tracker
All repositories: gitoria
13.4 KB
# hl:http1 PluginHTTP/1.1 server plugin for Hybriel. Epoll-based acceptor with I/O thread pool, optional TLS, keep-alive support.## Two Usage Patterns### 1. Event-based (class with `on request`)Extend `NativeHttpServer`, override `on request`, call `listen()`.```hybrielimport { NativeHttpServer, Response } from 'hl:http1';MyServer < NativeHttpServer {Number port = 8091;on request(req) {return new Response('<!DOCTYPE><html><head><title>Hello World!</title></head><body><h1>Hello World!</h1><p>This could be your front page!</p></body></html>', {status = 200;headers = {"Set-Cookie" = 'session=test';};});}}myServer = { port = 8091; } > new MyServer();```The server runs in the event loop. Each incoming request is dispatched to `on request` as a fiber, so the process stays alive and can handle concurrent requests.### 2. Iterator-based (`for (req of server)`)Call `listen(port)` directly, iterate requests in a loop.```hybrielimport { listen, Response } from 'hl:http1';server = listen(9123);for (req of server) {if (req.path == "/hello") {emit req.response(new Response('<!DOCTYPE><html><head><title>Hello World!</title></head><body><h1>Hello World!</h1><p>This could be your front page!</p></body></html>', {status = 200;headers = {"Set-Cookie" = 'session=test';};}));} else {emit req.response(new Response('Not found', { status = 404; }));}}```This blocks on each iteration waiting for the next request. Simpler but sequential.## Request ObjectEach request has these fields:| Field | Type | Description ||-------|------|-------------|| `method` | String | `"GET"`, `"POST"`, etc. || `path` | String | `"/hello"`, `"/api/users"` || `query` | Object | Parsed query params (keys/values) || `headers` | Object | HTTP headers (keys/values) || `body` | String or null | Request body — a `Content-Length` one, or a `Transfer-Encoding: chunked` one de-chunked || `bytes` | Bytes | The same body as a Bytes, byte for byte (empty when there is none) || `respond` | Handle | Response handle |## Response ObjectCreate a `Response` with body and optional options:```hybriel// Basicnew Response("Hello!")// With status and headersnew Response('{"ok":true}', {status = 200;headers = {"Content-Type" = 'application/json';"Set-Cookie" = 'session=abc';};})// Error responsenew Response('Not found', { status = 404; })```| Field | Type | Default | Description ||-------|------|---------|-------------|| `body` | String or Bytes | `''` | Response body; a Bytes is sent raw, as `application/octet-stream` unless `headers` name a type || `status` | Number | `200` | HTTP status code || `headers` | Object | `{}` | HTTP headers (key-value pairs) |## NativeHttpServer Properties| Property | Type | Default | Description ||----------|------|---------|-------------|| `port` | Number | 8080 | Listen port || `host` | String | "0.0.0.0" | Bind address || `tls_cert` | String | null | Path to TLS certificate || `tls_key` | String | null | Path to TLS private key || `threads` | Number | 4 | I/O worker threads |## Architecture```Client → epoll acceptor thread → PARK (epoll) → I/O thread pool (TLS + parse)→ request queue → event loop → on request fiber```- Acceptor thread uses epoll to accept connections without blocking- I/O workers handle TLS handshake and HTTP parsing in parallel- Parsed requests go into an MPSC queue- The runtime event loop polls the queue (non-blocking) and spawns a fiber per request### Parking: an idle connection costs no thread (mission 084)A worker's `read()` is blocking, so a connection handed to the pool occupies a wholeworker until bytes arrive. Keep-alive connections used to be re-enqueued straight into thepool after each response, so an idle one still held a thread — with the default`threads = 4`, **four quiet sockets starved the server** and every later request hung(mission 075 saw this as "the second browser session's page never fires `load`").Connections are therefore **parked** in a dedicated epoll instance (`ParkedConns`) and onlyenter the worker queue once they are actually readable — or have hung up, which a workerreads as EOF and closes. Both the acceptor and the post-response keep-alive path park.An idle connection now costs one fd and zero threads.- A parked connection that stays silent for `PARK_IDLE_TIMEOUT_MS` (60s) is closed.