OLTP vs OLAP Latency Benchmark

OLTP-versus-OLAP latency benchmarking compares the two fundamentally different access patterns databases serve: short transactional operations (OLTP) and long analytical queries (OLAP). It matters most for hybrid transactional/analytical processing (HTAP) systems and translytical databases that claim to do both, where the key question is whether running heavy analytics degrades transactional latency. The category is conceptual — it combines transactional and analytical benchmarks rather than being a single standardized test.

What It Measures

The benchmark captures point-query latency for OLTP operations (single-row lookups, inserts, updates), scan-and-aggregate latency for OLAP queries (large group-bys and joins), the 99th-percentile latency of transactions, and crucially the throughput and latency under mixed load when both workloads run simultaneously. The defining measurement is interference: how much OLTP tail latency rises while analytics run, and how much analytical throughput drops while transactions run.

Methodology

A typical setup pairs an OLTP workload (such as TPC-C or a YCSB read-write mix) with an OLAP workload (such as TPC-H or TPC-DS queries) running against the same data set. Three phases are measured: OLTP alone, OLAP alone, and both concurrently. The concurrent phase reveals resource contention on CPU, memory bandwidth, buffer pool, and I/O, and tests whether the engine isolates the two — for example via separate row and column stores, workload scheduling, or resource governors. CH-benCHmark, which fuses TPC-C and TPC-H over a shared schema, is a well-known instantiation of this idea. Metrics are reported per workload and as the degradation between isolated and mixed runs.

How to Interpret Results

The headline number is interference, not isolated speed: a system can be fast at each workload separately yet collapse when they overlap. Compare OLTP p99 latency in the mixed phase against the isolated phase, and likewise for OLAP throughput. Examine how the engine maintains analytical freshness — column stores that lag behind the row store trade consistency for speed. State the data size, the OLTP and OLAP workloads used, and the concurrency, since all strongly affect contention.

Limitations

There is no single accepted standard, so methodologies vary and results are hard to compare across studies. Combining two benchmarks over one schema can create artificial patterns unlike any real application. Tuning can favor one workload at the other's expense, so a single configuration rarely represents all use cases. Use this benchmark category to evaluate HTAP isolation and freshness trade-offs, defining workloads that resemble your actual transactional and analytical mix.