Changelog

Copy Markdown View Source

All notable changes to this project are documented here. Format based on Keep a Changelog, versioning follows Semantic Versioning.

[0.5.2] - 2026-10-28

Added

  • Terminate a user's connections (POST /apps/:app_id/users/:user_id/terminate_connections, or Socketelx.terminate_user_connections/3). Closes every connection of a signed-in user on every node, for logout or a ban. Works across Redis and Erlang cluster nodes.

[0.5.1] - 2026-10-23

Fixed

  • REST request bodies are capped at 1 MB; a bigger one gets 413. Before, the body was read in full (up to Plug's 8 MB) before the signature was checked, and anything larger crashed the request with a 500.
  • WebSocket frames are capped at 100 KB; a bigger one closes the connection (code 1009). Cowboy's default is no limit, so one client could make the server buffer an arbitrarily large frame. Real Pusher frames are a few hundred bytes (client events are limited to 10 KB by Pusher).

[0.5.0] - 2026-10-23

Added

  • Redis Cluster support. Set redis_cluster: true (SOCKETELX_REDIS_CLUSTER=true) and point redis_url at any node of the cluster; redis_cluster_nodes (SOCKETELX_REDIS_CLUSTER_NODES) adds more places to look. All of an app's keys carry one hash tag, so they stay in one slot and the atomic presence and occupancy scripts keep working. Socketelx finds the node that owns the slot, follows it through failovers and resharding, and moves its pub/sub subscription with it. Tested against a real 3-master, 3-replica cluster, including a manual failover and killing the owning master under load. Standalone Redis, Sentinel and single-endpoint managed services work as before and are unchanged. See Multi-server.
  • CI runs the Redis tests a second time against a real Redis Cluster.

Notes

  • One app's state lives on one shard. Several apps can spread over shards.
  • Delivery between nodes pauses for a few seconds during a Redis failover, like any other Redis interruption. Delivery is still at-most-once.

[0.4.3] - 2026-10-23

Changed

  • Setting a limit to 0 now means "no limit" (rest_rate_limit, client_event_rate_limit, max_connections, max_connections_per_ip, in options and in the matching SOCKETELX_* variables). Before, 0 refused every request or connection. Negative numbers also mean no limit. If you used 0 on purpose to block everything, use 1 instead.
  • bench/load.mjs: new --workers N runs the generator in several processes, so a single Node process no longer caps the test at about 50k messages per second.
  • Going to production guide: new section on CPU cores and Erlang schedulers.

[0.4.2] - 2026-10-17

Changed

  • Idle connections use about 3.5 times less memory. WebSocket processes now hibernate while they wait: 10,000 subscribed connections went from about 570 MB to about 200 MB in a measurement (about 15 KB each). No change in latency or CPU per message.
  • bench/load.mjs: accepts --option=value, rejects unknown options, sends keep-alive pings like real clients, checks credentials before starting, reports why sends and connections failed, and takes --targets a:6001,b:6001 to test several nodes.

[0.4.1] - 2026-10-17

Added

  • Guides: Going to production (open-file limits, proxy and OS settings, monitoring) and Capacity planning (what a connection and a message cost, with an 8 GB example).

[0.4.0] - 2026-10-16

Added

  • Several apps in one server (apps: [...] / SOCKETELX_APPS JSON). Each app has its own id, key, secret and settings, and is fully isolated (channels, presence, webhooks, limits, metrics). Metrics gain an app label when there is more than one app. Single-app setups are unchanged. Socketelx.trigger/5 and authorize_channel/4 take app: on multi-app servers.
  • Allowed origins (allowed_origins / SOCKETELX_ALLOWED_ORIGINS): exact origins, bare hosts and *. wildcards. Default is to accept every origin. Refused connections get pusher:error 4009.
  • GET /apps/:id/channels/:channel/users (presence channel members).
  • User sign-in: pusher:signin, events to a user through #server-to-user-<id>, and Socketelx.authenticate_user/4.
  • max_connections_per_ip and client_ip_header (error 4100).
  • webhook_queue_path: failed webhooks are kept on disk and sent again after a restart.
  • Built-in TLS (tls_certfile / tls_keyfile).
  • Chat example demo GIF.