- Workers set `SO_RCVTIMEO` (30s) on every connection as a backstop against a client thattrickles headers; it is no longer the idle-timeout mechanism.- **An established TLS connection is never parked**: OpenSSL may hold already-decryptedplaintext that epoll on the raw fd cannot see. Those go straight to a worker, as before,so a TLS server is still starvable by idle keep-alive connections — the plain-HTTP path(and the pre-handshake TLS socket, whose ClientHello does arrive on the raw fd) is fixed.- Regression test: `tests/browser/tests/13-keepalive-starvation.mjs`.### WebSocket liveness sweep (mission 091)The WS reader loop (one thread, epoll, level-triggered EPOLLIN) already wakes on a 500mstimeout, so liveness costs no timer and no extra thread: on every tick it walks theupgraded sockets, pings the ones that have gone quiet, and drops the ones whose pong isoverdue.A drop enqueues the **same `close` event a real FIN would have produced**, so everyconsumer above this plugin — hl:web's push subscription registry included — prunes throughthe path it already had. No new concept travels upward, and nothing at the `.hl` levelneeds a timer facility.Why it is needed: a peer that vanishes without closing (dead wifi, a suspended laptop, ahost that never sent the FIN) leaves a socket that stays open and **writable**. A serverwith nothing to push at it never discovers the corpse, so before this the subscription wasimmortal. A closed tab is not this case — the browser's stack sends the FIN, and even underCDP offline emulation it still answers protocol pings (measured, mission 091), which is whythe regression test's dead peer is a raw socket that goes silent rather than a tab.| Env var | Default | Meaning ||---------|---------|---------|| `HL_WS_PING_MS` | `15000` | a socket quiet this long is pinged || `HL_WS_PONG_TIMEOUT_MS` | `10000` | a ping unanswered this long ⇒ the peer is gone |Read once when the engine starts; timeouts are measured on `CLOCK_MONOTONIC`, so an NTPstep cannot make one fire early or never. Any inbound frame — of any opcode — counts asproof of life and clears an outstanding ping.- Regression test: `tests/browser/tests/23-ws-liveness-sweep.mjs` (includes thefalsification: the same silent peer against a sweep-disabled server stays registered).### The handshake's cookies + a cookie-grade random (mission 093)A WebSocket upgrade is an ordinary HTTP request, so the browser attaches the site's cookiesto it. That is the one moment an upgraded socket can be attributed to whoever loaded thepage — after the 101 the request and its headers are freed and the connection carries noidentity of its own. The `connect` event therefore gained a sixth field:```{ kind, id, data, binary, code, cookie }````cookie` is the handshake's `Cookie:` header verbatim, non-empty on `connect` only. Nothingis parsed here: hl:web's session layer (`plugins/web/web_session.hl`) owns the cookie's nameand meaning, and this plugin owns only its delivery.```randomToken(Number n) // n lowercase hex chars, kernel CSPRNG (getrandom(2)), max 128```A session id is the only thing between a stranger and someone else's session, so it may notcome from a seeded PRNG — `hl:math`'s `random()` is a clock-seeded xoshiro whose state ahandful of outputs reveals, which would make every other id derivable from one's own. Thislives here because a session cookie is an HTTP artifact and this is the plugin that parsesand writes one; it is currently the only CSPRNG reachable from `.hl`.- Regression test: `tests/browser/tests/25-sessions.mjs` (two real browser instances = twocookie jars = two sessions).Ticket #74 added `host` (the handshake's `Host:` header) and ticket #105 `headers`: everyheader of the handshake as `name: value` lines, names lowercase, non-empty on `connect` only.`NativeWebSocketServer` hands them to the `WsClient` as `client.headers`, a hash by name —which is where hl:web reads the headers a navigation over that socket is constructed with.## The outgoing WebSocket client (ticket #109)Everything above is the SERVER half — a browser (or anything else) dialling in.