mnm-sdk 0.1.1__py3-none-any.whl

mnm_sdk/__init__.py

@@ -0,0 +1,6 @@
+"""MNM SDK: log model predictions to MNM without slowing down inference."""
+
+from mnm_sdk.monitor import Monitor, Stats
+from mnm_sdk.transport import HttpTransport, Transport, TransportError
+
+__all__ = ["HttpTransport", "Monitor", "Stats", "Transport", "TransportError"]

mnm_sdk/monitor.py

@@ -0,0 +1,212 @@
+"""Monitor: the object a model service uses to log predictions.
+
+`log()` only appends to a bounded in-memory queue. A background thread sends
+the queue to the MNM ingestion API in batches. Nothing in here may slow down
+or break the caller's prediction path: `log()` never blocks on the network
+and never raises.
+"""
+
+from __future__ import annotations
+
+import atexit
+import logging
+import threading
+import uuid
+from collections import deque
+from collections.abc import Mapping
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any
+
+from mnm_sdk.transport import HttpTransport, Transport, TransportError
+
+logger = logging.getLogger("mnm")
+
+
+@dataclass(frozen=True)
+class Stats:
+    """Counters since the Monitor was created."""
+
+    logged: int  # events accepted by log()
+    sent: int  # events delivered to the ingestion API
+    dropped: int  # events discarded before sending (queue full, invalid, closed)
+    failed: int  # events given up on after the retries ran out
+    queued: int  # events waiting in the queue right now
+
+
+class Monitor:
+    """Logs predictions of one model version to MNM in the background.
+
+    Args:
+        api_url: Base URL of the MNM backend, e.g. ``http://backend:8000``.
+        model: Model name.
+        version: Model version.
+        batch_size: Send as soon as this many events are queued.
+        flush_interval: Send whatever is queued at least this often (seconds).
+        max_queue: Queue capacity. When full, the oldest event is dropped.
+        max_retries: Extra attempts for a batch that failed to send.
+        retry_backoff: Wait before the first retry (seconds); doubles each time.
+        timeout: HTTP timeout per request (seconds).
+        transport: Replace the HTTP transport (used by tests).
+    """
+
+    def __init__(
+        self,
+        api_url: str,
+        model: str,
+        version: str | int,
+        *,
+        batch_size: int = 100,
+        flush_interval: float = 1.0,
+        max_queue: int = 10_000,
+        max_retries: int = 3,
+        retry_backoff: float = 0.5,
+        timeout: float = 5.0,
+        transport: Transport | None = None,
+    ) -> None:
+        self._model = model
+        self._version = str(version)
+        self._batch_size = max(1, batch_size)
+        self._flush_interval = flush_interval
+        self._max_queue = max(1, max_queue)
+        self._max_retries = max(0, max_retries)
+        self._retry_backoff = retry_backoff
+        self._transport = transport or HttpTransport(api_url, timeout=timeout)
+
+        self._queue: deque[dict[str, Any]] = deque()
+        self._lock = threading.Lock()
+        self._drained = threading.Condition(self._lock)
+        self._wake = threading.Event()
+        self._closing = threading.Event()
+        self._in_flight = 0
+        self._logged = self._sent = self._dropped = self._failed = 0
+
+        self._thread = threading.Thread(
+            target=self._run, name="mnm-monitor", daemon=True
+        )
+        self._thread.start()
+        atexit.register(self.close)
+
+    # -- called on the prediction path -------------------------------------
+
+    def log(
+        self,
+        features: Mapping[str, Any],
+        output: Any,
+        latency_ms: float,
+        *,
+        request_id: str | None = None,
+        label: Any = None,
+        ts: datetime | None = None,
+    ) -> None:
+        """Queue one prediction. Returns immediately and never raises."""
+        try:
+            event: dict[str, Any] = {
+                "model": self._model,
+                "version": self._version,
+                "request_id": request_id or uuid.uuid4().hex,
+                "ts": (ts or datetime.now(timezone.utc)).isoformat(),
+                "features": dict(features),
+                "output": output,
+                "latency_ms": float(latency_ms),
+            }
+            if label is not None:
+                event["label"] = label
+        except Exception:
+            with self._lock:
+                self._dropped += 1
+            logger.debug("mnm: dropped an event that could not be built", exc_info=True)
+            return
+
+        with self._lock:
+            if self._closing.is_set():
+                self._dropped += 1
+                return
+            if len(self._queue) >= self._max_queue:
+                self._queue.popleft()
+                self._dropped += 1
+            self._queue.append(event)
+            self._logged += 1
+            batch_ready = len(self._queue) >= self._batch_size
+        if batch_ready:
+            self._wake.set()
+
+    # -- control -------------------------------------------------------------
+
+    @property
+    def stats(self) -> Stats:
+        with self._lock:
+            return Stats(
+                logged=self._logged,
+                sent=self._sent,
+                dropped=self._dropped,
+                failed=self._failed,
+                queued=len(self._queue),
+            )
+
+    def flush(self, timeout: float | None = None) -> bool:
+        """Block until everything logged so far was sent or given up on.
+
+        Returns False if that did not happen within `timeout` seconds.
+        """
+        self._wake.set()
+        with self._drained:
+            return self._drained.wait_for(
+                lambda: not self._queue and not self._in_flight, timeout
+            )
+
+    def close(self, timeout: float = 5.0) -> None:
+        """Send what is queued, then stop the background thread. Idempotent."""
+        self._closing.set()
+        self._wake.set()
+        self._thread.join(timeout)
+
+    # -- background thread ---------------------------------------------------
+
+    def _run(self) -> None:
+        while True:
+            self._wake.wait(self._flush_interval)
+            self._wake.clear()
+            self._drain()
+            if self._closing.is_set():
+                self._drain()  # anything logged while the last batch was sending
+                return
+
+    def _drain(self) -> None:
+        while True:
+            with self._lock:
+                if not self._queue:
+                    self._drained.notify_all()
+                    return
+                size = min(self._batch_size, len(self._queue))
+                batch = [self._queue.popleft() for _ in range(size)]
+                self._in_flight = size
+            delivered = self._send(batch)
+            with self._lock:
+                self._in_flight = 0
+                if delivered:
+                    self._sent += len(batch)
+                else:
+                    self._failed += len(batch)
+                self._drained.notify_all()
+
+    def _send(self, batch: list[dict[str, Any]]) -> bool:
+        """Try to deliver one batch, retrying with backoff. Never raises."""
+        delay = self._retry_backoff
+        for attempt in range(self._max_retries + 1):
+            try:
+                self._transport.send(batch)
+                return True
+            except TransportError as error:
+                reason: Exception = error
+                if not error.retryable:
+                    break
+            # A bug in a transport must not kill the logging thread.
+            except Exception as error:  # noqa: BLE001
+                reason = error
+            if attempt < self._max_retries:
+                # Returns early when close() is called: do not delay shutdown.
+                self._closing.wait(delay)
+                delay *= 2
+        logger.warning("mnm: gave up on %d events: %s", len(batch), reason)
+        return False

