synth-ai 0.2.9.dev0__py3-none-any.whl → 0.2.23.dev3__py3-none-any.whl

examples/multi_step/sse_metrics_streaming_notes.md

@@ -0,0 +1,357 @@
+# SSE metrics/event streaming design (RL + FT)
+
+## Goals
+- Near real-time push of job status, metrics, and logs during RL training, evaluation, and fine-tuning (FT)
+- Single streaming endpoint per job, resumable (Last-Event-ID), low overhead, widely compatible (HTTP/1.1)
+- Minimal client friction (CLI + Python helper), production-ready (auth, backpressure, rate limit)
+
+## Non-goals
+- Binary/frame multiplexing (use WebSocket if needed later)
+- Arbitrary high-frequency payloads (we will coalesce/limit ~2–4 Hz for metrics)
+
+---
+
+## Endpoint
+- Method: GET `/rl/jobs/{job_id}/stream`
+- Headers:
+  - Auth: `X-API-Key: <env key>` or `Authorization: Bearer <token>`
+  - Cache: `Cache-Control: no-cache`
+- Response:
+  - Content-Type: `text/event-stream; charset=utf-8`
+  - Transfer-Encoding: `chunked`
+  - Connection: `keep-alive`
+- Query params (optional):
+  - `since_id`: int; resume from a specific event id (inclusive)
+  - `types`: comma list `metric,status,log,artifact` (default: all)
+  - `heartbeat`: seconds between heartbeats (default 20)
+  - `split`: `train|eval` (filters metrics only)
+
+## Event framing (SSE)
+- Fields per message:
+  - `id: <int>` monotonically increasing per job
+  - `event: <status|metric|log|artifact|heartbeat>`
+  - `data: <JSON>` single-line JSON (compact)
+- Heartbeats: comment lines `: keep-alive` at configured interval
+- Flush: after each event write + heartbeat
+- Backpressure: if producer > consumer, coalesce metrics, keep status/logs, never buffer unbounded
+
+## Payload schemas
+- `status`
+  - `{ "state": "queued|running|succeeded|failed|canceled", "step": 123, "epoch": 3, "phase": "train|eval|ft", "message": "...", "ts": 173.12 }`
+- `metric`
+  - `{ "name": "avg_reward|loss|accuracy|success_rate|return", "value": 0.123, "step": 123, "epoch": 3, "split": "train|eval", "window": 100, "mean": 0.42, "std": 0.08, "ts": 173.12 }`
+  - Optional extras: `{ "tags": {"env": "crafter", "policy": "react"} }`
+- `log`
+  - `{ "level": "INFO|WARN|ERROR", "message": "...", "ts": 173.12 }`
+- `artifact`
+  - `{ "kind": "checkpoint|trace|plot|jsonl", "url": "/rl/jobs/{id}/artifacts/ckpt_0003.pt", "step": 123, "ts": 173.12 }`
+- `heartbeat`
+  - `{ "alive": true, "ts": 173.12 }`
+
+### Example stream (illustrative)
+```
+id: 101
+event: status
+data: {"state":"running","phase":"train","step":820,"epoch":4,"ts":173.12}
+
+id: 102
+event: metric
+data: {"name":"avg_reward","value":0.62,"step":820,"epoch":4,"split":"train","ts":173.13}
+
+id: 103
+event: metric
+data: {"name":"loss","value":1.84,"step":820,"epoch":4,"split":"train","window":100,"mean":1.90,"std":0.15,"ts":173.13}
+
+id: 104
+event: log
+data: {"level":"INFO","message":"checkpoint saved","ts":173.16}
+
+id: 105
+event: artifact
+data: {"kind":"checkpoint","url":"/rl/jobs/j_abc/artifacts/ckpt_0004.pt","step":820,"ts":173.16}
+
+: keep-alive
+```
+
+---
+
+## Server architecture
+
+### Components
+- Event bus per `job_id` (async queue) where producers (RL, FT, evaluator) `emit(Event)`
+- Ring buffer per job for replay (configurable: last N events OR last T minutes)
+- SSE handler:
+  1) Authenticate, pick job, determine resume cursor (`since_id` or `Last-Event-ID`)
+  2) Replay from ring buffer >= cursor
+  3) Attach to live queue; stream new events
+  4) Emit heartbeats; close after terminal `status` + grace
+
+### Concurrency & ordering
+- Single writer increments `event_id`
+- Replay preserves original order; live continues from last id
+- If consumer slow: drop/coalesce metrics (preserve last per metric name), always deliver status/log/artifact
+
+### Rate limiting & coalescing
+- Default target 2–4 Hz for metrics per split
+- Coalesce by metric name within a small interval (e.g., 250–500 ms)
+- Status events limited to phase changes or every 5–10s
+
+### Auth & security
+- Accept `X-API-Key` or `Authorization: Bearer`
+- Validate job ownership/visibility
+- CORS: allow EventSource; set `Access-Control-Allow-Origin` appropriately
+- Timeouts: server idle timeout > heartbeat * 2; client reconnect on drop
+
+### Config knobs (env)
+- `SSE_HEARTBEAT_SECS` (default 20)
+- `SSE_RING_BUFFER_EVENTS` (e.g., 2000) OR `SSE_RING_BUFFER_WINDOW_SECS` (e.g., 600)
+- `SSE_MAX_METRIC_HZ` (e.g., 4)
+- `SSE_MAX_CLIENTS_PER_JOB` (protect from fan-out)
+
+---
+
+## Emit points
+
+### RL training/eval
+- On train step end: `metric` avg_reward/return/success_rate; `status` every N steps
+- On eval step end: `metric` eval_return/success_rate; `artifact` eval JSONL optional
+- On checkpoint: `artifact` + `log`
+- On phase transitions: `status` (train→eval, etc.)
