ClickHouse struggles with complex queries
Product manager at Firebolt on on scaling challenges and ACID compliance in OLAP databases
The key implication is that ClickHouse wins when a query can stay narrow and scan a few columns fast, but it gets harder to use as a general purpose analytics layer once teams start stacking many joins and multi step logic. That is why ClickHouse shines in logs, events, and dashboards, while vendors like Firebolt, StarRocks, Doris, and SingleStore focus so heavily on optimizer, caching, and HTAP features that reduce the manual work needed for deeper analytical workflows.
-
In practice, ClickHouse often performs best when data is denormalized ahead of time, with materialized views and careful schema design doing work before the query runs. The interview describes a power user workflow where teams tune engines, projections, merges, and views up front, then hit friction when they later want more ad hoc joins and deeper query trees.
-
That is the gap newer OLAP systems try to close. Firebolt adds query rewriting and subquery caching on top of a ClickHouse like core. StarRocks and Doris emphasize cost based optimizers and automatic materialized view refresh. SingleStore pitches one engine for transactional rows and analytical columns, so teams do less shuffling between Postgres and a separate analytics store.
-
The broader market signal is that fast scan speed alone is no longer enough. ClickHouse has grown quickly on open source adoption and real time analytics demand, but much of the next buying decision happens on day two issues, meaning whether a smaller data team can keep performance predictable as workloads expand from simple event slicing into mixed transactional and analytical use cases.
Going forward, the center of competition will be the optimizer layer rather than the raw engine alone. As more companies want one system for dashboards, customer facing analytics, and near transactional updates, products that can hide join complexity, isolate noisy workloads, and scale compute only when queries arrive will keep taking share from databases that still expect users to hand tune the hard parts.