Every measurement published here was captured on isolated physical hardware using native OS kernel accounting (/usr/bin/time -l). We publish full methodology, memory breakdowns, and hardware configurations.
Empirical benchmark harness metrics measured via isolated OS kernel accounting (/usr/bin/time -l).
| Workload | Engine Mode | Runtime | Peak RSS | Throughput | Notes & Conditions |
|---|---|---|---|---|---|
| 100,000 Records (Parquet) | In-memory (Optimized) | 0.113s | 42.1 MB | 884,955 rows/s | 2.83× speedup over un-warmed baseline (0.320s median). Warm protocol. |
| 1,000,000 Records (Parquet) | In-memory Reference | 4.05s | 2,803 MB | 246,913 rows/s | Baseline in-memory key-indexed hash table; full dataset retained in RAM. |
| 1,000,000 Records (Parquet) | External Sort/Merge (64MB) | 2.03s | 98.0 MB | 492,610 rows/s | 64MB sorter budget; streaming chunk spill to NVMe; 96.5% lower RSS. |
| 5,000,000 Records (Parquet) | External Sort/Merge (64MB) | 10.19s | 97.3 MB | 490,677 rows/s | Constant bounded memory footprint despite 5× dataset volume scaling. |
Too many benchmarks in data infrastructure rely on cherry-picked un-warmed cold runs or unverified synthetic comparisons. Rift follows strict peer-reviewable auditing guidelines.
Initial runs that include OS page faults and dynamic linker resolution are explicitly recorded and discarded. Across 5 identical measured trials, optimized 100K Parquet reconciliation achieves a median runtime of 0.113s, reflecting a 2.83× representative speedup over baseline (0.320s).
We avoid in-process heap sampling, which can overlook external C/Arrow allocator fragmentation. Peak RSS is captured through OS kernel process accounting (/usr/bin/time -l maximum resident set size).
All benchmark records are generated deterministically using a fixed random seed (seed = 42) generating 5 realistic enterprise columns: id (Int64 PK), account_id, amount (exact decimal), timestamp, is_settled (bool).
We provide our benchmark harness scripts to prospective enterprise evaluators to reproduce identical measurements on your target Linux or macOS hosts.