+
+### Fine-tuning (FT)
+- On optimizer step: `metric` loss (and optional lr)
+- On validation: `metric` val_loss/accuracy; optional `artifact` (curves)
+- On checkpoint: `artifact` + `log`
+
+---
+
+## Client (synth-ai)
+
+### CLI
+- `synth-ai jobs stream <job_id> [--jsonl out.jsonl] [--types metric,status]`
+- Prints compact lines: `t=18:22:40 step=820 avg_reward=0.62 loss=1.84`
+- Writes raw events to JSONL if specified
+
+### Python helper
+```python
+from synth_ai.client import TaskAppClient
+
+with TaskAppClient(base_url, api_key) as c:
+    for ev in c.stream_job(job_id, types=["metric","status"], since_id=None):
+        handle(ev)
+```
+- Handles reconnect with `Last-Event-ID`
+- Dedupes by event id; optional local aggregation windows
+
+---
+
+## Error handling & edge cases
+- 404 (job not found): immediate error, no stream
+- 401/403: immediate error, no stream
+- Producer stalls: heartbeats continue; client shows "connected: no new data"
+- Ring buffer miss (client too far behind): send a summary `status` snapshot, then live only
+- Buffer overflow: set `dropped_events: true` field in a `status` update
+
+---
+
+## Testing plan
+- Unit: SSE formatting, ring buffer indexing, resume semantics, coalescing
+- Integration: run toy RL + FT loops; assert ordered sequence and types
+- Load: simulate 100 Hz metric emissions; verify rate limiting and non-blocking behavior
+- Auth: invalid key, missing key, cross-job access rejection
+
+---
+
+## Rollout plan
+1) Implement event bus + ring buffer per job
+2) Add SSE route with auth, resume, heartbeat
+3) Emiters in RL/FT loops (train/eval steps, loss/reward, checkpoints)
+4) synth-ai client: CLI + Python helper; optional JSONL sink
+5) Docs and example snippets (curl, Python, browser EventSource)
+
+---
+
+## Open questions / decisions
+- Do we persist events beyond ring buffer? (optional artifact JSONL)
+- Multi-tenant limits per org/user? (SSE_MAX_CLIENTS_PER_ORG?)
+- Artifact URLs: signed vs. authenticated direct
+- Backward compatibility with current polling endpoints
+I'll scan the backend to find where RL jobs are created, status is served, and whether there's any existing streaming/event infrastructure to extend. Then I'll identify exact files/functions to change for adding an SSE endpoint and metric emitters.
+
+[2 tools called]
+
+Here’s the exact backend surface you’ll need to touch to add a single SSE stream for status + metrics, with where to hook metric emitters.
+
+### Where jobs and events live
+- Job and event tables (legacy RL path):
+```1:20:monorepo/backend/app/db/models/jobs.py
+"""Job orchestration models (RL/SFT shared).
+"""
+```
+```31:58:monorepo/backend/app/db/models/jobs.py
+class Job(Base):
+    __tablename__ = "jobs"
+    job_id = Column(String, primary_key=True)
+    ...
+    status = Column(String, nullable=False, default="QUEUED")
+    ...
+    last_seq = Column(BigInteger, nullable=True, default=0)
+    events = relationship("JobEvent", back_populates="job", cascade="all, delete-orphan", passive_deletes=True)
+```
+```66:88:monorepo/backend/app/db/models/jobs.py
+class JobEvent(Base):
+    """Append-only job event log."""
