Mixed Workload Stress Tests

Stress tests that exercise all Neumann engines simultaneously with realistic workload patterns.

Test Suite

Test	Scale	Description
`stress_all_engines_concurrent`	25K ops/thread, 12 threads	All engines under concurrent load
`stress_realistic_workload`	30s duration	Mixed OLTP + search + traversal

Results

Test	Key Metric	Result
All engines	Combined throughput	841 ops/sec
All engines	Relational p50	12ms
All engines	Graph p50	5us
All engines	Vector p50	< 1us
Realistic workload	Mixed throughput	232 ops/sec
Realistic workload	Read rate	91 reads/sec
Realistic workload	Write rate	68 writes/sec
Realistic workload	Search rate	72 searches/sec

Running

# Run all mixed workload stress tests
cargo test --release -p stress_tests --test mixed_workload_stress -- --ignored --nocapture

# Run specific test
cargo test --release -p stress_tests stress_all_engines_concurrent -- --ignored --nocapture

All Engines Concurrent

Tests all engines (relational, graph, vector) under simultaneous heavy load from 12 threads.

What it validates:

Cross-engine concurrency safety
Shared TensorStore contention handling
No deadlocks or livelocks
Correct results under maximum stress

Workload distribution per thread:

Relational: INSERT, SELECT, UPDATE
Graph: NODE, EDGE, NEIGHBORS
Vector: EMBED, SIMILAR

Expected behavior:

No panics or assertion failures
All operations complete (no hangs)
Data consistency verified post-test

Realistic Workload

Simulates a production-like mixed workload over 30 seconds.

What it validates:

Sustained throughput over time
Memory stability (no leaks)
Latency consistency

Workload pattern:

40% Reads (SELECT, GET, NEIGHBORS)
30% Writes (INSERT, UPDATE, NODE)
30% Searches (SIMILAR, PATH)

Expected behavior:

Throughput variance < 20%
Memory usage stable
No degradation over time

Engine Latency Breakdown

Engine	Operation	Typical p50	Notes
Relational	SELECT	1-10ms	Schema lookup overhead
Relational	INSERT	3-15ms	Index maintenance
Graph	NEIGHBORS	5-50us	Adjacency list lookup
Graph	PATH	100us-5ms	Scales with path length
Vector	EMBED	1-5us	Hash insert
Vector	SIMILAR	1-100ms	Scales with corpus size

Bottleneck Identification

When mixed workload throughput is lower than expected:

Vector search dominates: Use HNSW index for SIMILAR queries
Relational scans: Add hash/B-tree indexes on filter columns
Graph traversals: Add LIMIT to NEIGHBORS/PATH queries
Contention: Check hot shards with instrumentation

Scaling Considerations

Bottleneck	Solution
CPU-bound	Add more cores, enable rayon parallelism
Memory-bound	Enable tiered storage, use sparse vectors
I/O-bound	Use NVMe storage, increase buffer sizes
Network-bound	Batch operations, use local cache

Keyboard shortcuts

Neumann