Two percent of pairs miss a one-hour match bound and the business asks for a 24-hour bound. How do you decide?

## The shape of the decision This is not a tuning question. Cost and benefit here scale differently, and recognising that is most of the answer: - **Cost is linear in the bound.** Held records are arrival rate × holding time on each side. Going from one hour to 24 multiplies the retained set, the memory behind it, the volume a saved picture of the job carries and the time a restart needs to reload it, all by roughly 24. - **Benefit follows a tail.** The extra pairs recovered are exactly those whose two moments lie between one hour and 24 hours apart. That distribution is usually heavily front-loaded, so most of the missing 2% is typically recovered within a few extra hours and the last sliver costs the rest of the day's memory. So the request as phrased asks for 24 times the cost without anyone having looked at the curve that decides the benefit. ## Measure this before agreeing to anything One measurement settles most of it: **among the pairs currently missed, what is the distribution of the gap between the two records' moments?** Sample the misses over a representative period, bucket them by that gap, and read off the cumulative recovery at two hours, three, six, twelve, twenty-four. A three-hour bound that recovers 1.8 of the 2 points is a different conversation from a twelve-hour bound that recovers 1.2. Two confounders to rule out first, because either makes the whole exercise meaningless: - **The gap is not the same thing as lateness.** The gap is between the two matched records' own moments. How long a record takes to reach the job after its moment is a different quantity belonging to a neighbouring subject, and widening a match bound is the wrong remedy for it. - **Check what the two moments actually are.** If one side's moment is assigned from a field in the payload and the other's from arrival, the pairs are being compared on incompatible clocks and no bound fixes that. What a moment is and which one to use is a neighbouring subject; confirming that both sides use the same one is part of this decision. ## The menu, priced | Move | What it costs | What it recovers | |---|---|---| | Widen the bound to 24 hours | ~24× held records on both sides, the same multiple on a saved picture and on restart time | every remaining pair, most of them buyable far more cheaply | | Widen to the knee of the curve | the multiple you actually chose, typically 2–4× | the bulk of the missing 2%, as read off the measurement | | Make the bound asymmetric | close to nothing; one side is held only briefly when it can never precede the other | no matches lost at all — usually the cheapest real win | | Project held records to the needed fields | discipline | no matches, but it makes any bound cheaper per unit of time | | Reconcile the residue in a periodic pass over stored history | a second code path, and a result that lands hours later | all remaining pairs, at a latency the low-latency job was never going to offer | | Accept the residue as a stated error budget | a documented and counted wrongness | nothing, and it is frequently the right answer | The fifth row is the move that separates a principal answer from a senior one: **a low-latency match and a long-tail completeness guarantee are two different products, and making one system deliver both is what forces the 24-hour bound.** Let the running job serve the 98% in seconds, and let a periodic pass over stored history close the rest. Neither has to hold a day of both inputs in memory. ## Writing down the wrongness If the residue is accepted, it stops being a bug and becomes a stated property, which requires three things: a **measured** miss rate, a **counter** so the rate is watched rather than assumed, and a **named consumer** who has agreed to it. Without the counter, an upstream change that pushes the miss from 2% to 20% is invisible; with it, the budget is a control rather than an excuse. ## What does not work - **Removing the bound so nothing can be missed.** Then nothing ever expires, and the retained set grows for the life of the job; the failure arrives later and is worse. - **Adding workers to make the longer bound free.** More machines divide the same retained total; they do not shrink it. - **Holding only one input.** Both sides are held because either can arrive first — dropping one silently loses matches in that direction. - **Deciding from intuition about how late things are.** Every number in this decision is measurable within a day, and the whole argument turns on which shape that measurement has.