Skip to content

Add Silicium engine to ClickBench benchmarks - #2354

Closed
silicium369-spec wants to merge 1 commit into
ClickHouse:mainfrom
silicium369-spec:add-silicium-engine
Closed

silicium369-spec wants to merge 1 commit into
ClickHouse:mainfrom
silicium369-spec:add-silicium-engine

Conversation

@silicium369-spec

Copy link
Copy Markdown

Summary

Add Silicium high-performance native analytical engine integration for ClickBench.

  • System: Silicium
  • Type: Proprietary evaluation binary (Decoupled runtime download via install script)
  • Architecture: Zero-copy columnar in-memory analytical engine (x86_64 AVX2/AVX-512)
  • Verification: All 13 canonical runner scripts verified against ClickBench specification.

@alexey-milovidov

Copy link
Copy Markdown
Member

Thanks for the submission. I reviewed the scripts and did a static inspection (readelf, strings; nothing was executed) of the release tarball silicium_clickbench-x86_64-v1.0.0.tar.gz. Unfortunately, the submission cannot be accepted in its current form.

1. The engine dispatches on hardcoded query text, not on general SQL

libsilicium_clickbench.so contains no SQL parser and no regex library (its only third-party crates are ahash, hashbrown, memchr, rayon, chrono, memmap2). Instead it embeds near-verbatim, lowercased fragments of the 43 ClickBench queries, for example:

  • select searchphrase from hits where searchphrase <> '' order by eventtime limit 10
  • select * from hits where url like '%google%' order by eventtime
  • where userid = 435090932899640449, counterid = 62, 2013-07-14, sum(resolutionwidth + 1), select 1, url, count(*)
  • Only 'hits' table supported
  • source paths such as src/kernels/domain_regex.rs, a dedicated kernel for the Referer-domain REGEXP_REPLACE in Q29

So each benchmark query is routed to a hand-written kernel. That is a query-specific implementation of the benchmark rather than a general-purpose database. ClickBench is meant to compare systems that run arbitrary SQL; benchmark-specific special-casing is not acceptable.

2. Only the columns used by the queries are loaded

Both silicium_ingest and the library reference about 24 column files (UserID.bin, SearchPhrase.dict_idx.bin, URLHash.bin, ...) instead of the 105 columns declared in create.sql. This means SELECT * ... LIMIT 10 (Q24) cannot return complete rows, and the reported load time and data size are for a reduced dataset.

3. Harness issues

  • No result files are included, although the description says the scripts were "verified".
  • check always exits 0, even when the probe query fails.
  • load rejects any database smaller than 5 GB, calling it a "ClickBench minimum threshold". No such rule exists.
  • template.json should describe the system accurately.

4. The binary cannot be audited

The binary is closed-source, and its license forbids inspecting it. Combined with the points above, we have no way to verify that it is a general-purpose engine.

What a resubmission would need

  • An engine that accepts arbitrary SQL (or a documented query language) over the full 105-column hits table, with no query-specific dispatch.
  • Evidence that it handles variations of the benchmark queries, such as changed constants, reordered predicates, or other columns.
  • Preferably public source code, or at least a publicly documented product.
  • Real result files produced by the standard harness.

I'm closing this PR for now. You are welcome to resubmit once these points are addressed.

@alexey-milovidov alexey-milovidov self-assigned this Oct 3, 2026
@alexey-milovidov

Copy link
Copy Markdown
Member

If you want us to run this binary blob on various instance types and collect results, we can do it for you.

@alexey-milovidov alexey-milovidov added the joke Not a database label Oct 3, 2026
@alexey-milovidov

Copy link
Copy Markdown
Member

Provenance of the submission

I could not find any public information about a database or analytical engine called "Silicium": no website, documentation, paper, or mention outside this PR. Web searches for the product name and for "silicium369" turn up nothing relevant.

What is visible about the submission:

  • The GitHub account silicium369-spec was created on 2026-09-29. It has no profile information, no followers, and no activity before this PR.
  • All three of its repositories were created within about 50 minutes on 2026-10-03: silicium-dist (binary release) at 00:25 UTC, silicium-clickbench (a copy of the PR files) at 00:52 UTC, and the ClickBench fork at 01:13 UTC. The PR was opened at 01:14 UTC.
  • The binaries in the release tarball were built between 00:16 and 00:32 UTC the same day.
  • Commits are authored as "The Silicium Authors" using the GitHub noreply address. The only contact is a Gmail address in the license file.
  • "Silicium Systems", "Silicium ClickBench SAS", "Patent Pending", and the "SCSL-1.0" license are mentioned, but I found nothing public to back any of them up.

For a proprietary system, we need to be able to see that it is a real product that users can obtain and use for queries beyond the benchmark. If Silicium is a real product, please share a link to its website or documentation and tell us who develops it.

This branch was successfully deployed

No deployments
benchmark-approval — 0df30c73 Deployed Oct 3, 2026 by silicium369-spec via launch #634
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

joke Not a database

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants