Lost Update

A background job reads the store, computes, and writes back — destroying anything that arrived in between.

Why it matters in a memory layer

Consolidation writes back into the store it read. The output is not a cache of the input, it replaces it: duplicates are merged away, losers retired, confidences moved. So there is no authoritative copy to re-derive from, and a write lost here is lost for good — unlike an index, which can always be rebuilt from a corpus that never changed.

On this course's corpus, store.replace(consolidate(store.all())) destroys 33 memories summed over every position a one-turn job could occupy. The worst single turn takes all four memories announcing the job change: the job that exists to keep the store correct deletes the correction. Scheduling consolidation more often — the fix from the previous lesson — multiplies the exposure rather than reducing it.

Nothing errors, and the store ends with exactly the number of memories the job expected.

Connections

Taught in: Background Job Mechanics

Do not confuse with: Snapshot Isolation