+    __tablename__ = "job_events"
+    job_id = Column(String, ForeignKey("jobs.job_id", ondelete="CASCADE"), primary_key=True)
+    seq = Column(BigInteger, primary_key=True)
+    ts = Column(DateTime(timezone=True), server_default=func.now())
+    type = Column(String, nullable=False)
+    level = Column(String, nullable=False, default="info")
+    message = Column(Text, nullable=False)
+    data = Column(JSONB, nullable=True)
+```
+
+- DB repo (atomic seq increment + append):
+```108:146:monorepo/backend/app/orchestration/jobs/repository_db.py
+async def append_event(...):
+    res = await self.session.execute(
+        update(LearningJob)
+        .where(LearningJob.job_id == job_id)
+        .values(last_seq=func.coalesce(LearningJob.last_seq, 0) + 1, updated_at=func.now())
+        .returning(LearningJob.last_seq)
+    )
+    seq = res.scalar_one()
+    ev = JobEvent(job_id=job_id, seq=seq, type=type_, level=level, message=message, data=(data or {}))
+    ...
+```
+
+- RL jobs REST (create + emit first events via PostgREST emitter):
+```214:233:monorepo/backend/app/routes/clustered_training/core/routes.py
+@router.post("/jobs", ...)
+async def create_job(...):
+    ...
+```
+```784:799:monorepo/backend/app/routes/clustered_training/core/routes.py
+await get_postgrest_emitter().append_event(job_id=job_id, type_="rl.job.created", message="RL job created", data={"work": payload.work})
+```
+
+- Shared storage path showing how “update” and “append event” are broadcast via the PostgREST emitter:
+```242:271:monorepo/backend/app/routes/simple_training/services/job_service.py
+async def update_job(...): ...
+async def append_job_event(...): ...
+```
+
+What to add/change (SSE + emitters)
+
+1) Add SSE endpoint under RL jobs
+- File: `monorepo/backend/app/routes/clustered_training/core/routes.py`
+- New route: `GET /api/rl/jobs/{job_id}/stream`
+  - Auth: `ValidatedAPIKey`
+  - Headers: honor `Last-Event-ID`; Query: `since_id`, `types`, `heartbeat`
+  - Flow:
+    - Validate job exists (use the DB-backed repo in production mode)
+    - Determine start seq (from Last-Event-ID or since_id, else 0)
+    - Replay: call repo `list_events(job_id, after=seq)` and stream as SSE (`id`, `event`, `data`)
+    - Live tail: loop with short sleep (e.g., 0.5–1.0s) fetching new events by `after=last_seq`
+    - Heartbeats as comments `: keep-alive` every N seconds
+  - Event mapping:
+    - `JobEvent.type` prefixes map to SSE `event`:
+      - `rl.job.*`, `job.updated` → `status`
+      - `rl.step.metric`, `ft.step.metric`, `eval.metric` → `metric`
+      - `system.log.*` → `log`
+      - `artifact.*` → `artifact`
+    - `Job.status` can be snapshotted once at connect (send a `status`)
+
+2) Ensure a consistent event source for reads
+- Prefer the DB repo (`JobsRepositoryDB`) in prod mode. If the current code path uses the PostgREST emitter for appends, verify that the repo’s `list_events` reads from the same canonical table (it does for `JobEvent`). If your RL path uses the “learning_shared” models instead, use the associated repository there (same pattern: list by job_id + seq).
+- If you must keep PostgREST for append-only, that’s fine; SSE can still read the DB rows inserted alongside (your outbox/emitter already supports both).
+
+3) Emitters in training/FT loops
+- File(s): `monorepo/backend/app/orchestration/hatchet/workflows.py` (RL workflow nodes), any FT job loops
+- After each meaningful step:
+  - Train: append `type="rl.step.metric"`, `data={"avg_reward":..., "return":..., "success_rate":..., "step":..., "epoch":..., "split":"train"}`.
+  - Eval: `type="eval.metric"` with eval metrics and split.
+  - FT: `type="ft.step.metric"`, `data={"loss":..., "lr":..., "step":..., "epoch":..., "split":"train"}`; validation as `split="eval"`.
+  - On phase changes/checkpoints: `type="job.updated"` or `artifact.checkpoint` with URLs.
+- Use the same helper used elsewhere:
+```236:276:monorepo/backend/app/routes/simple_training/services/job_service.py
+async def append_job_event(...): return await get_postgrest_emitter().append_event(...)
