Watch & compare
Canary deployments
Different ways to see the same idea. Play one cut at a time, then return to the guide to try it yourself.
Why we recommend Split the signal
Split the signal makes the chapter’s most useful release decision concrete: a green 0.4% pooled error rate conceals 40% failures in candidate exports, so promotion must wait for the right breakdown. Its same-window denominators and retained failed window make it the strongest main entry for the guide and exercise. The original is a simpler routing-and-latency introduction, while Canary the outcome adds a useful agent-prompt example where HTTP 200 coexists with failed PR tests and a visible gate returns new sessions to control.
Revised original · 26 sec
Original explanation
A canary deployment exposes a change to limited production traffic before expanding its rollout.
Read the video sequence
- All traffic uses v1, the control. Its illustrative p95 latency is 120 ms; the unexposed candidate has no result.
- The router allocates twenty percent to the candidate. Three stylized four-to-one traffic cycles begin; the animated packets are not a statistical sample.
- The candidate’s illustrative p95 is 900 ms, compared with 120 ms for the control. Its work bar remains busy longer.
- The 7.5-fold latency difference is highlighted. ROLLBACK appears at 17.4 seconds and a signal returns to the router; the share bar moves back to the control from 18 seconds.
- New requests use the control. The candidate receives zero percent, and its failed latency result remains visible. The stage fades from 24 seconds, retaining the title for the loop.
About this cut
Production: Original prototype; model not recorded.
Revision: gpt-6.1-sol comparative review.
Illustrative p95 values summarize an observation window; animated requests are not enough to estimate a tail percentile.
A weighted traffic router is assumed. The exposure percentage and latency ratio are examples, not universal rollout thresholds.
New traffic returns to control; in-flight work and existing side effects require separate handling.
Audio analysed, not physically heard. Original authoring model was not recorded.
Earlier transcript corrections: gpt-6-astra comparative review. Final comparative verification: gpt-6.1-sol comparative review. Original authorship is unchanged.
Recommended · Astra · 30 sec
Split the signal
Break down a canary’s measurements by version and changed operation before expanding exposure; a healthy pooled rate can hide a severe regression.
Read the video sequence
- The stable release receives all requests. Its previous observation window has no errors. The candidate is not yet receiving traffic.
- A release containing an agent-assisted change receives ten percent of traffic. Two routing cycles send nine requests to stable for each one to the canary. The previous result is cleared while a new window is collected.
- The new window shows forty errors out of ten thousand requests, or 0.4 percent overall. Promotion is still pending; the green pooled result is not approval.
- The same window is split by version for exports: stable has zero failures among nine hundred exports; the canary has forty among one hundred, or forty percent. The failure signal blocks promotion.
- The rollout is aborted and new traffic returns to stable. The candidate receives zero percent, while its failed measurement stays visible as historical evidence. This illustrates containment, not repair of earlier effects.
About this cut
Production: gpt-6-astra.
Revision: gpt-6.1-sol comparative review.
Illustrative export-service scenario; counts, percentages, and compressed observation time are not safety guarantees.
Routing supports weighted traffic in this example; replica counts alone do not guarantee exact traffic shares.
Abort redirects new traffic but does not undo persisted changes or external side effects; the failed window is retained.
Audio levels, timing and spectrogram were analyzed; physical listening was not performed.
The original author review was a self-review; the completed independent comparison is now saved in reviews/comparisons/canary-deployments.json.
Earlier visual repairs: gpt-6-astra comparative review. Final comparative verification: gpt-6.1-sol comparative review. Original authorship is unchanged.
Alternative · Sol · 28 sec
Canary the outcome
Compare an agent prompt canary with a concurrent control using required artifact checks, and withdraw it on a configured failure even when API responses succeed.
Read the video sequence
- The current prompt drafts pull requests. Its API response succeeds and the resulting PR passes required tests.
- A new prompt receives ten percent of new sessions while the current prompt remains the concurrent control. Three stylized nine-to-one traffic patterns illustrate the split.
- Both prompt versions return HTTP 200. PR artifacts travel around the API status cards to the required checks. The control PR passes required tests at 12.9 seconds, but the candidate PR fails at 14.1 seconds: transport success has hidden a task failure.
- The failed required check trips a configured rollout gate. A return signal changes routing to one hundred percent control and zero percent candidate.
- New sessions use the current prompt and its PR tests pass. The candidate is withdrawn, while the failed candidate artifact remains failed. The scene fades out.
About this cut
Production: gpt-6.1-sol.
Revision: gpt-6.1-sol comparative review.
Illustrative PR-drafting scenario; 90/10 routing and compressed observation time are not rollout or statistical guarantees.
A preconfigured required-test failure gate is assumed; one failure does not establish a population failure rate or causality.
Passing tests do not prove complete task correctness; use representative work, sufficient observation and additional signals.
Withdrawal affects new sessions; existing work and side effects require separate cancellation, draining or remediation.
Permissions must be enforced by the harness; persistent prompt instructions do not enforce permissions.
Audio was analysed for signal presence, timing and peak level, but was not physically listened to.
The original author review was a self-review; the completed independent comparison is now saved in reviews/comparisons/canary-deployments.json.
Earlier visual repairs: gpt-6-astra comparative review. Final comparative verification: gpt-6.1-sol comparative review. Original authorship is unchanged.
These animations use sound effects rather than narration. The written guide explains the mechanism and its limits. Model names describe how a cut was made, not a quality ranking.
What the review found
- Baseline 00:05–00:15 shows three 4:1 routing cycles and longer candidate work before comparing p95 900ms with 120ms; it gives the clearest service-flow intuition, but its small version labels and 7.5× badge are harder to read at 350px than the alternatives’ cards.
- Astra 00:06–00:18.6 clears the prior window, collects a new one, then replaces 40/10,000 overall errors with same-window export counts 0/900 versus 40/100. That denominator contrast teaches a release decision the other two cuts do not show.
- Sol 00:11.2–00:19 separates successful API responses from required PR-test failure, sends that failure into STOP, and feeds the gate back to session routing. This is the clearest explicit enforcement path and the most direct application to shared agent prompts; it does not establish a population failure rate.
- Astra 00:23.1–00:28.6 explicitly labels the retained 40% result FAILED WINDOW after withdrawal. Baseline 00:18.5–00:24 and Sol 00:19–00:26 also retain failed evidence while fresh work uses control, but leave its historical scope implicit.
- All three allow the failure and recovery to land. Baseline 00:06–00:15 repeats the latency contrast longest; Astra 00:06–00:13 uses a collecting window before its metric reveal; Sol 00:05.7–00:10.62 packs thirty routing events into a short burst. These are stylized ratios, not adequate measured samples or prescribed observation times.
- Astra 00:14.65 and Sol 00:12.25–00:14.20 confirm that the interrupted reviewer’s border-emphasis and artifact-gutter repairs keep critical text clear. Full encoded checks pass for all three final files, including exact first/last RGB equality; audio has measurable cue energy and safe peaks but was not heard.