DB Migrations in CI/CD

What GoFr provides

GoFr ships a migration system that you wire up via app.Migrate(migrations.All()). It supports MySQL, PostgreSQL, Redis, ClickHouse, Cassandra, and Elasticsearch. Records are kept in a gofr_migrations table (or Redis hash). A distributed lock (gofr_migration_locks table or Redis SETNX) prevents two replicas from running the same migration concurrently — see Handling Data Migrations for the full mechanics.

That means any deployment shape works correctness-wise. The CI/CD question is operational: do you want migrations tied to app startup, or separated?

Option A: Migrations on app start (default)

The simplest setup, and the default GoFr lifecycle. Every replica calls app.Migrate(...) in-process before serving traffic — there is no separate migration binary or subcommand. The first replica to acquire the lock runs the migration; the others observe the populated gofr_migrations table and no-op. After migrations finish, all replicas continue normal startup.

Pros:

• Zero extra infra. One artifact per service.
• Migrations cannot drift from code — they ship in the same image.
• Idempotent under concurrency: the lock plus the version table guarantee each migration runs exactly once across replicas.

Cons:

• A migration error fails the readiness probe of every replica simultaneously, which can take down healthy old pods if the rollout strategy isn't careful.
• Slow migrations delay every pod's start.
• Logs are mixed with normal application logs.

Use this for small services and early-stage projects.

Option B: Separate cmd/migrate binary as a Helm pre-upgrade Job

For multi-replica production services, run migrations as a Kubernetes Job triggered by Helm before the Deployment rolls forward. There is no built-in gofr migrate CLI or MIGRATE_ONLY env mode in the framework. Instead, organize your application as two binaries built from the same Go module: the serving binary (cmd/server or your existing main.go) and a small dedicated migration binary (cmd/migrate). The migration binary calls gofr.New(), registers migrations, calls app.Migrate(...) (which is synchronous and runs to completion before returning), and exits without ever calling app.Run(). This is application code organization — not a framework knob.

// cmd/migrate/main.go
package main

import (
    "gofr.dev/pkg/gofr"

    "yourmodule/migrations"
)

func main() {
    app := gofr.New()
    // app.Migrate runs migrations synchronously using GoFr's distributed lock
    // and returns once they have completed (or failed). No app.Run() — this
    // binary is intended to be invoked as a one-shot Job.
    app.Migrate(migrations.All())
}

Build it as a separate binary in the same image (or a slimmer migrate-only image), and invoke it from the Helm pre-upgrade Job:

apiVersion: batch/v1
kind: Job
metadata:
  name: {{ .Release.Name }}-migrate-{{ .Values.image.tag | replace ":" "-" }}
  annotations:
    "helm.sh/hook": pre-install,pre-upgrade
    "helm.sh/hook-weight": "-5"
    "helm.sh/hook-delete-policy": before-hook-creation
spec:
  backoffLimit: 0
  template:
    spec:
      restartPolicy: Never
      containers:
        - name: migrate
          image: "{{ .Values.image.repo }}:{{ .Values.image.tag }}"
          # Invoke the dedicated migration binary built from cmd/migrate.
          command: ["/app/migrate"]
          envFrom:
            - secretRef:
                name: {{ .Release.Name }}-db

Pros:

• Failed migration fails the Helm release atomically; the rollout never starts.
• Job logs are clean and separately addressable: kubectl logs job/....
• Application pods see the new schema by the time they boot.
• The migration binary is just Go code — no framework feature to learn beyond app.Migrate.

Cons:

• One more binary to build and template. The migrator Job's image SHA must always match the Deployment image SHA.

Use this for production.

Init container: usually not the right tool

It is tempting to put migrations in an initContainer. Don't, in multi-replica deploys. Each pod's init container will race for the lock. GoFr's lock makes that safe, but it also means N-1 pods just wait for nothing while the rollout takes longer than necessary, and a failing migration manifests as N pods CrashLooping. A pre-upgrade Job centralizes the failure into one Pod and one log stream.

Init containers are fine for single-replica services or local dev.

Expand-contract for zero downtime

Schema changes that break the previous app version are dangerous during a rolling deploy because both versions run simultaneously. Use the expand-contract pattern:

1. Expand — release migration A that adds the new column/table without removing the old one. Old code keeps working.
2. Migrate code — release the app version that writes both old and new, reads new (or vice versa).
3. Backfill — copy data from old to new in a background Job if needed.
4. Contract — once the app version is stable, release migration B that drops the old column/table.

This typically means at least two deploys per breaking change. It is the price of zero-downtime.

Rollback strategy

GoFr currently runs migrations in UP mode only (verified against pkg/gofr/migration semantics described in the migrations doc). That has implications for rollback:

• App-level rollback (image SHA): always safe if the schema change was expand-only.
• Schema rollback: write a new forward migration that reverses the change. Treat the database as append-only history.
• Snapshots before destructive migrations are a safety net for genuine emergencies.

Idempotency

Write migration SQL so re-running it is harmless: CREATE TABLE IF NOT EXISTS, ALTER TABLE ... ADD COLUMN IF NOT EXISTS (PostgreSQL), or guarded IF checks. The lock prevents concurrent runs, but idempotency protects against partial failures and manual re-runs.

CI ergonomics

• Run migrations against an ephemeral database in CI on every PR. If the migration fails in CI, it never reaches prod.
• Tag the migration Job with the Helm release name and image SHA so old Jobs are identifiable: migrate-{{ .Release.Name }}-{{ .Values.image.tag }}.
• Pin the database driver version in go.mod and treat upgrades as their own change.