise, a random individual-record pattern can defeat a layout designed for sequential scans. Measure the useful columns and rows obtained per physical read and preserve the file-version identity used to interpret offsets.

Larger sequential reads can amortize request latency while overfetching unused data; smaller reads can reduce bytes while multiplying operations and overhead. Coalescing nearby requests or caching blocks may improve both for a suitable workload. Batch-aware fetching can exploit known locality without changing the declared sample order if results are reassembled correctly. If changing shuffle or shard order is the optimization, record and validate that semantic change. A fast stream dominated by one source block may be statistically different from the intended global traversal even when every fetched byte is eventually useful.

Compare cold and warm behavior, and record whether a hit returns an immutable block from the intended dataset snapshot. Reusing a stale block after an object is replaced can make a faster pipeline incorrect. Include request failures and retries in operational accounting, and distinguish fewer remote bytes from fewer decoded records. A decompression cache may save CPU without reducing the first remote transfer, while a remote block cache can reduce subsequent network traffic. The relevant optimization depends on which resource currently limits delivered batches and which reuse identities are trustworthy.

The cache lab isolates one locality effect. For each separate view, Reset, set both capacity C and request stride s=4, predict initial misses, then inspect. The sixteen requests cycle through four addresses. Capacity four gives four misses and twelve hits; capacity two misses all sixteen. This exact empty LRU trace does not know object sizes, byte ranges, compression, or training samples. If each abstract miss required a remote fetch, the counts would indicate different fetch demand, but actual byte amplification still requires the size and useful payload of each real read.