Roadmap

What ships today is a deterministic rules engine. That is a deliberate choice, not a stepping stone we are embarrassed about — and here is where the analysis goes from here.

  1. Tier 0

    Rules

    Deterministic SQL thresholds

    Built — ships today

    “This warehouse is idle 62% of the time, and its auto-suspend is 600s.”

    Measured thresholds evaluated first-match-wins. Every verdict is reproducible by hand from the numbers shown beside it, which is exactly why it is the tier that ships first: a cost recommendation you cannot check is one nobody acts on.

  2. Tier 1

    Anomaly detection

    Statistical baselines with seasonal decomposition

    1–2 weeksDesigned, not built

    “This spend spike is abnormal, and here is what drove it.”

    Decompose credits into level, weekly seasonality and residual; flag on the residual rather than the raw series, so Monday morning is not an anomaly every week. Then attribute the delta by query pattern, user, tag and role against a trailing same-hour median — the top three become the explanation.

  3. Tier 2

    Scaling-law regression

    Supervised prediction of runtime against warehouse size

    3–4 weeksDesigned, not built

    “This query will cost X and take Y seconds on a MEDIUM.”

    The fixed downsize threshold is a single pair of numbers standing in for a question that varies per query: how well does this particular work parallelise? This estimates that per pattern, learned from resize events the account has already performed, with partition count as a hard ceiling on achievable parallelism.

  4. Tier 3

    Counterfactual simulation

    Discrete-event replay of real arrivals against a candidate config

    4–6 weeksDesigned, not built

    “Resize to MEDIUM: 22% cheaper, p95 latency up 1.4s.”

    Replay weeks of actual query arrivals against a candidate size, cluster count and auto-suspend — modelling queueing, cluster spawn, resume, and the 60-second billing floor. Enumerate the configurations and show the cost/latency frontier, so right-sizing becomes a choice with both axes visible rather than a single recommendation.

  5. Tier 4

    Sequential control

    Contextual bandits

    MonthsDeliberately deferred

    “I adjusted your warehouses 400 times this week.”

    Continuous automatic tuning against a reward that prices both credits and SLA violations. Deferred on purpose, and if it is ever built it ships in shadow mode first — proposing, being scored by the simulator, and applying nothing — for as long as it takes to earn the permission.

Why rules first, and mostly rules

Most of what this product decides is not a judgement call. Whether a query says SELECT *, whether a filter column is wrapped in a function, whether the same pattern is running on two warehouses — these are facts about text and measurement. More data will never make a rule better at them, and a model would only make the answer harder to check.

There is one place a fixed threshold genuinely cannot capture the question, and that is right-sizing: “would this workload be fine on a smaller warehouse” depends on how well the work parallelises, which varies per query. That is what Tier 2 is for, and it is the tier we would build next.