Migrating from the Old Site

The old site is live and has bookmarks, search results and links from other places pointing at it. A migration that just switches it off turns every one of them into a 404. This chapter does the opposite: it measures exactly what the old site answers, finds what is missing on the new sites, suggests redirects for a person to approve, and moves one site at a time. The numbers come from the real old site and the real new database.

Step 1: Measure, Don't Remember

The old site is a folder of built files: /hungary/x/y/ is the file hungary/x/y/index.html. That folder is the complete list of old addresses, so a command reads it and asks each new site about each address:

That is the useful result: 97.5% of old pages already have the same address, because the folder structure was kept, and 80 do not. Reading the list showed why:

• 32 moved or renamed lessons, for example hungary/hungarian-language/… became hungary/hungarian-lessons/…, and ten Japanese lessons that used to sit under ai/claude-tools/claude-lessons/ are now under japan/.
• 23 _print pages: the old site had a printable twin of many pages, and the new one does not.
• 25 others: index pages of renamed folders, old “tentative” course folders that were removed, and japan/kanji-tiles.

The 166 missing files are mostly older copies of Node.js and Express solution files (120, in text-files/ folders the courses no longer use) and 36 in projects/nextjs-rebuild/solutions. The command exits with an error until the list is empty, so it can be run again after every change.

Step 2: Redirects That Someone Reviews

An old address that moved needs a permanent (301) redirect. Redirects are decisions, not text imported from files, so they live in the database, are editable in the admin, and come from a CSV. The matching is deliberately modest:

1. If exactly one new page has the same file name (ignoring case, and - versus _), propose it.
2. If several do, list them for a person. If none do, say so.
3. A name ending in _print falls back to the page without it.

A file name is not proof of the same lesson, so every row of the CSV starts with an empty decision column, and only rows where you write yes are loaded. Loading a file with no “yes” does nothing (“0 created, 0 updated, 32 skipped”). On the real data the first version proposed 32 of the 80, and the _print rule added 23, giving 55 proposals, none ambiguous, 25 with nothing to suggest.

The second row is a redirect to a different site: the old ai address becomes a full address on the languages site, built from the same site-address setting as the global bar. The redirect is tried only when there is no page and no folder listing at that address, so a redirect can never hide a real page (a test checks this), and a redirect belongs to one site.

Step 3: Links Between Sites

A page written for one site that links to /linux/… with a plain absolute path breaks after the split: that path is on the systems site. Editing thousands of files, and writing a host name into them, would be the wrong fix. Instead the link is rewritten when the page is shown: a link into another site's folders gets that site's address in front, taken from the same setting that differs between development and production. The stored page is unchanged (a test checks this).

On the real content the rewriter found nothing to do: of 11,691 internal links, every one points into its own site. That is a good result for the site map, and it means the rewriter is exercised only by tests today. It leaves alone external links, //cdn… links, relative links, #fragments, /search/ and static files, and keeps query strings.

Broken Links

A second tool reads every stored page and checks each absolute internal link against the pages, folders, redirects and mirrored files. On the real data (3 seconds): 11,715 links checked, 483 broken.

Step 4: One Site at a Time

While the old site still answers on osztromok.com, a site moves by telling the old host's Apache to send that site's folders to the new subdomain. A command prints the line for one or more sites:

Without Apache, I ran the same pattern as a Python regular expression over every old address: of 8,892 page and file addresses that a site owns, none was matched by the wrong site or by two sites, and none of the 144 addresses that no site owns was matched. Tests check that no folder is claimed by two sites, that /linuxfoo/ does not match the systems rule, and that sidebar pages follow their subject (for example /sidebar/football/ goes to the humanities site).

Hands-On Exercises