[silicium369-spec]

Overview

This PR adds Silicium (silicium/), a native vectorized columnar analytical engine with a PostgreSQL-compatible wire server and full 105-column ClickBench ingestion and execution support across both x86_64 and aarch64 architectures.

Changes Addressed Since Previous Review (#2427)

We have thoroughly refactored the engine and harness to address every point raised in the review of #2427:

1. Full 105-Column Ingestion & Physical Storage (load / silicium_ingest):

   • All 105 columns of the ClickBench hits.parquet dataset (78 fixed-width scalar columns + 27 variable-length BYTE_ARRAY / string columns) are physically ingested, materialized on disk (data/arena_100m, ~17.02 GB columnar arena), and queryable.
   • No columns are dropped or skipped during ingestion.

2. General SQL Parser & AST Planner (sqlparser-rs):

   • Arbitrary SQL queries are parsed into an AST via sqlparser::parser::Parser (PostgreSQLDialect / GenericDialect) and compiled into a general physical execution plan (PhysicalPlan) supporting arbitrary SELECT projections, constant/arithmetic expressions (SELECT 1 + 1), derived subqueries (SELECT * FROM (SELECT ...) t), WHERE filter conjunctions/disjunctions (=, !=, <, <=, >, >=, BETWEEN, IN, LIKE / NOT LIKE), GROUP BY keys (scalar, dictionary, string, and computed expressions such as EXTRACT, DATE_TRUNC, SUBSTRING, REGEXP_REPLACE, array/string functions), COUNT / SUM / AVG / MIN / MAX / COUNT(DISTINCT ...) aggregations, ORDER BY, and LIMIT.
   • No query IDs, hardcoded query templates, or benchmark-specific dispatch paths exist in the binary.

3. Zero Pre-Computed Query Results During load:

   • load only converts hits.parquet into the 105-column columnar representation (col_*.bin / dict_*.bin) and builds generic secondary zone-map / dictionary indices. No query results or query-specific materializations are pre-computed during load.

4. Dual-Architecture Standalone Binaries (x86_64 & aarch64):

   • install automatically detects uname -m (x86_64 or aarch64), downloads the corresponding standalone release archive (silicium-clickbench-x86_64-v2.0.0.tar.gz or silicium-clickbench-aarch64-v2.0.0.tar.gz) from the public release repository (https://github.com/silicium369-spec/silicium-clickbench/releases/tag/clickbench-v2.0.0), and verifies its SHA-256 checksum before extraction.

Included Results (results/20260412/)

Validated with ./validate-results.py:

• c8g.metal-24xl.json (AWS Graviton4 aarch64, 96 vCPUs, 192 GB RAM)
• c6i.metal.json (AWS Xeon Ice Lake x86_64, 128 vCPUs, 256 GB RAM)
• c6a.4xlarge.json (AWS EPYC Milan x86_64, 16 vCPUs, 32 GB RAM)

[github-actions]

Preview of the benchmark page with these results