[0.3.0] - 2026-10-16

Added

  • Experimental: Erlang cluster mode (cluster: :erlang / SOCKETELX_CLUSTER=erlang). Several nodes share events, presence, channel occupancy and webhooks over Erlang's own clustering, with no Redis. Nodes can find each other from SOCKETELX_CLUSTER_NODES or SOCKETELX_CLUSTER_DNS, or use a cluster your app already has. Survives a node dying, the leader dying and a network split that later heals. Tested across real separate nodes. See the Multi-server guide for setup and limits.
  • docker-compose.cluster.example.yml, a two-node example with no Redis.

[0.2.0] - 2026-10-09

Changed (breaking)

  • Renamed from pusher_server to Socketelx. The Hex package, OTP app and release are socketelx, the main module is Socketelx (was PusherServer), and the Docker image is socketelx. Update {:pusher_server, ...} to {:socketelx, ...} and {PusherServer, ...} to {Socketelx, ...}.
  • Environment variables are now SOCKETELX_*. The old PUSHER_* names are still accepted, and SOCKETELX_* wins when both are set.
  • Metric names start with socketelx_ (was pusher_server_), and Redis keys start with socketelx: (was pusher:). Redis state is ephemeral, so nothing needs migrating.

Added

  • Socketelx.trigger/5 publishes an event straight from your Elixir code, and Socketelx.authorize_channel/4 signs private and presence channel joins, so an embedded app needs no HTTP round trip.
  • examples/chat: a complete chat app (presence, messages, typing indicator) as a standalone Docker version and an embedded Elixir version, with a smoke test.
  • bench/load.mjs and a load-testing guide to simulate many connections.
  • Documentation reworked around pusher-js and the official Pusher server libraries; Laravel setup moved to its own guide.
  • POST /apps/:app_id/batch_events (up to 10 events per request) and GET /apps/:app_id/channels (with filter_by_prefix and info=user_count), both working across nodes with Redis.
  • The Docker workflow can also publish to Docker Hub, and tags semver versions and latest.
  • SOCKETELX_BATCH_LIMIT (default 10) sets the batch_events size limit.
  • Webhook retries with exponential backoff (SOCKETELX_WEBHOOK_RETRIES, default 3). 5xx, 408, 429 and network errors are retried; other 4xx are not.
  • Optional limits, all off by default: SOCKETELX_REST_RATE_LIMIT, SOCKETELX_CLIENT_EVENT_RATE_LIMIT, SOCKETELX_MAX_CONNECTIONS.
  • Optional Prometheus metrics at GET /metrics (SOCKETELX_METRICS, SOCKETELX_METRICS_TOKEN).
  • Multi-server mode: set SOCKETELX_REDIS_URL (or the :redis_url option) and several nodes share broadcasts, presence and channel occupancy through Redis. A janitor heartbeats each node, removes state left by dead nodes and rebuilds a node's state if Redis loses its data. Single-node use is unchanged and still needs no Redis.
  • deploy/nginx.conf and a two-node docker-compose.example.yml
  • Redis service in CI so the multi-node tests run there
  • User guides (guides/), grouped API docs, CONTRIBUTING, SECURITY, code of conduct, and issue / PR templates
  • A warning is logged when the public default app secret is in use

Changed

  • The server now advertises a 30 s activity_timeout and closes connections silent for 60 s. Both were 120 s, so a quiet client's ping could race the server's own idle timeout.

Fixed

  • Several ExDoc warnings (hidden-function links, missing Config.t() type).
  • Graceful shutdown never reached clients: the drain ran in Application.stop/1, which OTP calls after the supervision tree is gone. It now runs in prep_stop/1, with a regression test.
  • CI's mix format --check-formatted step would have failed; the code is now formatted and .formatter.exs knows the plug macro.
  • config/runtime.exs no longer starts a listener under mix test.
  • Dockerfile copies mix.lock when present.

[0.1.0] - 2026-07-09

Added

  • Pusher protocol websocket handshake, subscribe/unsubscribe, ping/pong
  • Private and presence channel auth (HMAC-SHA256)
  • Presence membership tracking, deduped by user_id across connections
  • REST trigger endpoint (POST /apps/:app_id/events)
  • Client events (client-*) on private/presence channels
  • Webhooks (channel_occupied/vacated, member_added/removed, client_event)
  • Read-only channel info endpoint (GET /apps/:app_id/channels/:channel)
  • Graceful shutdown (drains connections on SIGTERM) and idle connection timeout
  • Test suite for signature verification, presence semantics, and webhook payloads

Known limitations

  • Redis Cluster (sharded) is not supported; use a single Redis endpoint
  • No SSL termination built in (put a reverse proxy in front)
  • No rate limiting or connection caps
  • No batch events endpoint