`WebSocketClient` is the other direction: Hybriel dialling OUT to someone else'sWebSocket server. Subclass it and answer its three events. A subclass does NOT inheritthe parent's run body (same rule as `NativeHttpServer`/`NativeWebSocketServer`: a parent'sown root never runs for it) — end your own file's root with `connect()`, exactly as thosetwo end theirs with `listen()`:```hybrielimport { WebSocketClient } from 'hl:http1'inherit WebSocketClientString url = "wss://example.com/socket"Hybrid headers = { "Authorization" = "Bearer …" }on open() {send('hello')}on message(text) {console.log('got: ' + text)}on close(code) {console.log('closed: ' + code)}connect()```Or attach handlers to an instance directly, without subclassing at all — the same`on instance.event(...)` form used throughout this plugin's own tests:```hybrielimport { WebSocketClient } from 'hl:http1'client = new WebSocketClient(url = 'ws://127.0.0.1:8091/')on client.open() { client.send('hello') }on client.message(text) { console.log('got: ' + text) }on client.close(code) { console.log('closed: ' + code) }```### Properties| Property | Type | Default | Description ||----------|------|---------|-------------|| `url` | String | `null` | `ws://host[:port][/path]` or `wss://…` || `headers` | Object | `{}` | Extra headers sent on the Upgrade request || `caFile` | String | `null` | `wss://` only — an EXTRA trust anchor for a self-signed fixture, exactly hl:fetch's field of the same name (never a replacement for the system trust store) || `connected` | Boolean | `false` | True between `open` and `close` |### Methods| Method | Answers | Description ||--------|---------|-------------|| `connect()` | `true`/`false` | Dial `url` (idempotent — a no-op while already dialling or open). The file root calls this for you. || `send(text)` | `1`/`0` | One text frame to the peer; `0` when not open || `close([code])` | `1`/`0` | Start the close handshake; `close` still fires once the peer's echo (or a timeout) completes it |### Events```emit open() // the connection is open and ready for send()emit message(String data) // one complete text (or binary) messageemit close(Number code) // the connection ended — RFC 6455 §7.4 code```At most one `open` ever precedes a `close`: a connection that never opened still getsexactly one `close` (`1006` for a dial that never reached a peer — DNS failure, refused,TLS handshake failure; `1002` for a handshake the peer answered but got wrong).### `wss://` uses hl:fetch's TLSThe client half of `plugins/http/tls_common.zig` — the SAME module hl:fetch uses for`https://`. The system trust store, hostname verification always on, and `caFile` for anextra anchor. There is no verify-off switch, on this client either.### Ping/pong is automatic; reconnect never isA quiet connection is pinged after `HL_WS_PING_MS` and dropped if the pong doesn't landwithin `HL_WS_PONG_TIMEOUT_MS` — the SAME two env vars the server's liveness sweep reads(mission 091, above). Any inbound ping is answered with a pong immediately. Neither everreaches `.hl`: the protocol requires both, and they are not application information.**Reconnect is never automatic.** A dropped connection reaches you as exactly one `close`event and nothing here dials again on its own — call `connect()` again yourself (typicallyfrom your own `on close`) if you want a reconnect policy; `connect()` after a `close` opensa fresh connection, it is not a one-shot object.### ArchitectureOne dedicated thread per connection — the same shape as hl:fetch's one-thread-per-ticketmodel, not the server's shared epoll pool: a client dials a handful of long-lived peers,not thousands of short ones, and a private thread means a private event queue with no riskof misdelivering one connection's frames onto another instance's `on open`/`onmessage`/`on close`. The thread does DNS + TCP connect, the TLS handshake for `wss://`, theHTTP Upgrade handshake, then a read loop over the shared RFC 6455 framing(`plugins/http/ws_common.zig`) — this side's own frames go out through`writeFrameMasked`/`writeCloseMasked` (RFC 6455 §5.1: "a client MUST mask all frames"),freshly masked per frame from the kernel CSPRNG.- Fixtures: `tests/pass/plugins/039_http1_ws_client.hl` (open, send, a server-pushedframe, close from the client) and`tests/pass/plugins/040_http1_ws_client_close_reconnect.hl` (close from the SERVER'sside, and an explicit reconnect proving it is never automatic).