mnm_sdk/transport.py

@@ -0,0 +1,76 @@
+"""How batches leave the process: HTTP POST to the MNM ingestion API."""
+
+from __future__ import annotations
+
+import json
+import urllib.error
+import urllib.request
+from typing import Any, Protocol
+
+INGEST_PATH = "/v1/ingest"
+
+
+class TransportError(Exception):
+    """A batch could not be delivered.
+
+    `retryable` is False when sending the same batch again cannot succeed
+    (the server rejected it as invalid).
+    """
+
+    def __init__(self, message: str, *, retryable: bool = True) -> None:
+        super().__init__(message)
+        self.retryable = retryable
+
+
+class Transport(Protocol):
+    def send(self, events: list[dict[str, Any]]) -> None:
+        """Deliver one batch or raise TransportError."""
+
+
+def _json_default(value: Any) -> Any:
+    """Encode values `json` does not know, e.g. numpy scalars and arrays."""
+    for method in ("item", "tolist"):
+        convert = getattr(value, method, None)
+        if callable(convert):
+            return convert()
+    return str(value)
+
+
+def encode(events: list[dict[str, Any]]) -> bytes:
+    """Serialise a batch as strict JSON.
+
+    NaN and +/-Infinity are not JSON. Python would write them as bare tokens,
+    which the ingestion database rejects for the whole batch, so they are sent
+    as null (a missing value). The round trip also catches the ones that only
+    appear after numpy values have been converted.
+    """
+    loose = json.dumps(events, default=_json_default)
+    if "NaN" not in loose and "Infinity" not in loose:
+        return loose.encode("utf-8")
+    strict = json.loads(loose, parse_constant=lambda _token: None)
+    return json.dumps(strict, allow_nan=False).encode("utf-8")
+
+
+class HttpTransport:
+    def __init__(self, api_url: str, *, timeout: float = 5.0) -> None:
+        self._url = api_url.rstrip("/") + INGEST_PATH
+        self._timeout = timeout
+
+    def send(self, events: list[dict[str, Any]]) -> None:
+        request = urllib.request.Request(
+            self._url,
+            data=encode(events),
+            headers={"Content-Type": "application/json"},
+            method="POST",
+        )
+        try:
+            with urllib.request.urlopen(request, timeout=self._timeout) as response:
+                response.read()
+        except urllib.error.HTTPError as error:
+            # 4xx means the batch itself is wrong, except "slow down" answers.
+            retryable = error.code >= 500 or error.code in (408, 429)
+            raise TransportError(
+                f"MNM ingest answered HTTP {error.code}", retryable=retryable
+            ) from error
+        except OSError as error:  # URLError, timeouts, refused connections
+            raise TransportError(f"MNM ingest unreachable: {error}") from error

