[silicium369-spec]

Summary

This pull request integrates Silicium, a vectorized columnar analytical database engine engineered in Rust, into the ClickBench suite, along with official 100M rows benchmark results measured on AWS c6a.4xlarge.

Architectural Highlights

• PostgreSQL v3 Wire Protocol: Native compatibility with standard psql (psql -h 127.0.0.1 -p 5433 -U postgres -d clickbench -c "$query").
• Full 105-Column Schema: Full schema coverage for all 105 columns declared in create.sql, including SELECT * projection queries (Q23, Q24).
• Decoupled Binary Distribution: In accordance with ClickBench standards for proprietary engines (pattern Umbra / ElosDB / CedarDB), no ELF binaries are committed into the Git tree. Evaluation binaries are downloaded and checksum-verified at cold start during install from the dedicated public release repository (silicium369-spec/silicium-clickbench).
• Standard Harness Interface: Conforms strictly to all ClickBench contracts (benchmark.sh, check, data-size, install, load, queries.sql, query, start, stop, template.json).

Benchmark Results (AWS c6a.4xlarge, 100M Rows, 105 Columns)

• Machine: AWS c6a.4xlarge (16 vCPUs, 32 GiB RAM, AMD EPYC 7R13 Milan, gp3 EBS).
• Cold Run (Sum of 43 Queries): 2.0581 s
• Hot Run (Sum of 43 Queries): 2.0236 s
• Data Footprint: 17,024,446,464 bytes (~15.8 GiB)
• Load Time: 286 seconds
• Results File: silicium/results/20261006/c6a.4xlarge.json

Verification

All result files have been validated with ClickBench official validate-results.py without errors.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Preview of the benchmark page with these results

[alexey-milovidov]

Let's check every item here: #2354 (comment)

[alexey-milovidov]

I went through every item from the previous review (#2354 (comment)) against this resubmission. This is a static inspection only (readelf, strings, byte searches; nothing was executed) of the scripts and of the current release tarball silicium-clickbench-x86_64-v2.0.0.tar.gz.

1. Query-specific dispatch — partially addressed, still present

Progress: silicium_runner now links sqlparser-0.54.0 and regex-1.13.1, has cb_kernels/src/sql/planner.rs, universal_aggregator.rs, fused_filter.rs, and the numeric literals (435090932899640449, 3594120000172545465, 2868770270353813622) are no longer embedded.

Still present:

• Per-query modules: cb_kernels/src/queries/q00_to_q09.rs, q10_to_q19.rs, q20_to_q29.rs.
• The literal google and the exact Q29 pattern ^https?://(?:www\.)?([^/]+)/.*$ are embedded as constants in the binary. A general engine gets those from the query text.
• Next to Only basic SELECT supported in query body, the planner contains a hardcoded list of exactly the ~20 columns used by the benchmark queries (AdvEngineID, ResolutionWidth, UserID, SearchPhrase, CounterID, URLHash, WindowClientWidth, ...).
• CLI help: "Query index to run (0..=42)", "Run all 43 queries".
• Hardware profiles aws_c6a_4xlarge / aws_c7g_4xlarge with calibrated_knobs.json and a "pareto-point" option, while the result says "tuned": "no".
• Internal paths like benchmarks/03_sanctuaires_internes/cortex_clickbench_v2/....

No evidence was provided that the engine handles variations of the queries.

2. All 105 columns — addressed

silicium_ingest now references all 105 columns from create.sql.

3. Harness — still broken

• The SHA-256 in install (72b0fd3f…) does not match the published tarball (50a692ed…); the asset was re-uploaded on 2026-10-10, after the results dated 2026-10-06. install fails.
• query prints no timing to stderr, which is the harness contract. benchmark-common.sh would record null for every query.
• Hence the results file was not produced by the standard harness. Also: the third run equals the second exactly in all 43 of 43 queries; times like 22 µs (Q1) and 30 µs (Q7) are impossible through a psql round-trip; the "cold" sum is nearly identical to the hot one (2.058 s vs 2.024 s), e.g. Q5 COUNT(DISTINCT UserID) in 0.116 s cold.
• benchmark.sh sets an empty BENCH_DOWNLOAD_SCRIPT (the dataset is never downloaded), sets BENCH_RESTARTABLE=no for a server-based system (cold runs aren't cold), and has an "autonomous" fallback mode.
• install, start, query fall back to development paths like ../../target/release/cb_runner; install installs gcc to probe the CPU; load modifies the system THP setting.
• Fixed: check now fails properly, the 5 GB "minimum" is gone, a results file is included.

4. Auditability — unchanged

Still a closed binary under a noncommercial license ("Patent Pending"), with no public source code or product documentation.

Conclusion

The requirements for a resubmission are the same as before: no query-specific kernels or embedded query literals; arbitrary SQL over the full table, with evidence on modified queries; results produced by benchmark-common.sh (timing on stderr, the real download script, proper restarts); a working install. Closing.