refactor: update benchmark results and configuration for improved performance in go-migrate

benchmark-results.md

@@ -0,0 +1,75 @@
+# Benchmark go-migrate — 2,000,000 filas
+
+**Tabla**: `Cartografia.MANZANA`  
+**Fecha**: 2026-05-29  
+**Entorno**: Docker local (MSSQL 2022 Developer / PostgreSQL 16 + PostGIS)
+
+---
+
+## Resultado final — 5 pasadas cada dirección
+
+| Métrica | MSSQL → PostgreSQL | PostgreSQL → MSSQL |
+|---|---|---|
+| **Promedio** | **8.37s** | **16.77s** |
+| **Mediana** | 8.16s | 16.33s |
+| **Mínimo** | 7.75s | 16.03s |
+| **Máximo** | 9.17s | 18.46s |
+| **Desv. estándar** | 0.56s | 1.01s |
+| **Throughput promedio** | **~238,892 filas/seg** | **~119,261 filas/seg** |
+| **Factor** | 1x | **~2x más lento** |
+
+---
+
+## Evolución del tuning PG → MSSQL
+
+| Etapa | Config | Tiempo | Throughput | Δ |
+|---|---|---|---|---|
+| Corrida 1 — original | conservadora | 236.8s | ~8,446 /seg | baseline |
+| Corrida 2 — igualada | mismos parámetros | 21.94s | ~91,148 /seg | +10.8x |
+| Tuning A | 4ext/8load 50k | 17.37s | ~115,200 /seg | +1.27x |
+| Tuning C | 16 loaders | 17.26s | ~115,900 /seg | +1.28x |
+| **Tuning D — óptimo** | **8ext/8load 50k** | **~16.77s** | **~119,261 /seg** | **+1.37x** |
+| Tablock + 8 loaders | lock exclusivo serial | ~44s | ~45,000 /seg | ❌ regresión |
+| Tablock + 1 loader | minimal logging | ~47s | ~42,000 /seg | ❌ regresión |
+
+---
+
+## Configuración óptima — `config-reverse.yaml`
+
+```yaml
+max_parallel_workers: 4
+defaults:
+  batches_per_partition: 4
+  max_extractors: 8        # ← mayor lever de mejora
+  extractor_batch_size: 25000
+  extractor_queue_size: 32
+  max_transformers: 8
+  transformer_batch_size: 50000
+  transformer_queue_size: 32
+  max_loaders: 8
+  loader_batch_size: 50000  # sweet spot — 75k y 100k peores
+```
+
+---
+
+## Análisis de la brecha final (~2x)
+
+La diferencia residual entre ambas direcciones es estructural y está en el protocolo de escritura:
+
+| Protocolo | Mecanismo | Overhead |
+|---|---|---|
+| `pgx.CopyFrom` (→ PG) | PostgreSQL COPY protocol — streaming binario sin SQL | mínimo |
+| `mssql.CopyIn` (→ MSSQL) | BCP protocol — row-by-row dentro de un bulk statement | mayor por fila |
+
+`mssql.CopyIn` itera fila a fila via `stmt.ExecContext(row...)` antes del flush final, lo que introduce overhead por fila independientemente del batch size. `pgx.CopyFrom` hace streaming puro.
+
+---
+
+## Hallazgos sobre Tablock
+
+`Tablock: true` en `mssql.BulkOptions` resultó contraproducente en ambos escenarios:
+
+- **Con 8 loaders paralelos**: cada loader compite por un lock exclusivo de tabla → serialización completa (~44s)
+- **Con 1 loader + batch enorme**: sin contención de locks, pero overhead de log + gestión de la lock exclusiva superó el beneficio de minimal logging (~47s)
+
+**Conclusión**: para este patrón de carga (múltiples loaders concurrentes), `Tablock: false` (default) es siempre mejor.

config-reverse-original.yaml

@@ -0,0 +1,35 @@
+max_parallel_workers: 4
+source_db_type: postgres
+target_db_type: sqlserver
+
+defaults:
+  batches_per_partition: 4
+  max_extractors: 2
+  extractor_batch_size: 5000
+  extractor_queue_size: 8
+  max_transformers: 2
+  transformer_batch_size: 12500
+  transformer_queue_size: 8
+  max_loaders: 4
+  loader_batch_size: 25000
+  partition_calculation_strategy: EXACT
+  truncate_target: true
+  truncate_method: TRUNCATE
+  retry:
+    attempts: 3
+    base_delay_ms: 500
+    max_delay_ms: 10000
+    max_jitter_ms: 500
+    max_failed_partitions: 5
+    max_failed_batches_load: 5
+
+jobs:
+  - name: cartografia_manzana_reverse
+    enabled: true
+    source:
+      schema: Cartografia
+      table: MANZANA
+      primary_key: GDB_ARCHIVE_OID
+    target:
+      schema: Cartografia
+      table: MANZANA

config-reverse.yaml

@@ -1,17 +1,17 @@
-max_parallel_workers: 2
+max_parallel_workers: 4
 source_db_type: postgres
 target_db_type: sqlserver
 
 defaults:
-  batches_per_partition: 2
-  max_extractors: 2
-  extractor_batch_size: 500
-  extractor_queue_size: 4
-  max_transformers: 2
-  transformer_batch_size: 2500
-  transformer_queue_size: 4
-  max_loaders: 2
-  loader_batch_size: 5000
+  batches_per_partition: 4
+  max_extractors: 8
+  extractor_batch_size: 25000
+  extractor_queue_size: 32
+  max_transformers: 8
+  transformer_batch_size: 50000
+  transformer_queue_size: 32
+  max_loaders: 8
+  loader_batch_size: 50000
   partition_calculation_strategy: EXACT
   truncate_target: true
   truncate_method: TRUNCATE

scripts/mssql-copy-in/main.go

@@ -12,10 +12,9 @@ import (
 )
 
 const (
-	// totalRows int = 1_000_000
-	totalRows int = 1000
-	chunkSize int = 200
-	queueSize int = 4
+	totalRows int = 2_000_000
+	chunkSize int = 5000
+	queueSize int = 8
 )
 
 func main() {