mnm_sdk-0.1.1.dist-info/METADATA

@@ -0,0 +1,73 @@
+Metadata-Version: 2.4
+Name: mnm-sdk
+Version: 0.1.1
+Summary: MNM client: log model predictions asynchronously, without slowing inference
+License: MIT
+Project-URL: Repository, https://github.com/Sohan-Narewadi/MNM
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: dev
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: ruff>=0.6; extra == "dev"
+Requires-Dist: black>=24; extra == "dev"
+Dynamic: license-file
+
+# sdk/ — MNM Python client
+
+A tiny library a model service imports to log predictions to MNM **without slowing down inference**.
+
+## Install
+```bash
+pip install mnm-sdk
+```
+
+## Usage
+```python
+from mnm_sdk import Monitor
+
+monitor = Monitor(api_url="http://backend:8000", model="churn", version="3")
+
+@app.post("/predict")
+def predict(req: Request):
+    t0 = time.perf_counter()
+    y = model.predict(req.features)
+    monitor.log(features=req.features, output=y, latency_ms=(time.perf_counter() - t0) * 1000)
+    return y
+```
+
+## How it behaves
+- `log()` only appends to an in-memory bounded queue and returns: it never waits on the network and never raises, whatever is passed in.
+- A background daemon thread sends batches to `POST /v1/ingest` when `batch_size` events are queued or every `flush_interval` seconds, whichever comes first.
+- A failed batch is retried with exponential backoff (`max_retries`, `retry_backoff`). Each event carries its own `request_id` and `ts`, and ingestion is idempotent on those, so a retry never stores an event twice. A batch the server rejects as invalid (4xx) is not retried.
+- `NaN` and `Infinity` (for example a missing feature from pandas or numpy) are sent as `null`, because JSON has no such numbers.
+- If the queue is full, the **oldest** event is dropped and counted. Nothing blocks.
+- `close()` runs at interpreter exit (`atexit`) and sends what is still queued.
+- `monitor.stats` reports `logged`, `sent`, `dropped`, `failed` and `queued`.
+- No dependencies outside the standard library. Python 3.9+.
+
+| Option | Default | Meaning |
+|---|---|---|
+| `batch_size` | 100 | Send as soon as this many events are queued |
+| `flush_interval` | 1.0 s | Send whatever is queued at least this often |
+| `max_queue` | 10 000 | Queue capacity; beyond it the oldest event is dropped |
+| `max_retries` | 3 | Extra attempts for a batch that failed |
+| `retry_backoff` | 0.5 s | Wait before the first retry; doubles each time |
+| `timeout` | 5.0 s | HTTP timeout per request |
+
+`log()` also accepts `request_id=`, `label=` and `ts=`. The event format is the ingest contract in [`docs/api-spec.md`](../docs/api-spec.md).
+
+## Development
+```bash
+cd sdk
+pip install -e ".[dev]"
+pytest          # no network and no backend needed
+```
+
+## Layout
+```
+sdk/
+├── mnm_sdk/__init__.py, monitor.py, transport.py
+├── tests/
+└── pyproject.toml
+```

mnm_sdk-0.1.1.dist-info/RECORD

@@ -0,0 +1,8 @@
+mnm_sdk/__init__.py,sha256=M0TvE7VTyrhpcXwmx-xc6WslKOcLP_mUg4DsZ-IO5bk,271
+mnm_sdk/monitor.py,sha256=FMjJtqws9mOwwhAMrSCKuiYTRf3HdByF7DuwUGxaFkc,7489
+mnm_sdk/transport.py,sha256=yNusnEleSlyxFA_ucQjhOWSAFbBbhe_kK8BOwOhgGF4,2710
+mnm_sdk-0.1.1.dist-info/licenses/LICENSE,sha256=2CUzD54rDtWKLl7gKJeCFSHsU7oTH8jV1CIoE1W18QQ,1088
+mnm_sdk-0.1.1.dist-info/METADATA,sha256=WTm9uYRKlIWQIZxAZC97yTp59PElGuYnYu1kUYgpORY,2849
+mnm_sdk-0.1.1.dist-info/WHEEL,sha256=mOdXR9C3wMVov7cSY8PY-XCzo3SmmWCHvGQr5csJoao,91
+mnm_sdk-0.1.1.dist-info/top_level.txt,sha256=SkAyUB74_REBoZcpRbaZuNR3oZFyAmPPb2xcHgcopeM,8
+mnm_sdk-0.1.1.dist-info/RECORD,,

mnm_sdk-0.1.1.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (84.2.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

mnm_sdk-0.1.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sohan Narewadi and Radhesh Nair
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mnm_sdk-0.1.1.dist-info/top_level.txt

@@ -0,0 +1 @@
+mnm_sdk