Skip to content

Benchmarks · /benchmarks

10.4× faster than Spark, on one machine.

The same 22 TPC-H-derived queries, the same Parquet files, the same laptop. Every number below comes with its method, and you can rerun all of it with the open harness.

Derived from TPC-H, not an audited TPC-H result. Gate runs of 2026-11-12/13 on Krishiv commit f345c5e, approved for the 0.2 release.

10.4×geomean speed-up over Spark 4.2.0 (local[*], 24 shuffle partitions, 20g driver), derived from TPC-H, SF100
228 sKrishiv, all 22 queries · Spark 4.2.0: 2,675 s · Spark with 12 shuffle partitions: 3,137 s
22 / 22SF1000 queries Krishiv completes · Spark finished 6 of 17 within 30 minutes
1.18× / 1.19× / 1.26×DuckDB 2.0.0, for reference: still faster today. Krishiv / DuckDB geomean at SF10 / SF100 / SF1000 (SF1000 over 21 queries; per-query range 0.59× on Q03 to 2.61× on Q16)

Per query, derived from TPC-H, SF100, cold (build of 2026-11-10; the gate build is faster)

log scale · lower is better · ratio = Spark / Krishiv
Krishiv 0.2Spark 4.2.0 (Spark Connect, local[*], 24 shuffle partitions, 20g driver)DuckDB 2.0 (reference)
Q01
15.2 s
298 s
8.3 s
19.7×
Q02
2.4 s
34 s
1.9 s
14.0×
Q03
10.5 s
110 s
8.6 s
10.4×
Q04
5.9 s
80 s
5.5 s
13.4×
Q05
13.0 s
194 s
10.1 s
15.0×
Q06
4.2 s
31 s
4.3 s
7.3×
Q07
11.9 s
89 s
10.0 s
7.5×
Q08
14.4 s
139 s
12.3 s
9.7×
Q09
25.7 s
262 s
17.7 s
10.2×
Q10
13.5 s
81 s
9.5 s
6.0×
Q11
2.2 s
33 s
1.6 s
14.5×
Q12
6.6 s
55 s
5.2 s
8.4×
Q13
11.5 s
74 s
11.9 s
6.5×
Q14
8.0 s
39 s
7.3 s
4.9×
Q15
14.1 s
76 s
7.2 s
5.4×
Q16
3.4 s
50 s
2.0 s
14.9×
Q17
12.5 s
296 s
10.3 s
23.7×
Q18
24.7 s
281 s
14.0 s
11.4×
Q19
8.4 s
39 s
7.0 s
4.7×
Q20
10.3 s
50 s
8.2 s
4.8×
Q21
26.1 s
340 s
16.1 s
13.0×
Q22
2.5 s
24 s
2.3 s
9.6×

SF1000: can it finish?

At 1 TB, Krishiv completes all 22 queries inside its 20 GB memory limit by spilling to disk. Spark, on the same machine and limit, finished 6 of the 17 queries it ran within the 30-minute cap per query.

All 22 SF1000 queries complete, the slowest (Q18) in 654 s. DuckDB runs out of memory on Q01 (a run needing 35 GB) and Q18 at 28 GB; those two were run alone without a cgroup cap. The SF1000 geomean (1.26×) covers the 21 queries with a DuckDB time. Peak resident memory: Krishiv 23.5 GiB (Q09), DuckDB 22.0 GiB (Q21), both above the 20 GB engine limit but inside the 28 GB container.

Method

MachineOne laptop (i7-9750H, 12 threads, 64 GB RAM, NVMe SSD, Linux, powersave CPU governor)
DataDerived from TPC-H, not audited: Parquet generated with tpchgen 3.0.0 (not dbgen), SF10, SF100 and SF1000, read in place (no loading step)
RunsCold: dataset evicted from the OS page cache before every run; two cold runs per query, sequential, one engine after another
Limits12 threads and a 20 GB engine memory limit for Krishiv and DuckDB, inside a 28 GB cgroup; Spark driver memory 20g. Peak RSS can exceed the engine limit (Krishiv 23.5 GiB, DuckDB 22.0 GiB).
VersionsKrishiv 0.2 (commit f345c5e) · Apache Spark 4.2.0, local[*] over Spark Connect, spark.sql.shuffle.partitions=24, Spark 4 defaults otherwise (a run with 12 partitions was slower) · DuckDB 2.0.0
CorrectnessKrishiv's results match DuckDB 2.0.0 on tpchgen data (row count and checksum): 22 of 22 at SF10 and SF100; at SF1000 20 of 22 against a fresh DuckDB run, Q01 against an earlier DuckDB run, Q18 consistent across repeated Krishiv runs only. Spark's answers were not verified; Spark times include the Connect round trip.
TPC noticeDerived from the TPC-H benchmark. These are not TPC-H results: they are unaudited and not comparable with published TPC results. TPC-H is a trademark of the Transaction Processing Performance Council.

Reproduce it

git clone https://github.com/KrishivAI/krishiv-engine
cd krishiv-engine/bench
uv run krishiv-bench reproduce --suite tpch --sf 100

Full harness and engine configurations are in the repository.

★ Star on GitHub

Distributed

Coming in a later release. It will measure: Scale-out, derived from TPC-H, across a cluster: speed-up per added worker and shuffle cost.

Lakehouse

Coming in a later release. It will measure: Reads and writes on Iceberg and Delta tables: scan pruning, commit time and compaction.

Streaming

Coming in a later release. It will measure: Event throughput and end-to-end latency on Nexmark-style queries, with and without checkpoints.

Apache Spark and Spark are trademarks of the Apache Software Foundation. DuckDB is a trademark of the DuckDB Foundation.