Hosting + data-layer spike

Each button measures one thing the port plan assumes about the platform it is running on, and keeps the timings. Run them here, on this origin, then paste the report at the bottom back into the plan.

Target

node · page /

proberesultttfrtotalchunks
Which region / pod answered (health)
Magic Containers injects BUNNYNET_MC_REGION and BUNNYNET_MC_PODID. Run it twice: a new podId means the container was recycled, a new region means your endpoint spreads requests over regions (and each region runs its own worker).
————
H1 async SSR, boundary + pending snippet
ideal: shell+pending first, rows in a later chunk. Svelte 5.57 docs say SSR renders the pending snippet and ignores the rest (no streaming yet) — expect 1 chunk, rows on the client.
————
H1 async SSR, no boundary
same query, no boundary: the response should block on the await, so one chunk with rows and a ~2.5 s time to first byte.
————
H1 raw chunked body
40 chunks 100 ms apart. Buffered proxy = 1 chunk after ~4 s.
————
H1b slow stream (idle timeout) (slow)
20 chunks 1 s apart: a proxy that drops idle connections fails here even though the fast stream passed.
————
H2 server-sent events
alternative to live queries if they turn out to be unusable behind the CDN.
————
H3 request lasting 3 s
baseline for the slow path.
————
H3 request lasting 11 s (slow)
past the 10 s figure in the plan; Gemini OCR calls land here.
————
H3 request lasting 31 s (slow)
past the 30 s cap that would force jobs+polling.
————
H3 request lasting 61 s (slow)
a minute-long request; bulk sends look like this if they ever run in-request.
————
H3 request lasting 91 s (slow)
the outer limit; tells you where the platform timeout really is.
————
H4 upload 1 MB
a phone photograph is 2–8 MB; the capture flow posts it.
————
H4 upload 8 MB
high-quality photo plus voice note.
————
H4 upload 20 MB (slow)
beyond this the plan says: presigned PUT straight to storage.
————
H4 upload 60 MB (slow)
finds the rejection limit and what the rejection looks like (413 from whom?).
————
H6 database info
transport (file vs http) and sqlite version — tells you which numbers below are meaningful.
————
H9 database latency
per-row queries only work if a round trip is single-digit ms. 20 rounds.
————
H6 read-your-writes
if a fresh connection does not see the row it just wrote, every action in the app shows stale data.
————
H6 transactions + batch
capture writes lead + photo rows atomically.
————
H6 concurrent job claim
8 parallel claims must produce 8 distinct jobs — the worker depends on this.
————
H5 worker heartbeats
gaps or many boots while idle = the worker cannot live in the web container.
————
H8 cold start
bootId changes after idle = scale-to-zero; the reported load time is your cold-start cost.
————
H7 outbound egress
Gemini, OpenAI, Tavily and Resend must be reachable from the container.
————
H10 shipped bytes
what a first visit downloads, straight from the served build output.
————
H11 what the CDN forwarded
origin/CSRF and the adapter body limit — a wrong ORIGIN blocks every mutation.
————

Report

Everything above, with origin and timestamps. This is what to paste back.