tractor/ai/tpt-backends/README.md

57 lines
2.8 KiB
Markdown

# next-gen `tractor.ipc` transport backend plans
Implementation specs for three prospective `.ipc` transport
backends, written so each can be worked independently (by a
different model/provider) without design or lib-selection drift.
**Read [`00_shared_backend_contract.md`](./00_shared_backend_contract.md)
first** — it is the normative description of what a `tractor`
transport backend *is* as of `main@83b34884` (the backend
duck-type, the 10-item registration checklist, the test-harness
plumbing, the code-style rules). The three plans assume it and
document only their own deltas.
| plan | issue | dep | size | lands |
| --- | --- | --- | --- | --- |
| [01 — TIPC](./01_tipc_backend.md) | [#378] | **none** (stdlib) | small | **landed**, PR [#493] — see the [handoff](./01_tipc_HANDOFF.md) |
| [02 — QUIC/`iroh`](./02_quic_iroh_backend.md) | [#353] | `iroh` (uniffi FFI) | large | needs a prep PR |
| [03 — `wg` bindspace](./03_wg_tunnel_bindspace.md) | [#482], [#443] | `pyroute2` | medium, 3 layers | layer A now |
Headline conclusions:
- **TIPC is the cheap win.** Verified: `trio.SocketStream` and
`trio.SocketListener` are address-family agnostic (only
`SOCK_STREAM` + a trio socket), and CPython ships `AF_TIPC` +
23 `TIPC_*` constants. So the backend is ~one module of
contract boilerplate, zero new deps, and it buys
*kernel-native* service discovery: `bind()` publishes,
`connect()`-by-name resolves — no registrar in the loop.
(`modprobe tipc` is required; hard-gate everything.)
- **QUIC's cost is entirely in two adapters**, not in QUIC. The
`iroh` python bindings are `uniffi`-generated asyncio, but the
asyncio dependency is confined to *one* future-poll callback —
a ~40-line `trio` bridge (`TrioToken.run_sync_soon`) replaces
it. The second cost is that an iroh listener isn't a socket,
which needs a small, independently-reviewable prep PR to
`_server.py`/`_types.py`.
- **WireGuard is not a transport.** It's an iface-layer tunnel,
so it belongs as a *nested bindspace* (`TunnelledAddress` +
`open_bindspace()` `@acm`s) wrapping whatever L4 tpt is in
use — which is also what finally implements the long-spec'd
`Address.namespace`, and what generalizes to
`veth`/`vxlan`/`gre`.
Ordering rationale: plan 01 first as the cheap proof the
table-registration story generalizes to a genuinely new proto;
plan 03 layer A is already deployable-today doc/example work;
plan 02 last (and gated on its prep PR). Plans 01 and 02 both
want the same `Address.rebind_from_sockname` gate — whichever
lands first ships it.
[#378]: https://github.com/goodboy/tractor/issues/378
[#353]: https://github.com/goodboy/tractor/issues/353
[#482]: https://github.com/goodboy/tractor/issues/482
[#443]: https://github.com/goodboy/tractor/issues/443
[#493]: https://github.com/goodboy/tractor/pull/493