ClickHouse's Merge Join Skips Whole Runs: A Fused Compare-and-Run-Search Call and an Index Map
ClickHouse’s merge join advances two sorted cursors in lockstep and uses a fused compare-and-run-search to skip whole equal-key runs.
Clickhouse
ClickHouse's Hash Join Keeps the Right Table Columnar: One Tagged 64-bit Word per Key
The ClickHouse hash join keeps the right table as blocks and stores one tagged 64-bit row word per hash-table cell.
ClickHouse Caches Remote Objects as 32 MiB Segments: The Size Lives in the Filename
ClickHouse’s filesystem cache stores remote objects as size-encoded local segment files ordered by an in-memory LRU.
ClickHouse Encrypts Files at Rest with a 64-Byte Header: AES-CTR Counts from the Plaintext Offset
ClickHouse’s encrypted disk prepends a self-describing 64-byte header and AES-CTRs the payload beneath it.
ClickHouse Publishes a MergeTree Part with One rename(2): The fsync Order Makes It Durable
A MergeTree part is written to a temp directory through a stack of write buffers and published by one atomic directory rename.