orneryd / NornicDB / 37209550980

89%

Build Type

push

github

Committed by orneryd

Commit Message

perf(storage,cypher): cheaper label-less scans, and WHERE / relationship forms take them (#857)

A label-less property match still cost 0.8 µs per scanned node against
Neo4j's 0.15 µs, and two common forms skipped the projected scan:
MATCH (a) WHERE a.id = $id, and relationship patterns starting from a
label-less node with properties (graphify's stale-edge delete), which
decoded every node in full.

projectedNodeDecoder decodes one scan's projected properties. It
resolves the projected names to key tokens once per database (lookupID),
skips other properties by token without a dictionary lookup, decodes
into one map it clears per node, reuses one reader, and reads the
database from the key bytes, so a node the filter rejects allocates no
map, reader or node ID. decodeNodeFiltered's filter is gone with it;
decodeNodeProjected decodes without one.

collectNodesWithStreaming's label-less projected scan now requires the
pattern's properties and the WHERE's top-level equalities on the
variable (labellessScanRequiredProperties), and the pipeline passes those
equalities when it evaluates the WHERE itself. Start-node loads pass the
pattern's properties (loadPatternNodes) instead of filtering a full
label-less load afterwards; filterNodesByProperties uses
nodePropertiesMatch, the one pattern-property test.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Coverage Stats

176 of 182 new or added lines in 12 files covered. (96.7%)

960 existing lines in 11 files now uncovered.

179142 of 206231 relevant lines covered (86.86%)

1.02 hits per line