On Tuesday we pushed a one-line bug fix to five hosts. One approved it before lunch, one took the night, one sent back a screenshot asking why a button had moved, one approved a stale build, and the last sat in review until Friday. The commit was identical. Only the calendar wasn't.
The part nobody warns you about: you don't run five releases, you run one release that five different gates have to clear. The slowest gate sets the tempo, and no amount of CI polish changes that. We stopped asking how fast we can ship and started asking what has to be true before the slowest host says yes.
One train, five timetables
The first surprise isn't review time, it's that there is no single artefact. Each host wants its own bundle: different app IDs, different signing, different SDK shims, and conditional branches for capabilities that only exist on one of them. We keep one codebase, but the build step fans out into five outputs, and every one of them needs to be traceable back to the same commit. Lose that mapping and you'll spend an afternoon arguing about which binary a reviewer is actually looking at.
Version numbers are the quiet version of the same problem. A single incrementing number stops meaning anything when host A is on the third revision of 3.2 and host B is still reviewing the first. We ended up with a small mapping table: one product version, and a per-host build that declares which features it contains. It's boring bookkeeping, and it's the only thing that makes a support ticket interpretable.
| Pipeline step | One host | What five hosts add | Who owns it |
|---|---|---|---|
| Build | One bundle, one config | Per-host variants, conditional code, separate signing | Build engineer |
| Version number | One incrementing integer | A mapping table so 3.2 means the same features everywhere | Release manager |
| Upload and review | One submission | Five consoles, five queues, five ways to phrase a rejection | Release manager |
| Staged rollout | One percentage dial | Five dials that can't be moved together | Release manager and on-call |
| Rollback | Redeploy the previous build | Some hosts only roll forward, with a fresh review | On-call |
| Config and flags | Shipped with the code | Remote config becomes the only lever between reviews | Platform team |
The shortcut that costs you the account
Now the uncomfortable part. When review cycles don't line up, remote config starts looking like the answer: flip a flag, change a layout, fix a price, no new submission. And to a point it's legitimate — copy, ordering, feature flags, kill switches belong outside the binary.
Push it further and you're not configuring anything. You're shipping unreviewed behaviour into a build a reviewer approved for something else. Hosts read that as a policy violation, and the penalty isn't a rejected version. It's the developer account. We've watched a team lose its distribution for weeks over a toggle that quietly rewrote a purchase flow. That's not a rounding error on a release schedule; it's the whole channel gone.
The other cost is that the slowest host really does set the pace. You can land a fix on four hosts on Tuesday and still tell users it's fixed only when the fifth clears on Friday. Either you hold the announcement, or you ship a feature that behaves differently depending on which app opened it. We've tried both. Holding is less confusing.
Rollback is where the fantasy of symmetry dies completely. On some hosts you redeploy an older build; on others there's no backward gear at all, only a new version that has to pass review again. Your incident runbook can't have one page. It needs five.
Make the slowest gate visible
We ended up building Cross Mini App because keeping five release calendars in sync by hand was making us miserable. It collapsed the plumbing into one pipeline. It did not collapse the five review queues, and we'd be lying if we said it had.
So our board now shows the hosts as lanes, not steps, with the slowest one pinned at the top. Nothing gets announced until that lane clears. Config changes follow the same discipline as code: reviewed, logged, and reversible. It's slower than the demos suggest. It's also the only version of this we've found that doesn't end with someone asking, on a Friday, why the button moved in one app but not the others.
Sources
- WeChat Mini Program framework docs — upload, review and staged release rules, plus what counts as a version.
- Alipay Mini Program developer docs — a second review queue with its own release and rollback controls.
- Apple Developer Documentation — a useful comparison for how review gates shape release cadence.
- uni-app documentation — one build, multiple platform outputs, and where conditional compilation comes in.