mirror of
https://github.com/scylladb/scylladb.git
synced 2026-06-04 14:03:06 +00:00
05f36422d8
Motivation: The prepared statement cache is memory-bounded. Every byte saved per prepared statement means more statements can fit in the cache before evictions occur. Cache evictions force re-preparation of statements, which is expensive (schema lookups, type checking, optimization). Workloads with many distinct prepared statements (multi-tenant, ORM-generated queries) benefit directly from smaller per-statement memory footprint. Nature of change: The _column_operations vector in modification_statement holds polymorphic operation objects (setters, adders, deleters, etc.) via seastar::shared_ptr, but these objects are never shared — each is created during prepare(), stored in the vector, and only accessed by reference during execution. Replace shared_ptr<operation> with unique_ptr<operation> throughout: - operation::raw_update::prepare() return type - operation::raw_deletion::prepare() return type - modification_statement::add_operation() parameter - modification_statement::_column_operations vector element type - All make_shared<T>(...) → std::make_unique<T>(...) Memory savings: seastar::make_shared wraps each object in shared_ptr_count_for<T>, which prepends a vtable pointer (8 bytes) + refcount (8 bytes) = 16 bytes overhead per operation. A typical INSERT/UPDATE has 3-5 column operations, saving 48-80 bytes per prepared statement. This allows statements to fit in the cache to fit in the cache for a typical workload (assuming 2KB average prepared statement size). unique_ptr also makes ownership semantics explicit in the type system. Performance: perf_cql_parser shows no regression (110K tps). Signed-off-by: Yaniv Kaul <yaniv.kaul@scylladb.com> AI-assisted: Yes Backport: no, improvement Closes scylladb/scylladb#30010