Fanning 14,000 accounts across 500 Lambdas keeps BMW's daily cost forecast near $50 a month

A fixed dollar threshold is the design everyone tries first, and the AWS post, co-written with BMW Group's Philipp Karg and Data Reply's Christopher Masurek, is specific about why it fails at this size ^[18]. Accounts grow, adopt new services and ramp workloads on purpose, and a fixed rule treats all of it as anomalous ^[13]. Set the number high enough for the largest accounts to stay quiet and the smaller ones get no coverage at all ^[13]. So each account-service pair is measured against its own learned trajectory instead of a number chosen centrally ^[14].

The authors put the cost of that choice in one sentence: a model that adapts to a trend will eventually absorb one ^[15]. The text stops mid-sentence at that point, before any filter criteria ^[22]. Getting from a forecast to an alert needs a baseline, a daily comparison that flags the days falling outside it, and filters that decide which of those days are worth an owner's attention ^[16]. Impact is defined as actual spend minus expected spend, so a positive number means overspend against the forecast ^[17].

Fourteen thousand accounts over 500 concurrent slots is 28 accounts per slot, and 20 minutes of wall clock divided by 28 leaves roughly 43 seconds per account if every slot stays busy for the whole run ^[19]. In that window each function fits Prophet on 365 days of daily history for every service the account uses, with additive seasonality ^[9]. The account list itself is produced by a preparation Lambda that writes JSON to S3 for the Distributed Map to read ^[10].

The compute figure divides the same way. At 30 runs a month, about $50 works out to roughly $1.67 per pass over the estate ^[20], or about a third of a cent per account per month ^[21]. Prophet was chosen for its simplicity and its steady performance on cost time series, not for accuracy claims ^[9].

The part I would steal outright is the module boundary. Forecasting is held in place by two formats: daily cost per account-service going in, per-day predicted values with confidence intervals coming out, so the engine can be replaced without touching the detection and alerting layers owners depend on daily ^[12]. A replacement therefore has to emit intervals, not just point predictions. Those predictions feed two consumers, a 12-month rolling forecast and the per-day expected baseline that detection reads ^[11]. Upstream of all of it sit roughly 3 billion billing rows across 500 columns a month, aggregated to one grain of daily cost per account per service ^[6].