A dispatch review uses three structures with different guarantees. An alias table samples a static weighted route list, MinHash bands retrieve possible near-duplicate incidents before an exact Jaccard check, and a run-length bitmap combines consecutive alert IDs without expanding them. The alias table has exact table probabilities up to floating-point arithmetic; MinHash candidate recall is approximate; bitmap set operations are exact. A route-weight edit needs a rebuilt alias table, while an incident replacement needs old bucket memberships removed before new ones are inserted.
Make this comfortable
Project: audit weighted draws, incident candidates, and alert spans
Acceptance trace
For weights 19, 47, and 29, inspect fixed unit draws 0.03, 0.31, 0.69, and 0.94. An incident query with four dock tokens returns exact overlap 1.0 for I-19 and 0.8 for I-47 among collided candidates. North and east alert runs intersect at [23,31) and [43,47), and their union contains fifty IDs. Compare route draw frequencies over many fixed unit points to the normalized weights, but do not treat a short sample as proof of distribution quality.
Failure and cost review
Reject empty or all-zero route weights and nonfinite totals; rebuild the table when any route changes. Generate token sets and compare every reported MinHash candidate score with direct intersection-over-union, while acknowledging that a non-colliding true match can be missed. Generate random run intervals and compare union, intersection, difference, membership, and cardinality with small direct sets. Include adjacent runs, exact right boundaries, and empty intersections. Check that published records never present candidate retrieval as an exhaustive similarity search.
Common Mistakes
- Do not use old aliases after weight changes.
- Do not claim that a missing MinHash collision proves no match.
- Do not return an unverified similarity estimate as an exact score.
- Do not count half-open run endpoints as members.
Connected lessons
data structures
projects