Deployment Strategies

Strategies

Feature flags separate deploying code from releasing a feature: the code ships switched off and is turned on for some users later, without a deployment.

Automated canary analysis

Compare the canary's error rate and latency with the stable version, and roll back automatically if it's worse.

import random

random.seed(11)

def error_rate(version, requests):
    p = {"stable": 0.004, "canary-good": 0.005, "canary-bad": 0.030}[version]
    return sum(random.random() < p for _ in range(requests)) / requests

def analyse(canary):
    stable = error_rate("stable", 20_000)
    new = error_rate(canary, 1_000)            # 5% of traffic
    verdict = "promote" if new <= stable * 2 + 0.002 else "ROLL BACK"
    print(f"{canary:12} stable {stable:.2%} vs canary {new:.2%} -> {verdict}")

analyse("canary-good")
analyse("canary-bad")

canary-good  stable 0.42% vs canary 0.40% -> promote
canary-bad   stable 0.41% vs canary 1.80% -> ROLL BACK

Rollbacks and databases

Rolling back code is easy only if the database still works with the old version. Use the expand and contract pattern: first add new columns or tables (both versions work), deploy the new code, and remove old columns only in a later release.