Performance

Drift's analysis time depends on repository size (files and functions). This page provides empirical timing data and tuning guidance.

Benchmark matrix

Measurement context

Single-threaded, cold start, all 23 signals enabled, embeddings on, default configuration, measured on a single developer workstation. Times include AST parsing, git history extraction, all signal detectors, and scoring. Score column omitted where timing is the primary metric.

Repository	Files	Functions	Duration (s)	Score
sqlmodel	171	399	1.4	—
drift (self) pre-ADR-007	66	436	0.4	—
requests	34	670	4.9	—
uvicorn	78	790	2.8	—
httpx	60	1,134	3.6	—
flask	65	1,405	4.7	—
starlette	67	1,474	4.7	—
drift (self)	197	1,676	4.1	—
rich	176	1,790	7.8	—
poetry	437	3,109	13.3	—
NestJS	1,667	3,472	17.5	—
sanic	296	3,641	11.3	—
FastAPI	664	3,902	13.1	—
pwbs-backend	484	5,057	9.3	—
celery	371	7,546	30.5	—
pydantic	399	8,350	60.7	—
Django	2,890	31,191	35.9	—

Scaling characteristics

Analysis time grows roughly as \(O(n \times m)\) where \(n\) is the number of files and \(m\) is the average function count per file. Individual signals have different scaling profiles:

• AST parsing (ingestion): linear in files × LOC
• Pattern fragmentation, mutant duplicates: quadratic in functions within each module (pairwise comparison), but bounded by module size
• Temporal volatility, co-change coupling: linear in git commits × files
• Embedding-based signals: dominated by embedding computation, batched by embedding_batch_size

Resource guardrails

Drift includes built-in safety limits:

Parameter	Default	Config key
Max discovered files	10,000	thresholds.max_discovery_files
Embedding batch size	64	embedding_batch_size
ECM max files	50	thresholds.ecm_max_files
ECM lookback commits	20	thresholds.ecm_lookback_commits

Tuning for large repositories

If analysis takes too long, consider these options in order of impact:

1. Restrict scope: drift analyze --path src/core analyzes only a subdirectory
2. Use diff mode: drift check --diff HEAD~1 analyzes only changed files — typically 10-100× faster
3. Disable embeddings: Set embeddings_enabled: false in drift.yaml — removes the most memory-intensive operation
4. Lower discovery cap: Set thresholds.max_discovery_files to cap file enumeration
5. Reduce git depth: Lower thresholds.ecm_lookback_commits for faster co-change analysis

CI performance

For CI pipelines, drift check (diff mode) is the recommended entry point. It analyzes only files changed in the current branch or commit, making it predictably fast regardless of repository size.

Typical CI timings:

• Small PR (1-10 files): < 3 seconds
• Medium PR (10-50 files): 3-10 seconds
• Large PR (50+ files): 10-30 seconds

Related pages

• Troubleshooting
• Configuration