Open-source experiments in vectorised execution on the JVM: compile a whole query step into one SIMD loop, keep the math bit-exact, and write down what was measured and what failed.
Everything here targets JDK 25 and the Vector API, is licensed Apache-2.0, and is research code: the contracts are stated in each README, and the numbers in the posts trace to results files committed next to the code that produced them.
| Varka | A research fork of Apache Spark that compiles an eligible projection into a single fused vector loop, emitted at runtime with the JDK 25 Class-File API and run over zero-copy views of Arrow columns. Behind one config flag; a batch it cannot serve falls back to stock Spark, and a Varka failure never fails a query. |
| vfdlibm | fdlibm on the Java Vector API: math functions over DoubleVector lanes whose every lane is bit-identical to java.lang.StrictMath. exp, expm1, log, log1p and log10 are on Maven Central as io.github.vecbricks:vfdlibm; functions are added as Varka needs them. |
| jegg | E-graphs and equality saturation for Java 25, a port of egg, with extraction over several roots that keeps a shared subterm in one form. |
| Engineering notes | The posts behind Varka: each is built from Markdown and figure scripts in the Varka repository, so a number or a drawing traces to a committed file. |
- Nine numbers between your query and machine code
- The 8000-byte cliff in Spark SQL
- Your sin() depends on your JDK and your instruction set
- Why a warmed Vector API kernel can still allocate
All posts are listed at vecbricks.github.io.