| name | lakebase-release-workflows |
| description | Opinionated branching + release methodology for Lakebase-paired projects. Use when designing a project's branch layout, cutting a release candidate, promoting between long-running tiers, rolling back, or asking 'where should this work happen?' Encodes the prod / staging / {feature,test,uat,perf} default and the N-tier-capable cut-RC / regression-test / cut-backup / migrate release flow. |
| compatibility | Requires the substrate's [lakebase-scm-workflows](../lakebase-scm-workflows/SKILL.md) skill for branch-pairing primitives, plus the substrate's migrate primitives (FEIP-7091, FEIP-7098). |
| metadata | {"version":"0.2.0"} |
| parent | databricks-lakebase |
Lakebase Release Workflows
The convention and release flow every Lakebase-paired project should follow. Composes on top of lakebase-scm-workflows (which gives you createBranch, getSchemaDiff, applyMigrations, etc.) and adds the opinionated answer for "how do those primitives compose into a release."
The full reasoning + decision record lives in references/branching-and-release-methodology.md. This SKILL.md is the agent-facing tldr.
Branch convention
A project's branches form a directed chain of long-running branches ending in prod. Each working-branch type targets a specific tier in that chain - the architect decides which. Two-tier is the default; N-tier is supported and configured per-project.
Two-tier (default):
prod (only updated by a release)
│
▼
staging (next release accumulates here)
│
▼
{feature, test, uat, perf} (working branches off staging)
Three-tier example (architect splits early work onto dev, validation onto staging):
prod (only updated by a release from staging)
│
▼
staging (only updated by a release from dev; validation work happens here)
│ ▲
│ └─── {test, uat, perf}
▼
dev (next release accumulates here; feature work happens here)
│
▼
{feature}
Any chain length works (dev → staging → preprod → prod etc.). The substrate stores both the chain and the per-type target-tier mapping as project metadata; agents and the extension read it rather than hardcoding.
| Branch | Purpose | Who merges in |
|---|
prod | Production state. Lakebase + git both authoritative. | Release promotion only (no PRs from working branches). |
Each intermediate tier (e.g. staging, dev) | Pre-promotion integration. Where the next release at this tier bakes. | Releases from the tier below + PRs from working-branch types whose configured target is this tier. |
feature/<n> | New work | Dev. |
test/<n> | QA / regression rehearsal | QA. |
uat/<n> | User acceptance | UAT runner / product. |
perf/<n> | Load / perf testing | Perf engineer. |
Each working-branch type pairs to its own Lakebase branch (via lakebase-scm-workflows's createBranch). Schema changes flow up the chain one release at a time.
A release is any merge into a long-running branch
The substrate maintains this convention uniformly. Every merge that lands in a long-running branch IS a release, regardless of where the source branch came from:
feature/X → staging = release to stagingfeature/X → dev (3-tier shop) = release to devdev → staging (3-tier shop) = release to stagingstaging → prod = release to prod
There is no separate "casual PR merge" vs "formal release" distinction. The same four-phase shape runs in all cases - what differs is whether the RC is implicit (feature branch IS the RC) or explicit (cut from to and merged with from).
Release-sprint flow
A release promotes one branch (the from source) into a long-running target (the to tier). from can be a working branch (feature/*, test/*, etc.) or a long-running tier; to is always a long-running tier. The shape is identical at every release, regardless of source kind - only the from/to labels change. to == prod adds the app-deploy step; intermediate-tier releases skip it.
A release proceeds in four ordered phases. The first two happen automatically when the PR opens (pr.yml on to); the second two happen on PR merge (merge.yml on to push). A PR is required for every release - no direct pushes to long-running branches:
- PR open → cut ci-pr-branch from
to. The substrate auto-creates an ephemeral Lakebase branch named ci-pr-branch (per PR) forked from the current to. This is the RC. The PR's migration changes are applied here. Cutting from to (not from) is what locks the release surface - the test runs against the current target's data shape, not whatever's accumulated on from.
- Regression test the ci-pr-branch. pr.yml runs
applyMigrations against ci-pr-branch, then the project's full test suite. Merge is gated on this passing - GitHub branch protection blocks the merge button until the check is green.
- PR merge → cut backup of
to. merge.yml on the to push fires lakebase-cut-backup, snapshotting current to (Lakebase branch + git tag). One-step revert target. Runs at every tier - staging-backup-<id> matters less than prod-backup-<id> but the same primitive runs for both.
- Migrate
to. merge.yml continues with substrate applyMigrations against to's Lakebase branch + git fast-forward of to + (only when to == prod) app deploy.
The 4-phase shape is identical for feature/X → staging and staging → prod. For staging → prod, the ci-pr-branch is cut from a fresh prod (NOT from staging) so the migration test runs against real production data shape - "ensure it will work on my prod" before merging.
When to load the full reference
Load references/branching-and-release-methodology.md when:
- A project is being bootstrapped and you have to recommend a branch layout (including deciding chain length).
- A release decision is in question ("can we cut the RC from
from this time?" - no, see the doc).
- A consumer asks why
to-backup exists separately from the RC.
- An N-tier shop's per-tier policy gates need to be designed.
- Drift from this convention is being proposed and you need the original reasoning to push back or extend.
Primitives this skill expects (future work)
The substrate doesn't yet ship the release orchestrator. These primitives are planned (FEIP-7059 roadmap). All of them are parameterized over the from/to tier pair so the same primitive serves every adjacent pair in any chain length:
bootstrap-branch-convention({chain, workingTypeTargets}) - given a fresh project, creates the configured long-running chain (default [staging, prod]) plus the working-branch types, all from prod, and writes parent-pair metadata + the type → target-tier mapping.cutRC({from, to, releaseId}) - branches the release candidate off current to and merges from in.regressionTest({rc, suite}) - runs the project's full e2e suite against the RC branch.cutBackup({to, releaseId}) - snapshots current to for rollback.migrate({rc, to, releaseId}) - applies substrate applyMigrations against to's Lakebase branch and fast-forwards the git pointer.release - orchestrator that calls the four phases in order with explicit gates between each.
Until these land, follow the manual procedure documented in the reference.
