Expand description
What this machine’s benches measured, keyed by what they were about.
App-level rather than per-project: a bench runs one model on this
machine’s own material, and the corpus is the machine’s too. The key
is (role, file, tier, corpus sha) - a result for a different corpus
is not a comparison, and overwriting the row for the same corpus is
what keeps a re-run from stacking a second copy of the same number.
Re-exports§
use redb::ReadableTable;🔒use redb::TableDefinition;🔒use crate::bench::BenchResult;🔒use super::Db;🔒
Constants§
- RESULTS 🔒
- One row per (role, file, tier, corpus), keyed by the four of them.
Functions§
- key
- The key one result is filed under.
- key_
parts - The key’s own spelling, for a caller that holds the parts rather than the row - the page’s delete names them and nothing else.
- remove
- Drop one result by its key. A key with no row is not an error: the row the page asked to delete is already gone.
- results
- Every saved result, newest first.
- save
- Save one result, replacing any row for the same (role, file, tier, corpus).