Bench Datastore

Datastore benchmarking harness for swamp. Two workflow-based test scenarios that exercise any swamp datastore backend under sustained load without lock contention.

Scenario A: Throughput (breadth)

Each worker owns isolated models. Cycles through write → read → query operations measuring per-step latency. Tests: sync push/pull speed, chunked storage, GSI queries, workflow overhead.

Scenario B: Write Stress (depth)

Each worker owns one model, writes continuously with varying payload sizes. Tests: sustained single-writer throughput, version accumulation, chunked vs inline performance, hot-partition behavior.

Usage

swamp extension pull @webframp/bench-datastore

# Create harness instance for worker 1 (throughput scenario)
swamp model create @webframp/bench-datastore/harness bench-harness-w001 \
  --global scenario=throughput --global worker_id=1 --global models_per_worker=50

# Setup (creates worker-owned models)
swamp model method run bench-harness-w001 setup

# Run throughput cycle once
swamp workflow run @webframp/bench-throughput-cycle \
  --input harness_model=bench-harness-w001 \
  --input probe_model=bench-probe-w001 \
  --input iteration=1

# Run write-stress once
swamp workflow run @webframp/bench-write-stress \
  --input harness_model=bench-harness-w001 \
  --input probe_model=bench-probe-w001 \
  --input iteration=1

Both workflows are designed to be called in a tight loop by an ECS Fargate worker or local script. Duration is controlled by the caller.

2026.07.24.1

Added: Initial release of the datastore benchmarking harness.

• Two workflow-based scenarios: throughput (breadth) and write-stress (depth)
• Zero lock contention by construction — each worker owns isolated models
• Backend-agnostic measurement of datastore I/O independent of swamp core locking
• Configurable at runtime via workflow inputs (scenario, worker count, duration controlled by caller)
• Deployable on ECS Fargate or runnable locally
• Payload delivery uses stdin streaming via temp files to avoid Linux MAX_ARG_STRLEN (131072 bytes) limit on large (500KB) write-stress payloads
• Probe measurements in workflows are scoped to the worker's harness model via --model filter to prevent cross-worker data inflation under concurrent load
• payload_size accepted as an optional workflow input on bench-write-stress for controlled single-size runs