8000 GitHub - donge/zighouse · GitHub
[go: up one dir, main page]

Skip to content
dongePublic

About

No description, website, or topics provided.

Resources

Stars

6 stars

Watchers

0 watching

Forks

Repository files navigation

ZigHouse

A schema-driven columnar OLAP engine written in Zig. Ingests Parquet and ClickHouse Native blocks into a MergeTree-compatible on-disk format and executes analytical SQL via an IR-pipeline executor.

Install

macOS (Homebrew)

brew tap donge/zighouse
brew install zighouse
brew services start zighouse

Linux / Docker

docker pull ghcr.io/donge/zighouse:latest

Binary releases are available on the releases page.

From source

Requires Zig 0.17+.

zig build -Doptimize=ReleaseFast -Dstrip=true -Dstatic-libs=true

No external parser dependency. All 43 ClickBench queries are handled by the native Zig SQL parser.

Quick start

# Import a Parquet file
./zighouse import hits.parquet --table=hits

# Start the HTTP server
./zighouse serve /var/lib/zighouse

# Run all 43 ClickBench queries
./zighouse bench --store=./store --query=9

ClickBench

ZigHouse participates in ClickBench. Submission files: clickbench-submit/zighouse/.

bash clickbench-submit/zighouse/install   # download binary
bash clickbench-submit/zighouse/run.sh    # run benchmark

Current results — Apple M2 Pro, 10M rows

Warm-best times from the official ClickBench driver (./query process per try, first try discarded; BENCH_TRIES=3):

Q ms Q ms Q ms Q ms
Q1 4.0 Q12 23 Q23 67 Q34 169
Q2 6.0 Q13 38 Q24 132 Q35 170
Q3 8.0 Q14 91 Q25 20 Q36 45
Q4 11 Q15 42 Q26 16 Q37 28
Q5 34 Q16 54 Q27 20 Q38 26
Q6 48 Q17 99 Q28 37 Q39 22
Q7 4.0 Q18 30 Q29 173 Q40 61
Q8 8.0 Q19 331 Q30 7.0 Q41 61
Q9 73 Q20 11 Q31 76 Q42 53
Q10 96 Q21 79 Q32 62 Q43 28
Q11 21 Q22 49 Q33 274

Warm total: 2.71 s, 0 null results.

Against DuckDB 2.0 on the same 10M dataset:

  • Query sum: 0.76× DuckDB (ZigHouse 1.73 s vs DuckDB 2.27 s without process startup)
  • Per-query gate: ≤3× DuckDB on every query
  • 3M gate: 32/32 deterministic query results match

zighouse query, zighouse bench, and zighouse serve now auto-detect both Store V2 parts and streamed ClickHouse MergeTree parts, so file imports and INSERT streams run through the same planner and pipeline. See scripts/compact-part-query-test.sh.

Tests

zig build test

Architecture

ZigHouse uses a single execution path:

Parquet / ClickHouse TCP Native blocks
    ↓  loader.zig / ingest/
    ↓  MergeTree-compatible column files (.bin, .str.bin)
    ↓
SQL parse  →  IR plan  →  pipeline executor
    ↓           ↓               ↓
sql/parser  planner.zig   pipeline.zig
                             (hash agg, TopK, scan, filter)
    ↓
serializer.zig  →  CSV / ClickHouse Native Block

Key modules: schema.zig, catalog.zig, generic_sql.zig, core/exec/pipeline.zig, core/exec/planner.zig, loader.zig, ingest/tcp_server.zig.

License

MIT — see LICENSE.

About

No description, website, or topics provided.

Resources

Stars

6 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0