cable_room 0.6.2.beta3 → 0.7.0.beta1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 5984faba13ce4bcbfc7cdcde1eee9a268ca8ea59227efd2cbe41532d67892782
-  data.tar.gz: 651af8589e9709f27e3cfee877fe2359c931919d3c853c74d097147b245d37f1
+  metadata.gz: 890aa545d8211663c2270322da9440b0dbd685e7cdcfa9858df04390d06e7533
+  data.tar.gz: 074e6c49a9886c90483500bdb61cdf40fa4d3c9e83ddc583415cc9ac051e90f6
 SHA512:
-  metadata.gz: 9f85aa97a935a9cc11049b5ce408f06cacca2ff1309800e13b4b4a9b750c68612d3991410f081e5c99a43b0282bbc42071418d856d728c3bda03381ec48ba85c
-  data.tar.gz: f94b0ac04d23f0bbb411b7f1e9432b09f778756046bcc6b6f3ed02557ca1020f97dc62e6c238b4259d581a92647cba726251037d2b59fd2c93c884c139a59ed4
+  metadata.gz: 3304bd77c3d455a2b867d83d82d4985ff7e1a3ec10bd80ec6a097f50be040fadbbaac501d7f26234384ba50dbe9f09864c92e1114d8cecff75d308ae2a770261
+  data.tar.gz: 408c73ab717cd8aaa5a7fbc8ce6b5d209ee2169fe50a731ac8b6d7b2b417b61e64c5a9ceac55155e07e677f6b3a952bb56edd4043a16b624bb264cd6ca1fb21d

data/CHANGELOG.md

@@ -0,0 +1,122 @@
+# Changelog.
+
+## 1.0.0
+
+Rooms can now run in their own process pool, and browsers can be served by anycable-go. Every
+Room definition from 0.6 works unchanged. The channel and deployment contracts do change; the
+README's "Upgrading from 0.6 to 1.0" section is the checklist.
+
+### Breaking.
+
+- **Clients must send `hello`.** A membership no longer announces itself to the Room when the
+  channel subscribes. The browser performs `hello` once ActionCable confirms the subscription
+  (`connected()` in the JS client), and only then does `port_connected` go out. There is no
+  fallback: a client that never says `hello` receives broadcasts on the shared stream, but the
+  Room never sees it and everything it sends is dropped. Browser tabs open across the deploy
+  speak the old handshake and stay deaf until reloaded.
+- **`ChannelBase`, `ChannelTracker`, and `DummyConnection` are gone.** Rooms no longer run inside
+  a synthetic ActionCable channel. `CableRoom::Host` (one per process) and `Host::Runner` (one
+  per Room) own the work queue, worker pool, scheduler, Redlock renewal, watchdog, and shutdown.
+  Internal names apps may have reached for moved with them:
+  - inside a Room, `@cable_channel` is now `@runner` (a `Host::Runner`);
+  - the `CableRoom.error_handler` context key `channel:` is now `runner:` (with `room:`,
+    `room_class:`, and `room_key:` alongside it);
+  - `Room#params` is gone (it delegated to the synthetic channel, which had none);
+  - `ChannelTracker::BEAT_INTERVAL` is `Host::BEAT_INTERVAL`;
+  - `Room::ChannelAdapter` is `Room::HostAdapter`, and the `MyRoom::Channel` constant no longer
+    exists;
+  - `CableRoom::Room.locally_open_rooms` and `MyRoom.locally_running_instances` read from the
+    Host and return `[]` in a process that hosts no Rooms.
+- **Provisioning happens over the bus.** `join_room(..., create: true)` never starts a Room in
+  the member's process. It publishes a `provision` request (at join and on every ping until the
+  Room acknowledges the port), and one of the hosts starts the Room after a load-weighted
+  delay. `Room::Base.ensure` still starts a Room directly, but only in a process that hosts
+  Rooms: in a `room_host = :remote` web process it raises `CableRoom::Host::NotHosting`.
+- **Member→room traffic rides cable_room's own Redis.** Everything a member sends (`to_room`,
+  `port_connected`, `port_ping`, `port_disconnected`, custom ports) and `Room::Base.send_message`
+  publish on `CableRoom::Bus`, over the `CABLEROOM_*` connection. The Room side no longer
+  touches ActionCable pubsub for inbound. That connection must be reachable from every web
+  process and every rooms host.
+- **`redis` is a runtime dependency** (`>= 5.0`). It was a development dependency; rediconn
+  needed it all along, and the bus uses it directly.
+- **`PortManagement::PORT_TIMEOUT` is 45 seconds**, up from 30. A member that stops pinging
+  keeps its port 15 seconds longer.
+- **`room_closed` always carries a reason.** In 0.6 a close the host decided (process exit,
+  watchdog timeout, lost lock) sent `reason: nil`; only `shutdown!` from inside the Room set one.
+  Now the runner sets it too: `"Server shutting down"`, `"Watchdog timeout"`, the signal name on
+  a `cable_room server` drain, and the failure on a migration that closed instead of moving.
+- **Instrumentation payloads.** `error.cable_room` and the error handler get the new context
+  keys above. Every 0.6 event still fires with the same name.
+
+### Added.
+
+- `CableRoom.configure` / `CableRoom.config` (`CableRoom::Config`) with five knobs:
+  `room_host` (`:inline` | `:remote`), `broadcaster` (`:action_cable` | `:anycable`),
+  `provision_delay_ms` (20), `drain_timeout` (10 minutes), and `handoff_timeout` (10 seconds).
