Running ClickHouse at scale with high-cardinality telemetry data reveals a deceptive failure pattern: ingest stays healthy while query-time costs accumulate silently. Three distinct failure modes emerge at scale — memory exhaustion during GROUP BY aggregations, part explosion from fine-grained grouping keys in materialized views, and full-text scans on unindexed high-cardinality columns. Real production numbers are shared, including a spike to 400M rows/minute that caused 160 concurrent merges and 40-50 second insert latencies. Mitigations include designing ORDER BY keys around real query patterns, adding tokenbf_v1 bloom filters for high-cardinality text lookups, coarsening timestamp granularity in materialized views, promoting hot Map keys to typed LowCardinality columns, and avoiding SELECT * on wide tables.