+```
+
+4) Optional shared service abstraction
+- File: `monorepo/backend/app/routes/simple_training/services/storage_shared.py`
+  - Add a small `stream_job_events(job_id, after)` helper that wraps `repo.list_events(...)` and normalizes schemas (legacy vs learning_shared). The SSE route can call this.
+
+5) Wire the router
+- File: `monorepo/backend/app/routes/main.py`
+  - Include the new GET route (under the RL router you already mount).
+- CORS: ensure EventSource allowed if frontend will use browser SSE.
+
+Indexes/Perf you already have
+- Jobs/events tables include per-job indexes and a `last_seq`; reads by `(job_id, seq)` are efficient:
+```84:88:monorepo/backend/app/db/models/jobs.py
+Index("idx_job_events_job_ts", "job_id", "ts"),
+Index("idx_job_events_data_gin", "data", postgresql_using="gin"),
+```
+- For learning_shared events (if used), there are sequence indexes too:
+```159:165:monorepo/backend/app/db/models/learning_shared.py
+sa.Index("idx_learning_job_events_job_seq_idx", "job_id", "seq"),
+```
+
+Summary of minimal backend edits
+- Add SSE route:
+  - `monorepo/backend/app/routes/clustered_training/core/routes.py` (GET `/api/rl/jobs/{job_id}/stream`)
+- Consume events via repo:
+  - `monorepo/backend/app/orchestration/jobs/repository_db.py` (use `list_events`)
+- Emit metrics from loops:
+  - `monorepo/backend/app/orchestration/hatchet/workflows.py` (append_event at train/eval steps)
+- Optional shared helper:
+  - `monorepo/backend/app/routes/simple_training/services/storage_shared.py` (normalize event reads)
+- Wireup:
+  - `monorepo/backend/app/routes/main.py` (include SSE route)
+- No schema changes required; you already have `JobEvent`/indexes and event append plumbing.
+
+### SDK additions for great terminal polling (with SSE fallback)
+
+- RlJobsApi extensions (synth_ai/jobs/client.py)
+  - stream(job_id, since_id=None, types=None, heartbeat=None) -> async iterator of events (uses SSE; falls back to polling)
+  - events(job_id, after=None, limit=500) -> list[JobEvent] (poll)
+  - status(job_id) -> JobSummary (single snapshot)
+
+- Event models (synth_ai/jobs/types.py)
+  - JobEvent base: {id, type, level, message, data, ts}
+  - StatusEvent, MetricEvent, LogEvent, ArtifactEvent (typed helpers)
+
+- JobsWatcher helper (synth_ai/jobs/watcher.py)
+  - constructor(client, job_id, interval=2.0, prefer_sse=True, jsonl_path=None, types=None)
+  - run(on_event, stop_when=None) → handles SSE connect/reconnect, polling fallback, Last-Event-ID cursor, dedupe
+  - metrics_tracker: rolling windows per metric name (mean/std/min/max, last_value, last_step)
+  - backoff policy: jittered reconnect; rate limiter for render
+
+- Terminal renderer (synth_ai/jobs/render.py)
+  - RichRenderer (or minimal TTY): compact line updates: t=HH:MM:SS | step/E | key metrics (avg_reward, loss, val_loss, success_rate)
+  - modes: one-line ticker vs. per-event lines; quiet mode; color by level/state
+  - JSONL sink: raw event writes without printing prompts/payloads
+
+- CLI command (synth_ai/api/train/cli.py)
+  - synth-ai jobs watch <job_id> [--types metric,status] [--interval 2] [--jsonl out.jsonl] [--since-id N] [--no-sse]
+  - exit codes: 0 on succeeded, 1 on failed/canceled, 2 on timeout
+
+- Utilities (synth_ai/jobs/utils.py)
+  - BackoffPolicy(retry, max) with jitter
+  - EventCursor(last_id, update)
+  - MetricsFormatter(map by job_type: RL vs FT metric labels)
+  - Coalescer: compress frequent metrics to ≤4 Hz
+
+- Defaults/behavior
+  - Prefer SSE; if 404/405/close → fallback to polling events() every interval
+  - Heartbeat support; show “connected/no data” when only heartbeats
+  - Resume: honor --since-id or Last-Event-ID; persist cursor optionally
+
+- Minimal backend assumptions
+  - GET /api/rl/jobs/{job_id}/stream (SSE) or /api/rl/jobs/{job_id}/events?after=… (poll)
+  - Events include metric/status/log/artifact with seq ids and ts
+
+- Extensibility
+  - Plugin renderers per job_type (rl, sft/ft)
+  - Hooks: on_status_change, on_metric(name, value), on_artifact(url)