Branches
- mainmain branch
Latest commits
- dc40d859tracker#31 (mission 026): duplicate titles merged — the 68 type+tmdbId pairs held by 157 records were the old tracker's (all migrated); merge.hl repair job (own clock, before the TMDB jobs) keeps one keeper per title (follows/watches > old short id > oldest), moves follows, watches, seasons, cast, credits, timelines, tombstones the rest (mergedInto, never deleted), slugs + short ids 301 to the keeper; stray seasons merged into their listed twin (Reacher S3 watches) or linked when watched; search import re-checks before its put; deploy.sh waits up to 90 s for 200; real copy 68 -> 0 dup ids, az5b2 follows/watches equal; gates 342/0, 32/0, 52/0mre
- 2667da05tracker#30 (mission 025): /my/ pages from slim cached title cards, episode rows and watch sets (after the jobs /my/series 1.8 s -> 0.06 s, /my/unwatched 4.1 -> 0.18 s); timeline page shows its name once; franchise widget under the poster/title; movies with TV leftovers (First Contact) go through the details repair; tools/count-tmdb-ids.hl; gates 327/0, 52/0, 32/0mre
- 3909810dantcolony#40: mission references in README/STATUS/docs point to the moved missionsmre
- e63b1d28antcolony#40: history (LOG.md), worker briefs (missions/) and reports moved here from antcolony, numbered per project; old numbers in antcolony docs/mission-map.mdmre
- 1cda451dtracker: Hybriel master 190aa11d (#127 both shapes, GC correctness fc838894) — conductor adopts despite /my/series 2.4x after jobs (memory 8.0 → 1.7 GB boot); see reports/071mre
- c1fa2f2etracker (mission 071): Hybriel master 190aa11d measured on the real copy vs the live binary 8590df63 — NOT adopted (after the first-start jobs /my/series 2.4x slower, /series 1.6x, RSS swings 7.4-12.2 GB; fresh it is flat at 1.7-2.2 GB and /my/unwatched faster), vendor stays 8efba065, candidate kept in .scratch/w071/vendor-190aa11d; tests/kinds.mjs: collection seed off (its TMDB request broke check 1 in 1 of 4 runs); tests/realdata-071.sh + realdata-071-bench.mjs + tools/realdata-071-table.py; README + STATUS (numbers, how to repeat); gates 325/0, 32/0, 50/0mre
- 8081350atracker docs (mission 070): README (summary, Config HL_GC_BYTES — kept at Hybriel's default, the 256 MiB setting is taken out of docker-compose.yml again: the jobs grew to 12+ GB with it too, see STATUS), Test (three gates), Deploy (first start ~50 min: kinds then seed, restart once after collections done, memory numbers), Vendored Hybriel 8efba065 + #48 audit, Files; STATUS mission 070 entry (merges, migrated counts, lambda audit, gates, RSS old vs new, how to repeat, open points); docs/kinds.md + docs/franchises.md job order; tests/realdata-070-*.sh, tools/count-migrated.hl, tools/ref-params.py, tools/lambda-audit.pymre
- 1ad19c8ctracker: re-vendor Hybriel master 8efba065 (#126 GC by bytes, #48 lambda parameters copy) (mission 070): bin/hybriel sha256 50361e95…, plugins core crypto data fetch fs http http1 mpackdb proc smtp time web; lambda audit: 13 lambdas change a passed record/list (11 through a local alias), no caller relies on it — unchanged; 319 read-only lambda parameters get & (no copy per call: /my/schedule 2.35 → 0.46 s, /my/unwatched 13.7 → 5.7 s on the real copy); /my/unwatched one merge sort instead of n² inserts; docker-compose HL_GC_BYTES=268435456; deploy.sh runs kinds.mjs + franchises.mjs too (default ports 8700–8710); tests/realdata-070.mjs; gates browser 325/0, kinds 32/0, franchises 50/0mre
- 46b21389tracker (mission 070, conductor): the sync never destroys migrated data — the one-time summary step MOVES a copied summary to migratedSummary (marker summaries-moved.txt) instead of clearing it; a migrated record's first title/genres/homepage/tagline TMDB replaces → migratedTitle/migratedGenres/migratedHomepage/migratedTagline, a migrated season's/episode's title/summary → migratedTitle/migratedSummary (set once); tests/peek-shows.hl prints them; gate 325/0mre
