supabase / supabase-flutter / 35344799042

90%

push

github

feat(supabase)!: keep idle connections open for 60 seconds by default (#1855)

## What

A `SupabaseClient` created without an `httpClient` now builds an
`IOClient` whose idle timeout is 60 seconds instead of the 15 second
`dart:io` default. The web build keeps the plain default `Client`. The
`supabase` README gains a server-side section that recommends one
long-lived `SupabaseClient` per process, shows how to act on behalf of a
user from it by setting the `Authorization` header per call, and covers
the idle timeout and disposing. The migration guide documents the
`runWithClient` change.

## Why

A Dart server talking to a remote Supabase project reported roughly
twice the latency of a direct database connection. Benchmarks showed the
SDK itself is not the cause: on a warm connection it is within 0.2 ms of
raw `package:http`, and there is no hidden extra round trip. The
doubling comes from connection setup, which happens whenever the pool
has no open connection.

Measured against a hosted project in eu-north-1 over TLS:

| Scenario | p50 |
| --- | --- |
| Reused connection | 24.8 ms |
| New TLS connection | 58.8 ms |

`dart:io` closes an idle connection after 15 seconds, so a server with
sporadic traffic pays the handshake on most requests. The servers on the
other end keep connections open much longer:

| Gateway | Idle connection still open after |
| --- | --- |
| Hosted (Cloudflare) | 300 s, the longest gap tested |
| Local CLI stack (Kong) | 60 s, closed by 90 s |

60 seconds is the largest value that stays below the self-hosted cutoff,
so a reused connection is never written to just as the gateway closes
it. Locally, two requests 16 seconds apart used one connection after
this change and two before it.

## Breaking change

The default client is built directly instead of through the `Client()`
factory of `package:http`, so a client installed with `runWithClient` is
no longer picked up. Pass it as `httpClient` instead. A script that
never calls `dis... (continued)

4 of 4 new or added lines in 2 files covered. (100.0%)

7866 of 8692 relevant lines covered (90.5%)

4.65 hits per line