Source
Mohan et al. (2021), LIT-059 — https://www.usenix.org/conference/fast21/presentation/mohan.
The source argues against this practice
This is the strongest case in the cluster and it goes the other way. LIT-059 opens on the observation that checkpointing “had been epoch-granular and hand-tuned”, and both halves are named as the problem:
- Epoch granularity is too coarse. CheckFreq’s contribution is to make checkpointing iteration-granular, because an epoch on a large corpus is hours and a failure inside one loses all of it.
- A hand-tuned interval is a guess. The frequency is derived from online profiling and adapted at runtime against an overhead bound.
“Every N epochs, N ~ sqrt(total_epochs)” is a hand-tuned interval at epoch granularity. It is precisely the practice the cited paper exists to replace, and no reading of LIT-059 supports it.
Nor is the formula the paper’s, or anyone’s that the record can point to. Where it came from is not recoverable from the document.
What this needs
Not a body. Either a source that argues for a closed-form epoch interval, or
retirement — and if it is retired, the successor is the mechanism in
SOTA-054’s discussion rather than a different constant. Left Active and
flagged rather than retired here, because changing a practice’s status is a
claim about the world and belongs in its own contribution, not in a pass
that was supposed to be writing bodies.
This is the finding the worklist in #107 anticipated: a practice whose source does not support it, discovered by trying to write down why it holds.