After a written allow line for each rule, the amarbaro.com rehearsal walk passed 1,301 URLs with 0 failures, 1,005 of them by stated exception, after 1,003 cookie-only failures before it.
The metadata walk over amarbaro.com failed 1,003 of its 1,301 URLs in rehearsal. The record labels all 1,003 as cookie-only. After we wrote an allow line with a reason for each rule, the walk ran with zero failures, and 1,005 URLs pass under a stated exception. The check is now only as good as that list.
What the walk does
Before a switch, we walk every public URL we can list, on a rehearsal copy of the site. The walk asks for each page and records a pass, a failure or an allowed difference. We call it the meta walk. It tests that the joined site answers the same way the old one did, for the addresses that matter to visitors and to search engines.
A walk like this fails on differences a person would never notice. A changed cookie is one. The joined site and the old site set different cookies in some responses, and the walk cannot know which differences are harmless. Someone has to say so.
The first run
The first run on amarbaro.com read FAIL=1003 cookies-only out of 1,301 URLs. The walk recorded no failure in the content. Its only recorded difference was in the cookies.
Trendytails showed the same pattern on a smaller set. Before its allow lines, the walk failed 27 URLs on cookies alone. The cause there was a CSRF token that the joined site sets lazily, which means only once a form is used. That is a real difference, and it is harmless to a visitor who never submits a form. It is also the kind of difference that can hide a real break when nobody reads the list.
The allow lines
Each allow line names a rule and gives a reason. On amarbaro.com the allow line went in with commit 8be6121a. The next run read urls=1301 FAIL=0 ALLOWED=1005. Trendytails went from 27 failures to FAIL=0 ALLOWED=28.
The 1,005 allowed URLs are 77 percent of the set. The other 296 passed with no exception. The allowed URLs are not hidden. They show in the count, and the rule behind each one is in the list.
The record does not say who wrote each reason. We should say that plainly. A reason written by the person who built the joined site is a different check from one written by someone who read the old site.
The walk is a gate, and later a log
Before a switch, the walk is a gate. The runbook calls the walk against the legacy site the T-1 gate, meaning the check that has to pass a day before the switch. After the switch, the check is health only. The smoke script runs with a flag that skips the legacy comparison, so it asks whether the new site answers and does not compare it with the old one.
The same walk ran again on later template updates, and its list grew with each one. On the 2026-11-13 update it covered 378 pages, all of them 200. The seven sitemap URLs for /<lang>/book were allowed as old 404s, because the record shows they already returned 404 on the live site before the update. Each new exception has to be written down with its reason, the same way the first 1,005 were.
What the walk does not show
The walk checks the URLs it asks for, on a rehearsal copy. It says nothing about production paths it did not request, or about pages that only a logged-in user sees. The count of 1,003 comes from the log's label, not from a list of the 1,003 URLs. So "every failure was a cookie" is a statement about how the walk labelled them. We have not reopened each URL to confirm the label.
The allow list is also a judgment. A rule that is wrong for one URL pattern passes every URL that matches it. A count of 1,005 exceptions can hide a bad rule as easily as it shows a good one.
Why keep the list at all
Without the list, the walk shows a wall of failures that tells us nothing. With it, the walk shows a short set of named exceptions, each open to argument. That turns the check from a number into a claim someone can dispute. For a rehearsal, that is the most useful thing it can give us before a switch.
How to prove this wrong
Empty the allow file and rerun the walk on the same copy. If any failure is not a cookie difference, the label is wrong. Then group the 1,005 allowed URLs by the rule that allowed them, open a sample from each group by hand, and check the stated reason against what the page does. A reason that does not match the page is a bug in the list. Until it is fixed, the passes in that group mean nothing.
Provenance
docs/cutover/LOG-cut-kit.md, line 13 (trendytails: FAIL=27 cookies only with a lazy CSRF token, then FAIL=0 ALLOWED=28 after the allow lines) and line 15 (amarbaro.com: FAIL=1003 cookies-only before allow line 8be6121a, urls=1301 FAIL=0 ALLOWED=1005 after).
docs/cutover/amarbaro.com.md, line 4 (the meta walk result).
Two gallery URLs failed in the first rehearsal walk. Mounting the live volume read-only crashed the container's startup, and mounting it read-write would have changed the owner of production files. The walk passed on a scratch copy.
The switches themselves were short. What slowed the night was a refused step on one site, a tunnel waiting for a dashboard invite on another, and a mail provider we had wrong until 23:30.
The cutover runbook for amarbaro.com stamps the database to revision 0001 before the switch. Read in full, the stamp makes the automatic tag rollback unable to start the old image. We found it by reading the runbook, not by a failed deploy.
Komentarze
Brak komentarzy.
Zaloguj się aby skomentować.