+  `CABLE_ROOM_HOST` and `CABLE_ROOM_BROADCASTER` override the first two, even when the block sets
+  them. Unknown values raise at boot and name the valid options. Defaults reproduce 0.6.
+- `CableRoom::Broadcaster`: room→member traffic goes through `ActionCableBroadcaster`
+  (`ActionCable.server.broadcast`, the default) or `AnyCableBroadcaster` (`AnyCable.broadcast`).
+  Stream names are identical in both. `anycable-rails` is optional; selecting `:anycable`
+  without it raises `Broadcaster::MissingDependency` naming the gem.
+- `CableRoom::Bus`: Redis pub/sub and the handoff storage on the `CABLEROOM_*` connection.
+  Channel names: `cr:{RoomClass:key}:in` (and `:in:{port}` for custom ports), `cr:provision`,
+  `cr:{RoomClass:key}:handoff`, `cr:{RoomClass:key}:snapshot`, `cr:{RoomClass:key}:adopted`.
+  `Room::Base.inbound_channel(key, port)` names a Room's inbound channel.
+- `CableRoom::Host`: `Host.current` (nil when the process hosts nothing), `Host.instance`,
+  `Host.start!`, `Host#open_ports` (the load gauge to scale on), `Host#open_rooms_count`,
+  `Host#drain!`, `Host.after_drain`, `Host::NotHosting`.
+- `cable_room server` (`exe/cable_room`, `CableRoom::CLI`): boots the Rails app in the current
+  directory (or `--require PATH`), refuses to run unless `room_host` is `:remote`, hosts Rooms
+  until SIGTERM or SIGINT, and migrates open Rooms to peers on the way out. `--workers N`
+  (default: core count) preloads the app and forks N single-GVL workers through
+  `Host::Supervisor`, which replaces a worker that dies (with a growing delay for one that dies
+  at boot), relays SIGTERM and SIGINT, and waits for every worker. `--version` prints the
+  version. `CableRoom.after_fork!` resets the bus, lock manager, Redis pool, and Host in a
+  forked child.
+- Load-weighted provisioning (`CableRoom::Placement`): every host hears each `provision`
+  request, waits `provision_delay_ms × (open_rooms + jitter)`, then races for the Redlock.
+  Exactly one wins. A draining host never claims.
+- Snapshot and restore: optional `snapshot_state` / `restore_state` hooks on a Room,
+  `restored?`, `CableRoom::Snapshot` (version 1, gem-owned ports, users, reaper deadlines, and
+  `app_state`), `Snapshot::NotSerializable` raised at snapshot time for non-JSON state, and
+  `Host#restore_room`. A restored Room runs `restore_state` instead of `startup` when it defines
+  it, and never broadcasts `room_opened`.
+- Live migration on planned shutdown (`CableRoom::Migration`): freeze → snapshot → offer (Redis,
+  60 second TTL) → peer adopts → relay and replay with per-stream markers → let go. Members keep
+  their streams and lose nothing. A Room nobody adopts within `handoff_timeout` closes with
+  `room_closed`; `drain_timeout` forces whatever is still waiting through. New lifecycle states
+  `:freezing` and `:frozen`.
+- Channel actions `hello` and `ping` on every `RoomMember` channel (and so on
+  `RoomProxyChannel`), plus the server-side `hello_room_memberships`, `RoomMembership#hello!`,
+  `#hello_received?`, `#accepts_input?`, and the protected `room_memberships` reader.
+- AnyCable channel state: `CableRoom::MembershipStore` keeps each membership's identity in
+  anycable-rails' `state_attr_accessor`, so `hello`, `receive`, `ping`, and `unsubscribed` work
+  across anycable-rails' per-call channel instances. Under plain ActionCable nothing changes.
+- New `ActiveSupport::Notifications` events: `hello_received.cable_room`,
+  `provision_requested.cable_room`, `provision_claimed.cable_room`, `room_snapshotted.cable_room`,
+  `room_restored.cable_room`, `host_draining.cable_room`, and `room_migrated.cable_room`.
+- CI: an `e2e-anycable` job runs `spec/e2e` against a real `anycable/anycable-go:1.6` under both
+  Appraisals. Locally it runs only with `CABLE_ROOM_E2E=1`.
+- `CHANGELOG.md` and `exe/` ship in the gem.
+
+### Changed.
+
+- `Room::Base.room_port_key` builds the stream name itself (`"RoomClass:key:port"`, records by
+  GlobalID). The format is the same one ActionCable produced in 0.6.
+- The ActionCable `restart` hook (development code reload) stops Rooms through the Host.
+- `RoomMembership#stream_port` takes `auto_close:`; the room-side `stream_port` takes `on_live:`.
+- The `Railtie` no longer declares empty `rake_tasks`, `console`, and `runner` blocks.
+- The README describes the 1.0 architecture: configuration, deployment modes, `cable_room
+  server`, provisioning, the hello handshake, AnyCable's browser contract, migration, every
+  instrumentation event, and the 0.6 → 1.0 upgrade guide.
+
+### Fixed.
+
+- The gemspec declares `redis`, which rediconn loads but never listed. An app that didn't have
+  `redis` in its own Gemfile could fail to boot with 0.6.
+
+## 0.6.2.beta1
+
+The last release before distributed rooms. Rooms ran inside a synthetic ActionCable channel in
+every process that joined them.