cloudflare-preview-benchmark#

The app behind the PR-lifecycle animation on the alchemy.run landing page: an R2 bucket (Photos), a KV namespace (Sessions), and an Effect-native Worker (Api) bound to both. It is deployed per pull request as a preview stage, tested over HTTP, and destroyed.

Route	Behavior
GET /photos	JSON gallery listing every object key in Photos
PUT /photos/:key	Uploads the request body to Photos
GET /session	Starts a session and stores it in Sessions; with an x-session-id header, reads it back

This example creates billable Cloudflare infrastructure. Destroy it when you are done.

Run it#

bun alchemy deploy --stage pr-147
STAGE=pr-147 bun test
bun alchemy destroy --stage pr-147

test/preview.test.ts deploys through the Alchemy test harness in beforeAll. That deploy is a no-op when the stage already exists. Without STAGE, the harness uses its default stage, test_$USER. The stack is destroyed after the tests only when CI is set.

Photos sets forceDestroy: true because the upload test leaves an object in the bucket, and R2 refuses to delete a non-empty bucket.

Benchmark#

The "full picture" animation on the landing page (website/src/components/marketing-islands/PRLifecycle.tsx) replays these timings in real time for the deploy, test and destroy logs. Re-run the benchmark when Alchemy gets faster or slower, and update the T_* constants at the top of that file to match.

Run it#

You need a Cloudflare account and an Alchemy profile with credentials for it (bunx alchemy login --profile <name> creates one).

# from the repo root, once
pnpm install

cd examples/cloudflare-preview-benchmark
ALCHEMY_PROFILE=<name> bun run bench

bun run bench runs scripts/bench.sh, which times one full preview lifecycle and prints one line per phase:

deploy   11.62s
url live 0.41s  https://api-….workers.dev
test     4.53s
destroy  3.80s
total    20.36s  (logs: .alchemy/bench/<timestamp>-*.log)

• STAGE picks the stage name (default pr-147). Use a unique one if several people share an account.
• Each phase's full CLI output is written to .alchemy/bench/. The per-resource breakdown below comes from the deploy and destroy logs.
• The script always destroys the stage at the end. If it fails midway, clean up with bunx alchemy destroy --stage <stage>.
• Run it three times and take the median; the first run after an install is slower while Bun warms its cache.

Results#

Median of three full cycles, wall clock:

Phase	Time
Deploy (create 3 resources)	11.6s
· CLI startup, stack import, state load	1.2s
· plan	1.3s
· Sessions created	0.7s
· Photos created	3.6s
· Api created (waits for Photos, bundles 2.2s, uploads and enables workers.dev 3.3s)	9.4s
Worker URL returns 200 after deploy	+0.4s
Tests, plain fetch against the URL	0.7s
Tests, through the Alchemy harness (bun test)	4.5s
· GET /photos renders the gallery	0.24s
· uploads a photo to R2	0.28s
· session survives a reload	0.26s
Deploy with no changes	4.0s
Deploy after a one-line Worker change	7.6s
· new code served at the URL	+0.2s
Destroy	3.8s
· Api deleted	0.7s
· Photos emptied and deleted	1.4s
· Sessions deleted	0.4s

Most of the harness test time is the no-op deploy in beforeAll: it re-reads every resource from Cloudflare (about 3.2s). The same three tests take 0.7s without it. Running the CLI under Node instead of Bun added about 0.7s per command.

These numbers come from one Linux machine and one Cloudflare account, using alchemy 2.0.0-beta.79, effect 4.0.0-rc.117 and Bun 1.3.13. Treat them as indicative.

The landing-page animation also shows a GitHub.Comment posting the preview URL. It is not part of this example or the benchmark because it needs a GitHub token and a target PR.