> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ellomas.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Performance

> Generation benchmarks, parallel execution, and optimization tips.

# Performance

Seedling is designed for high-throughput data generation. This page covers performance characteristics, benchmarks, and optimization strategies.

## Benchmarks

| Configuration                          | Rows      | Time | Throughput  |
| -------------------------------------- | --------- | ---- | ----------- |
| SQL writer, simple schema              | 10,000    | 0.3s | 33K rows/s  |
| SQL writer, simple schema              | 500,000   | 12s  | 42K rows/s  |
| SQL writer, complex schema (15 tables) | 500,000   | 18s  | 28K rows/s  |
| COPY (Postgres)                        | 1,000,000 | 8s   | 125K rows/s |
| Direct DB insert                       | 500,000   | 22s  | 23K rows/s  |

*Benchmarks on a standard laptop (M1 Pro, 16GB RAM).*

## Parallel Generation

When `--parallel` is enabled, Seedling identifies **independent table subgraphs** and generates them concurrently:

* Tables with no FK dependencies on each other run in parallel workers
* Performance scales with available CPU cores
* **Caution:** Parallel generation breaks determinism (row ordering varies between runs)

```bash theme={null}
seedling generate --count 500000 --parallel --verbose
```

## Batching

The `--batch-size` flag controls how many rows are generated per batch:

* Smaller batches use less memory but more overhead
* Larger batches are faster but use more RAM per table
* Default: 1000

## Optimization Tips

1. **Use COPY for Postgres**: `--copy` mode is 3-5x faster than batched INSERTs
2. **Increase batch size** for large tables: `--batch-size 5000` reduces per-batch overhead
3. **Use deterministic mode** with smaller samples for development: `--seed 42 --count 1000`
4. **Avoid parallelism** if you need deterministic output
5. **Truncate before insert** with `--truncate` to avoid unique constraint violations from existing data
6. **Dry-run first** with `--dry-run` to verify the plan before spending time generating