Hi database folks,

If you evaluated ClickHouse a few years ago, you were likely told to avoid JOINs at all costs. The standard industry advice was to denormalize everything into massive, flat tables.

In 2020, that criticism was fair. ClickHouse had a single hash join algorithm; if your right-side table exceeded available memory, the query simply crashed.

That advice is now outdated. I recently dug into the commit history, and between 2022 and 2026, the engineering team merged over 50 pull requests that dismantled almost every legacy limitation of the join engine.

What’s changed? (Shipping by default today)

Grace Hash Join: Inactive buckets now spill to disk.…