The siege: load-testing TYPO3 13 on Google Cloud before 414 dealer sites move
November 2026 · Bouletteproof engineering · Staging figures
GT2, a French digital agency, runs the websites of Polaris's off-road vehicle subsidiaries and of their dealers. We are moving them from TYPO3 8.7 to TYPO3 13.4 on Google Cloud. Before anything moves, the new platform has to carry the traffic the old one carries, floods included. So we laid siege to it: a separate machine sending the new platform production's real traffic levels, then the peaks.
The first steps went badly. One setting and one cache layer turned it around. Every figure here comes from load tests on staging in November 2026. Nothing runs on TYPO3 13 in production yet, and we will add production figures when New Zealand and Australia switch over.
Bouletteproof has been GT2's engineering provider for more than ten years. This is our own work.
The traffic to carry
We sized the test on 30 days of requests reaching the current servers, counted per second, floods included, because floods are traffic too.
| Level | Requests per second |
|---|---|
| A typical minute (median) | 12 |
| Average | 16.5 |
| 95% of minutes | under 37 |
| 99% of minutes | about 100 |
| 99.9% of minutes | 400 to 432 |
| The highest minute | 1,000 |
The platform under test
On TYPO3 13.4 the sites run in containers on Cloud Run, with a Cloud SQL MySQL database, Memorystore for Redis, and a Cloud Storage bucket for media. The containers keep nothing of their own, so Cloud Run can add and remove them as the load changes.
The siege ran against staging: a database with one CPU, Cloud Run allowed up to 20 instances, and the load sent from its own virtual machine. Production answer times did not move during any run.
Bad: the first steps
We started with one instance serving cached pages and raised the rate step by step. At 20 requests a second, 37% were answered and the rest timed out. At 40 a second, under 1% were answered. We stopped after three steps and looked at what each request was doing.
TYPO3 was opening a new encrypted connection to Redis for every cache, on every request, and each one cost about 55 ms. We changed it to keep the connections open, one per Apache worker. Same page, fresh container, before and after:
| A cached page | Before | After |
|---|---|---|
| Time per request, one at a time | 133 ms | 49 ms |
| CPU per request | 112 ms | 34 ms |
| Requests per second, 10 at once | 13 | 43 |
| A 404 page | 127 ms | 41 ms |
One instance with one CPU went from about 8 cached pages a second to 25 to 30. At 20 a second, every request was now answered, half of them within 52 ms.
What we checked and left alone: PHP's opcode cache was working. Keeping database connections open gave a small extra gain, from 49 to 44 ms, but we did not apply it, because with Apache's default of 150 workers per instance it could run the database out of connections. Apache's defaults and a 397-line .htaccess file read on every request are not tuned yet, and 34 ms of CPU is still a lot for a cached page.
Better, not good enough
Then the full siege, with up to 20 instances and still no CDN. Up to 100 requests a second, everything was answered. At 200 a second, the database reached 100% CPU. At production's 400-a-second peaks, 83 to 88% of cached pages were answered, and only 32 to 38% of requests in the production mix got the page they asked for. At 1,000 a second, the production mix failed.
In every failing case the limit was the one-CPU staging database, not Cloud Run.
Floods are their own case. With the scanner block production already uses, a flood of requests for pages that do not exist went, at 432 a second, from 26% timed out, with half waiting more than 7 seconds, to every request answered, half within 40 ms. At 1,000 a second, 37% still timed out. Our test flood was harsher than the real ones: only about 40% of its addresses match the blocking rules, while the flood of 4 November was almost all requests for .env files, which the rules stop completely.
How we cache HTML on Google Cloud
The database was the ceiling, so the answer was to stop sending it questions it had already answered. We put the HTML itself in Cloud CDN, in front of Cloud Run.
- A page that can be cached is served from Google's edge. Form pages, the stock list and 404s are not cached.
- Browsers are told to keep nothing, so a change never waits on a visitor's own cache.
- Saving a page in TYPO3 purges it at the edge, and Google completed each purge in about a second. A full cache flush in TYPO3 purges everything.
- Everything else expires after five minutes at the edge. News and vehicle records edited outside a page wait for that.
- The cache key includes the visitor's region, so each region can get its own consent banner. Google accepted the setting; we have not yet proven it from two countries.
Redis still holds TYPO3's own caches behind it. The CDN keeps whole pages; Redis keeps the pieces TYPO3 builds them from.
Good: the same siege with the CDN
| Step | Without the CDN | With the CDN |
|---|---|---|
| Cached pages, 100 a second | all answered, 95% within 120 ms, database at 73% | all answered, 95% within 3 ms, database at 13% |
| Cached pages, 400 a second | 88% answered, 20 instances | all answered in 1 to 3 ms, 5 instances, database at 15% |
| Production mix, 100 a second | all answered, 95% within 116 ms | all answered, 95% within 66 ms |
| Production mix, 432 a second | 38% got the expected page | all answered, 95% within 62 ms |
| Production mix, 1,000 a second | fails | all answered, 95% within 107 ms, 14 instances |
Cached pages now stop at the edge and never reach Cloud Run or the database. Every production level passes, including the highest minute, which failed in every earlier run.
What is still open
- At the top step the database reaches 100% CPU again, from the 404s in the mix that still reach TYPO3. That is the next ceiling.
- The stock list is not cached at the edge, so it is still slow and bound by the database.
- Four scenarios have not been re-run with the CDN on: one instance alone, right after a cache flush, a cold start, and the stock list.
- The region in the cache key still needs a check from a second country.
We will add the production figures when New Zealand and Australia switch over to TYPO3 13.
Engineering note from Bouletteproof. Related: The blast radius of a change and The boring parts that ship.