Regional LCP Regression on a Flash-Sale Storefront
Lumen Retail's storefront is a React single-page app instrumented with browser monitoring and a scripted browser synthetic monitor running from three regions. On the first day of a flash sale, the synthetic reports a largest-contentful-paint of 2.4 s from us-east, while real-user data shows p75 LCP of 5.9 s in ap-southeast and 2.1 s in us-east. JavaScript error rate is unchanged. Navigation timing shows the HTML document arriving within 300 ms in every region, but the LCP element on mobile, the hero image, is served by a third-party CDN that returns cache-miss responses from the Singapore point of presence. Session replay shows users tapping the hero area before it renders and then abandoning. A separate A/B testing flag service is render-blocking in the critical path for all users, adding roughly 400 ms. Marketing wants the CDN contract escalated; engineering wants to delete the A/B flag.
1. Explain why the synthetic monitor and real-user data disagree, and state which signal should drive the ap-southeast investigation.
Read the worked answer
Trust RUM for the regional regression; treat the synthetic as a smoke test. The synthetic runs one browser profile from cloud infrastructure with a different cache topology and, critically, may not be assigned the same A/B variant as real sessions. It measures the app as an operator sees it, not as a phone in Jakarta sees it. A p75 of 5.9 s across many real sessions in ap-southeast describes the population that is actually failing, so it wins the argument.
Data to pull before acting: LCP split by country, device class and connection type; resource timing for the hero image decomposed into DNS, TLS, time to first byte and download, split by cache hit versus miss; CDN cache-hit ratio per point of presence; and the resource initiator to confirm the hero image is genuinely the LCP element rather than a webfont or background asset. If cache-miss TTFB from Singapore is 3-4 s while us-east shows hits, the divergence is fully explained and the synthetic simply never experienced the miss. That comparison is also the evidence Marketing needs for a contract conversation, framed as a cache-hit-ratio problem rather than a vague performance complaint.
2. Rank the two suspected causes using the evidence given, and specify the measurement that would settle it.
Read the worked answer
Ranking: the CDN cache-miss is the differentiator; the A/B flag is a global floor.
- The regression is geo-specific. A render-blocking flag in the critical path affects us-east and ap-southeast equally, so it cannot produce a 3.8 s gap between them. It can only raise the absolute number in both regions.
- The LCP element is the hero image, and that image is the exact resource showing cache misses from the Singapore PoP. The blamed resource and the measured element match.
Measurement that settles it: a controlled test from a Singapore vantage point, on a throttled mobile profile, with a cold cache, loading the page with the flag forced off. If LCP remains well above 4 s, the CDN is the dominant term. Then repeat with a warmed CDN cache to size the recoverable portion. Separately, measure the flag's marginal cost with the CDN already warm to decide whether removing it is worth the experimentation cost. The two fixes are not exclusive, but they belong in different backlogs: CDN cache policy is a vendor and configuration action, the flag is an engineering action, and only the first explains why ap-southeast users abandoned during the sale.
3. What monitoring changes would make this class of regression visible before customers abandon?
Read the worked answer
Close the gaps the synthetic left open.
- Add a synthetic from the ap-southeast region using a mobile device profile, throttled network and a cold cache, asserting an LCP budget. The current monitor's region set and warm path are why it reported 2.4 s while users saw 5.9 s.
- Alert on RUM p75 LCP segmented by region and device class, not on a global average that the healthy us-east traffic masks. Alert on the segment, with a short evaluation window during sale events.
- Instrument the A/B variant as a custom attribute on browser events so any future regression can be attributed to a variant instead of debated.
- Track third-party resource timing as its own signal, including CDN cache-hit ratio per PoP, and put it on the sale-readiness dashboard.
- Define an SLO on LCP for the storefront and alert on burn rate, so a slow drift is caught before a flash sale amplifies it.
Session replay is the qualitative backstop, not the detector; the detector has to be a segmented real-user alert on the metric that maps to abandonment.