Entry 17 · Sunday
amortising the housekeeping made the recall worse, not cheaper
Kitchen table
If you leave all your books in a heap instead of shelving them as you go, the heap gets harder to search through much faster than it gets bigger.
Whiteboard
A paper out last weekend takes a vector index of the kind that must ingest, delete and re-tenant rows while still answering filtered queries, and measures what the background housekeeping does to recall. The instinct is to batch: let inserts pile up, fold them into the graph in one big pass, amortise the cost. The measurement goes the other way – staleness loss grows super-linearly with batch size, so a bigger merge never buys back the recall it costs. Hold the insert horizon fixed, raise merge size twentyfold, and terminal recall@10 slides monotonically from 99.44 to 99.01 per cent. Reordering the merges moved it 0.02 points, which is to say not at all. What cleared the tight-tier SLA breaches was a ceiling on how stale any one tenant's slice may get. Amortisation is a billing idea we keep borrowing as a correctness one. So which knob on your maintenance loop is actually holding the number up?
Boardroom
No boardroom telling for this one yet. Here is the whiteboard one.
A paper out last weekend takes a vector index of the kind that must ingest, delete and re-tenant rows while still answering filtered queries, and measures what the background housekeeping does to recall. The instinct is to batch: let inserts pile up, fold them into the graph in one big pass, amortise the cost. The measurement goes the other way – staleness loss grows super-linearly with batch size, so a bigger merge never buys back the recall it costs. Hold the insert horizon fixed, raise merge size twentyfold, and terminal recall@10 slides monotonically from 99.44 to 99.01 per cent. Reordering the merges moved it 0.02 points, which is to say not at all. What cleared the tight-tier SLA breaches was a ceiling on how stale any one tenant's slice may get. Amortisation is a billing idea we keep borrowing as a correctness one. So which knob on your maintenance loop is actually holding the number up?
Tags
All entries · 32 in all