[iamralch]

Problem

Rust Workers can talk to D1 and Durable Object storage through sqlx, but not to Postgres through Hyperdrive. Hyperdrive speaks the Postgres wire protocol, so sqlx-postgres should just work, except stock sqlx 0.9 panics twice on wasm32-unknown-unknown: connecting ends in rt::missing_rt, since sqlx can only open sockets through tokio or async-io, and every query calls std::time::Instant::now(), which isn't implemented on that target.

Change

• Upstream fix. Postgres: connect over a caller-provided socket, and a wasm-safe Instant, for hosts like Cloudflare Workers transact-rs/sqlx#4426 proposes the two changes: PgConnection::connect_with_socket(&options, socket), and sqlx_core::rt::Instant (web-time on that target). The branch is iamralch/sqlx wasm-custom-socket.
• Workspace patch. Until that ships, the workspace [patch.crates-io] takes sqlx, sqlx-core and sqlx-postgres from that branch, pinned to 4313dae. The patch applies to the whole graph, so D1 and DO build against it too; their test Workers keep their own locks on crates.io sqlx.
• New crate sqlx-cloudflare-hd, with no driver of its own. Laid out like sqlx: one feature per database, none by default, so MySQL (which Hyperdrive also fronts) can come later. postgres::connect(&env, binding), postgres::connect_with(&hyperdrive, &options) and postgres::options(&hyperdrive) return sqlx's own PgConnection/PgConnectOptions, so consumers get sqlx::Postgres: its types, query_as, FromRow, transactions and sqlx::migrate!. The socket adapter and binding lookup are shared.
• TLS goes through the Workers runtime. The socket opens in StartTls mode. The crate sends the SSLRequest itself, upgrades with startTls() when the server accepts, and then hands the socket to sqlx with sslmode=disable so sqlx doesn't negotiate again. prefer/require behave as in libpq. It refuses bytes that arrive after S but before the upgrade (CVE-2021-23222).
• Socket adapter. HdSocket turns worker::Socket's AsyncRead/AsyncWrite into sqlx's readiness-based Socket.
• Consumer setup. getrandom's wasm_js feature is enabled by the crate, so consumers don't have to.
• Test Worker. 10 scenarios: types, 2,000 rows over many reads, unique-violation mapping, rollback and commit, statement reuse, FromRow, migrations, sslmode disable and require, and custom options. run.sh starts a scratch Postgres with SCRAM auth and no TLS and points the binding at it through CLOUDFLARE_HYPERDRIVE_LOCAL_CONNECTION_STRING_HD. The dev shell gains postgresql, and make test-worker gets a third runner.

Not mergeable yet

• Depends on an unreleased sqlx. A git dependency can't go to crates.io. The crate is publish = false and not in the release-please config. When sqlx releases #4426: drop the patch, bump sqlx, and add the crate to release-please.
• Never run against a real Hyperdrive. wrangler dev --local connects straight to Postgres, so the accepted-TLS path (S, then startTls()) and Hyperdrive's own behaviour are unexercised. That needs a real Hyperdrive config on a Cloudflare account.
• No Pool or PgListener. Both need a runtime to spawn tasks. One connection per request is the intended model; Hyperdrive does the pooling.

Verification

• make check lint test doc-check pass, host and wasm32.
• make test-worker passes for all three: D1 29 scenarios, DO 36, Hyperdrive 10.
• On the fork: its tests/postgres/postgres.rs (58 passed, 2 ignored, plus the new it_connects_with_socket) and tests/postgres/migrate.rs (4) pass. With only the Instant change reverted, the Worker panics with time not implemented on this platform.