The warehouse bill had grown faster than the product. Nobody could map spend to a workload, and the finance review kept getting deferred.
Schema conversion with codec selection, partition strategy taken from real query patterns, and checksum validation on both sides before a single reader moved. Zero downtime cutover.
Dashboards timed out during customer demos. The team had started building workarounds instead of features.
Kafka ingest partitioned for a 10x spike, MergeTree schemas with ordering keys aligned to the real queries, and TTL tiering so history stopped driving the bill.
Intermittent timeouts under load, traced to garbage collection pause spikes that only appeared at peak.
Gateway rebuilt in Go with connection pooling and backpressure, benchmarked against production traffic shapes before cutover.
Keyword search could not find answers that existed. Escalation rates climbed while the content library kept growing.
Multi stage retrieval with hybrid search, cross encoder re ranking, query rewriting and guardrails, plus embedding drift monitoring after launch.
Compute was sized for a peak nobody had measured, and there was no chart putting utilisation next to spend.
Utilisation instrumented first, then autoscaling on request load rather than CPU, with per query cost tracking so the savings stayed after we left.
Steady organic traffic on technical terms, almost no inbound. The only offer on the site was a meeting.
Crawl graph and index coverage fixed first, then capture added to the pages carrying real intent, with a free scan replacing the meeting as the first ask.
Illustrative examples while real engagements clear review. Distribution figures come from a sample 142 page scan, not a named client; cost and latency numbers are placeholders until a client signs off.
SIVARO