---
type: product_context
title: Performance Plan & SLOs
tags: [performance, benchmarking, slo]
---
# Performance Plan & SLOs
Initial targets; will calibrate after baseline runs.
## SLOs (v0)
- get_info
- p95 < 500ms (small), < 2s (medium)
- translate (e.g., GeoTIFF → COG)
- p95 < 10s (medium ~512MB), single-thread default
- warp
- p95 < 20s (medium reprojection)
- Memory
- Steady-state RSS < 1.5× input size for translate/warp; no leaks across 100 runs
## Benchmark Harness
- Script: `scripts/bench.py`
- Cases: tiny (KB), small (MB), medium (hundreds MB)
- Metrics: wall/CPU time, RSS peak, IO bytes, GDAL error counts
- Output: `bench/` JSON per case + `bench/REPORT.md`
- Determinism: fixed seeds, temp dirs `TMPDIR/gdal-mcp/<run>/<case>`
## Tuning & Config
- GDAL creation options (COG tiling, block sizes), resampling choice
- Concurrency cap: `GDAL_MCP_MAX_CONCURRENCY`
- Env verification at startup: GDAL, PROJ data paths, temp/cache dirs
## Acceptance Criteria
- Baselines produced and stored in repo
- Regressions tracked; CI can run tiny/small cases on PRs
- Logs contain timings and error counts at INFO level (stderr)
