mirror of
https://github.com/scylladb/scylladb.git
synced 2026-05-13 03:12:13 +00:00
15fa1c1c7e
Applying lazy evaluation to the BTI encoding of clustering keys was probably a bad default. The possible benefits are dubious (because it's quite likely that the laziness won't allow us to avoid that much work), but the overhead needed to implement the laziness is large and immediate. In this patch we get rid of the laziness. We rewrite lazy_comparable_bytes_from_clustering_position and lazy_comparable_bytes_from_ring_position so that they performs the key translation eagerly, all components to a single bytes_ostream in one synchronous call. perf_bti_key_translation (microbenchmark added in this series, 1 iteration is 100 translations of a clustering key with 8 cells of int32_type): ``` Before: test iterations median mad min max allocs tasks inst cycles lcb_mismatch_test.lcb_mismatch 9233 109.930us 0.000ns 109.930us 109.930us 4356.000 0.000 2615394.3 614709.6 After: test iterations median mad min max allocs tasks inst cycles lcb_mismatch_test.lcb_mismatch 50952 19.487us 0.000ns 19.487us 19.487us 198.000 0.000 603120.1 109042.9 ``` Enhancement, backport not required. Closes scylladb/scylladb#26302 * github.com:scylladb/scylladb: sstables/trie: BTI-translate the entire partition key at once sstables/trie: avoid an unnecessary allocation of std::generator in last_block_offset() sstables/trie: perform the BTI-encoding of position_in_partition eagerly types/comparable_bytes: add comparable_bytes_from_compound test/perf: add perf_bti_key_translation