- 749019b0Merge t19 (tracker#19 franchises + timelines) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/browser.mjs/faketmdb.mjs, both sides kept; franchise/timeline pages get #20's typed heading (Franchise | …, Timeline | …), their title links via titlePath (/movies|/series|/shows); the collection seed waits for repair, kinds and credits too; franchises.mjs URLs + 2 new checks; gates browser 323/0, kinds 32/0, franchises 50/0mre
- daf49feaMerge t20 (tracker#20 typed headings + #21 series/shows split) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/search.hl/components/search.hl/browser.mjs/faketmdb.mjs, both sides kept; clock order backfill → repair → kinds → credits; withDetailsFields stores tmdbType + kind; gates: browser.mjs URLs → /series|/movies, typed h1 selectors; kinds.mjs repair/credits off, fixture name = seed name; browser 323/0, kinds 32/0mre
- 20e09d89Merge t18 (tracker#18 delta sync) into main (mission 070): conflicts README/STATUS/show.hl/project.hl/browser.mjs, both sides kept; pageShowOf summary = summaryOfmre
- f83571c3tracker#18 (mission 067): daily sync by change lists — TMDB /tv|movie/changes (since the stored day, paged) + TVmaze /updates/shows → only our changed titles (followed: full step, unfollowed: light step — changed seasons, no TVmaze), full walk on first run / gap > 14 days / failed list; show record refreshed (title, tmdbSummary, tagline, status, genres …; renamed titles re-indexed); summary = the creator's own text (page: summary > tmdbSummary > tvmazeSummary), one-time clear of copied summaries (9,647 on the live copy); gate 261, tests/realdata-018*.mjs, README + STATUSmre
- 1ed5457etracker#20 + #21 (mission 068): typed headings "<Type> | <name>" in type colours; TV titles split into Series (/series) and Shows (/shows) by TMDB type + Reality/Talk/News genres — kind stored by sync/import/adult backfill + new kind backfill (resumes), /movies/<slug>, /shows/<slug> of a series/movie → 301, /my/series + /my/shows, home 5 tiles + 3 rows, search/filmography labels; gates kinds 32 + browser 266, tests/realdata-068.mjs, tools/count-kinds.hl, docs/kinds.md, README + STATUSmre
- f2b00674Merge t26 (tracker#26 + #28) into main (mission 062): short ids for every new person (castPersonId, guest route), guest stars stored on the title and created as people only when opened (/person/tmdb/<id>?show=<id> → 302), lean watch/follow clicks (showRow a small object, cast/crew from the slug, watches cached per user, face rows only after a season toggle); gate 311, tests/realdata-062.mjs, README + STATUSmre
- 0553b51ftracker#19 (mission 066): franchises and timelines — tables, /franchises, /franchises/<slug>, /timelines/<slug> (Timeline | Release sort, series by last episode), the Prequel | Timeline | Sequel widget with the franchise above, the creator's editor, TMDB collection seed in the app (resumes, paced); gate tests/franchises.mjs 48/0 + browser.mjs 266/0, tests/realdata-066.mjs, docs/franchises.md, README + STATUSmre
- fa1f9dfatracker#29 (mission 063): unwatched check muted grey outline + check (accent only on hover), watched stays solid — no code regression, the accent outline read as ticked; gate checks real checks visibly (computed style + screenshot pixel) on /my/unwatched, show, movie, /my/movies; gate 266, tests/realdata-063.mjs, README + STATUSmre
- 10bb3f93tracker#28 (mission 061): full cast (all seasons, main cast by episodes, guest stars) + crew (created by, directed by, written by, screenplay, story, music) — stored by the details completion, the daily sync, the search import (one details request) and a background credits job (resumes, RSS limit); show page collapsed after 20 with client-side Show all; showBySlug via a slug map; gate 249, tests/realdata-028.mjs, README + STATUSmre
- d8b12d67tracker#27 (mission 060): short ids for movies, series and persons — old 702 kept (data/old-short-ids.json), new random [a-z0-9]{5} unique across both, claimed at creation, background backfill (resumes), shown under poster/photo, /<shortId> → 301; gate 259, tests/realdata-060*, README + STATUSmre
- 25a50bc4tracker#26 (mission 059): titles from a filmography are completed — on open (skeleton, step-wise face showComplete, no reload) and by the in-app details repair (resumes, TMDB-paced, series in parts); cast from TMDB credits; gate 231, tests/realdata-026.mjs, README + STATUSmre