What a Supabase Edge Function cold start actually costs, in milliseconds
Supabase's own engineering blog, not a third party, published the before/after boot-time numbers for Edge Functions: real milliseconds, by SDK, with the two root causes named. Here is what they measured and what a cold function's first request still costs today.
Last checked November 18, 2026.
The before/after numbers, from the people who run it
Two changes produced those drops, per the same post: lazy integrity checking (the runtime used to verify every module's checksum before the module loader ran at all; now it only checks a module's checksum when that specific module is actually loaded), and switching the eszip source-checksum algorithm from SHA-256 to xxHash-3, a non-cryptographic hash chosen for being computationally cheaper to run on every boot.
| SDK in the function | Boot time before | Boot time after |
|---|---|---|
| supabase-js client | 275ms | 25ms |
| OpenAI SDK | 459ms | 57ms |
| Drizzle / node-postgres | 301ms | 83ms |
When a function actually pays the cold-start cost
A cold start happens on a function's first invocation after it has been idle, or after a new deployment — the Deno isolate has to boot and load the function's module graph before the first request can run. A function getting steady traffic keeps its isolate warm between requests and mostly skips this cost; a function called rarely (a cron job, a low-traffic webhook) pays it closer to every time.
A function that imports a heavier SDK (the OpenAI client in the table above being the slowest even after the fix) pays a larger share of its cold-start time loading that SDK's own module graph, not Supabase's runtime itself.
Scribase functions run the same Deno family
Scribase Functions run on Deno, the same runtime family Supabase Edge Functions use, so the same general cold-start shape applies: a function with a lighter import graph boots faster than one that pulls in a large SDK on every cold invocation, regardless of which of the two platforms runs it.
Every Scribase project is $5 a month on Hobby or $25 on Pro with functions included in that price — no separate per-invocation or per-GB-hour function bill stacked on top of the project price the way some platforms meter functions as their own line item.
Sources: Scribase pricing
Questions people ask
What causes an Edge Function's cold start, specifically?
Per Supabase's own post: the Deno isolate booting and loading the function's module graph, with integrity-checking overhead on every module (before the lazy-check fix) and the cost of the cryptographic hash used to verify each module's checksum, before that hash was swapped for a faster non-cryptographic one.
Did the fix remove cold starts entirely, or just shrink them?
Shrink them. The numbers in Supabase's post are boot-time reductions (e.g. 275ms to 25ms), not a claim that cold starts no longer happen at all.
Does a heavier SDK import always mean a slower cold start?
The three examples in Supabase's own benchmark support that pattern directionally — the OpenAI SDK case stayed the slowest of the three even after the fix — but the post does not claim a universal rule covering every possible import.
Is cold start the same thing as a function timeout?
No — a cold start is time spent booting before a function can start handling a request; a timeout is the function running too long once it has already started. See the timeout fix guide for the separate wall-clock and CPU-time limits.