Quick answer
Progressive delivery is releasing a change to users in stages, such as 1%, 5%, 25% and 100% of traffic, and moving to the next stage only when measured gates (error rate, latency, a business metric) stay healthy, with traffic returned to the old version if they fail.
What is progressive delivery?
Progressive delivery is releasing a change to users in stages, starting with a small share of traffic and widening only when measured results stay healthy. It separates two things that used to happen together: deploying code to production and releasing it to people. The term was popularized by RedMonk analyst James Governor, whose August 2018 post argued that routing techniques such as feature flags and canarying deserve their own name beyond CI/CD.
In practice, each stage has a gate: a metric and a threshold the new version must meet before it gets more traffic. If the gate fails, traffic goes back to the old version. The rest of this page shows how to choose stages, how to size a gate so it can actually detect a problem, and what you can and cannot measure.
Progressive delivery vs related practices
| Practice | What it controls | Who sees the change | Typical rollback |
|---|---|---|---|
| Big-bang release | Everything at once | All users | Redeploy the old version |
| Canary release | Share of traffic per version | A small random share, then more | Route traffic back |
| Feature flag | Which users get a feature | Chosen users or percentages | Switch the flag off |
| Blue-green deployment | Which full environment is live | Everyone, at the switch | Switch back to the old environment |
| Beta program | Opt-in access | Volunteers who know it is unfinished | End the program |
| Progressive delivery | The whole staged process with gates | Growing share, by stage | Automatic or manual, at any stage |
Progressive delivery is not a rival to canaries and flags. It is the process that uses them, plus the measurement rule that decides when to continue.
Staged rollout: an example plan
Sample data: a web service receives 400,000 requests per day. The stable version has an error rate of 0.5%. The team will not accept more than 0.3 percentage points of extra errors, so the gate is a canary error rate of at most 0.8%.
| Stage | Traffic to new version | Requests per day at that share | Gate to move on |
|---|---|---|---|
| 1 | 1% | 4,000 | No crash loops; error rate not above 5% |
| 2 | 5% | 20,000 | Error rate ≤ 0.8% and latency within 10% of stable |
| 3 | 25% | 100,000 | Same gate, plus no drop in the main business metric |
| 4 | 50% | 200,000 | Same gate |
| 5 | 100% | 400,000 | Keep watching for 24 hours |
Argo Rollouts, an open source controller, expresses plans of this shape as setWeight steps with pause durations between them, and aborts the rollout if an analysis fails. Without a traffic router it approximates the percentage with replicas: with 10 replicas, a 10% step means one canary and nine stable.
How long should each stage last?
A gate is only useful if the stage collects enough requests to detect the problem you care about. For comparing two error rates (0.5% on stable, 0.8% if the canary is worse) with a two-sided 5% significance level and 80% power, the standard two-proportion sample size formula gives about 11,263 requests per arm. Treating that as a rough guide (it assumes equal arms; a small canary against a large stable group needs a little more):
| Stage | Share | Time to reach 11,263 canary requests at 400,000/day |
|---|---|---|
| 1 | 1% (4,000/day) | 2.8 days (67.6 hours) |
| 2 | 5% (20,000/day) | 0.56 days (13.5 hours) |
| 3 | 25% (100,000/day) | 0.11 days (2.7 hours) |
| 4 | 50% (200,000/day) | 0.06 days (1.4 hours) |
Two conclusions follow. First, the 1% stage cannot detect a 0.3-point rise in a sensible time, so do not use it for that: use it as a blast-radius check for failures that are large. Detecting a jump from 0.5% to 5% needs only about 206 requests per arm, which 1% of traffic provides in roughly an hour. Second, the cost of a bad release is bounded by the stage share. Google's SRE Workbook gives the arithmetic: a canary with a 20% error rate on 5% of traffic raises the overall error rate by about 1% (0.20 × 0.05 = 0.01).
Choosing metrics for the gates
The SRE Workbook's guidance on canary metrics is a good checklist:
- Start with signals users feel, such as failed requests and latency, not CPU.
- Keep the list short, "perhaps no more than a dozen."
- Leave out 400-level HTTP codes, which often reflect user error rather than a defect.
- Break monitoring down by canary versus control; whole-service averages can hide a canary's errors.
- Run one canary at a time, so you know which change moved the metric.
- Make sure metric intervals are no longer than the canary duration.
Add one business gate to the technical ones: for a web or app change, a core conversion or activation rate, compared at the same stage between old and new.
When progressive delivery is not enough
- Rare problems. A defect that affects 1 in 10,000 requests will not show up in a 4,000-request stage.
- Slow problems. Memory leaks, billing errors and retention drops appear after hours or weeks. Keep the final stage under watch.
- Shared state. Database schema changes cannot always be rolled back by moving traffic. Make them backward compatible first.
- Unequal arms. The sample size above assumes the canary and stable groups are the same size. A 1% canary against a 99% stable group needs more canary requests for the same power, so treat the table as a lower bound.
- Low traffic. With 4,000 requests a day, even the 100% stage is thin. Use fewer stages and longer holds.
What TapCub measures during a rollout
Imagine a mobile app update released to 5% of users first. TapCub does not split traffic or run flags, so the rollout itself happens elsewhere (your app store's staged release, a flag service or your deploy tool). TapCub only measures what each version does.
- Compare versions. App analytics › Users › Versions lists each version with its share of active users and crash count. Select the new version to filter the page to it.
- Compare stability. App analytics › Quality & attribution › Stability shows the crash-free share (1 − crash users ÷ active users); filter by the new version and read it against the previous one.
- Watch errors on web. Monitoring › Errors groups errors with counts, affected sessions and last-seen times, so a new error group after a deploy stands out.
- Get alerted. In All-platform analytics › Operations › Monitors, create a monitor on
error_ratewith a threshold such as 0.8%. A dry run tells you how often it would have fired over the past 30 days. A monitor only notifies; it does not roll anything back. - Mark each stage. Add a note per stage in Configuration › Annotations (1%, 5%, 25%...), so every chart shows where the rollout steps were.
Sources
- RedMonk, James Governor, Towards Progressive Delivery (6 August 2018).
- Google, Site Reliability Engineering Workbook: Canarying Releases.
- Argo Rollouts documentation, Canary strategy.
- NIST/SEMATECH e-Handbook of Statistical Methods, 7.3.3 Two-proportion tests (background for the two-proportion comparison used in the sample size).
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Product Experience Specialist
Explores usability, user journeys, and feature adoption through practical examples of onboarding and everyday product use.