Live experiment
Index advisor arena
Our survey traced index selection from 1985 heuristics to 2020s bandits. Here they run against each other: the same workload, the same storage budget, a real SQLite engine, and a stopwatch on everything — what-if calls, index builds and queries. Pick the TPC-H-like benchmark or the Louvre ticketing database designed in INFO20003.
Results
Pick a workload, then press run
The worker generates a TPC-H-like database (about 30,000 line items at size S), loads it into SQLite compiled to WebAssembly, and replays the same rounds of queries once per advisor. Before each round an advisor may build or drop indexes; every millisecond it spends recommending, building and querying is counted.
Offline tools (DROP, AutoAdmin, DB2 Advisor, CoPhy) are invoked at the start of round 2 with round 1 as their representative workload, and again after a shift — the protocol Perera et al. used for their commercial baseline. The bandit tunes every round from observed runtimes.