askrobots / dbbasic-object-server
Public
Runs live, versioned Python application objects on one VM, with a suite of apps: projects, notes, tasks, files and more.
danq9
signup: the person chooses the password and name on the confirm page
83c7aa6
2 weeks, 5 days ago
History
object_worker_pool.py · 19.5 KB · 567 lines
| 1 | """Persistent worker-process pool for near-in-process object execution. |
| 2 | |
| 3 | Objects that are not listed in ``DBBASIC_TRUSTED_IN_PROCESS_OBJECTS`` run |
| 4 | through ``object_execution.execute_python_object_subprocess`` today, which |
| 5 | spawns a brand new interpreter for every single request. That keeps process |
| 6 | isolation and wall-clock timeout enforcement, but a cold interpreter boot |
| 7 | dominates latency (roughly 100-200ms) for objects that could otherwise run |
| 8 | in a couple of milliseconds. |
| 9 | |
| 10 | This module keeps a small pool of long-lived worker subprocesses around |
| 11 | instead. Each worker runs a tiny read-request/write-response loop |
| 12 | (``_worker_main`` below, reachable as ``python -m object_worker_pool |
| 13 | --worker``) and re-loads the target object's source fresh on every request |
| 14 | via ``python_object_runtime.PythonObjectRuntime`` -- so a warm worker still |
| 15 | picks up source edits made between requests (the "live edit" guarantee |
| 16 | objects depend on). Isolation and timeout enforcement are preserved because |
| 17 | every request still runs in its own OS process; only the process-spawn cost |
| 18 | is amortized away by keeping the process around between requests. |
| 19 | |
| 20 | Disabled by default: ``DBBASIC_WORKER_POOL_SIZE`` unset or ``0`` means the |
| 21 | pool is never constructed and callers should keep using the existing |
| 22 | spawn-per-request path untouched. |
| 23 | |
| 24 | Wire protocol |
| 25 | -------------- |
| 26 | Parent and worker exchange one JSON message per request/response, each |
| 27 | framed as a 4-byte big-endian length prefix followed by that many bytes of |
| 28 | UTF-8 JSON (``_encode_message`` / ``_read_message_async`` / |
| 29 | ``_read_message_sync``). |
| 30 | |
| 31 | Request message:: |
| 32 | |
| 33 | { |
| 34 | "object_id": str, |
| 35 | "method": str, |
| 36 | "payload": dict, |
| 37 | "correlation_id": str | None, |
| 38 | "path": str | None, # pre-resolved by the pool manager |
| 39 | "base_dir": str | None, # optional PythonObjectRuntime base_dir override |
| 40 | } |
| 41 | |
| 42 | Response message is ``object_execution.ObjectExecutionResult.to_dict()``. |
| 43 | |
| 44 | Worker stderr is left connected to the parent's stderr (not piped), so |
| 45 | worker-side tracebacks/log noise land in the normal server logs. |
| 46 | """ |
| 47 | from __future__ import annotations |
| 48 | |
| 49 | import asyncio |
| 50 | import atexit |
| 51 | import json |
| 52 | import os |
| 53 | import signal |
| 54 | import struct |
| 55 | import sys |
| 56 | import time |
| 57 | from pathlib import Path |
| 58 | from typing import Any, Iterable |
| 59 | |
| 60 | import object_execution |
| 61 | from object_execution import ( |
| 62 | ObjectExecutionError, |
| 63 | ObjectExecutionRequest, |
| 64 | ObjectExecutionResult, |
| 65 | ) |
| 66 | from object_namespace import resolve_object_id |
| 67 | |
| 68 | WORKER_POOL_SIZE_ENV = "DBBASIC_WORKER_POOL_SIZE" |
| 69 | |
| 70 | _LENGTH_PREFIX = struct.Struct(">I") |
| 71 | _MAX_MESSAGE_BYTES = 64 * 1024 * 1024 # guard against a corrupt/desynced stream |
| 72 | |
| 73 | # Reuse object_execution's private timestamp/duration/format helpers so the |
| 74 | # results this module builds (timeouts, crashes, worker responses) are |
| 75 | # byte-for-byte identical in shape to the ones execute_python_object_subprocess |
| 76 | # produces. These are considered part of the internal contract between the |
| 77 | # two modules, not object_execution's public API. |
| 78 | _utc_timestamp = object_execution._utc_timestamp |
| 79 | _duration_ms = object_execution._duration_ms |
| 80 | _format_seconds = object_execution._format_seconds |
| 81 | |
| 82 | |
| 83 | # --------------------------------------------------------------------------- |
| 84 | # Wire protocol helpers |
| 85 | # --------------------------------------------------------------------------- |
| 86 | |
| 87 | |
| 88 | def _encode_message(payload: dict[str, Any]) -> bytes: |
| 89 | body = json.dumps(payload).encode("utf-8") |
| 90 | return _LENGTH_PREFIX.pack(len(body)) + body |
| 91 | |
| 92 | |
| 93 | async def _read_message_async(reader: asyncio.StreamReader) -> dict[str, Any]: |
| 94 | header = await reader.readexactly(4) |
| 95 | (length,) = _LENGTH_PREFIX.unpack(header) |
| 96 | if length > _MAX_MESSAGE_BYTES: |
| 97 | raise ValueError(f"worker message too large ({length} bytes)") |
| 98 | body = await reader.readexactly(length) |
| 99 | return json.loads(body.decode("utf-8")) |
| 100 | |
| 101 | |
| 102 | def _read_message_sync(stream: Any) -> dict[str, Any] | None: |
| 103 | header = _read_exact_sync(stream, 4) |
| 104 | if header is None: |
| 105 | return None |
| 106 | (length,) = _LENGTH_PREFIX.unpack(header) |
| 107 | body = _read_exact_sync(stream, length) |
| 108 | if body is None: |
| 109 | return None |
| 110 | return json.loads(body.decode("utf-8")) |
| 111 | |
| 112 | |
| 113 | def _read_exact_sync(stream: Any, size: int) -> bytes | None: |
| 114 | chunks: list[bytes] = [] |
| 115 | remaining = size |
| 116 | while remaining > 0: |
| 117 | chunk = stream.read(remaining) |
| 118 | if not chunk: |
| 119 | return None |
| 120 | chunks.append(chunk) |
| 121 | remaining -= len(chunk) |
| 122 | return b"".join(chunks) |
| 123 | |
| 124 | |
| 125 | # --------------------------------------------------------------------------- |
| 126 | # Worker process (child side) |
| 127 | # --------------------------------------------------------------------------- |
| 128 | |
| 129 | |
| 130 | def _worker_main() -> None: |
| 131 | """Entry point for a worker subprocess (``python -m object_worker_pool --worker``).""" |
| 132 | from python_object_runtime import PythonObjectRuntime |
| 133 | |
| 134 | stdin = sys.stdin.buffer |
| 135 | stdout = sys.stdout.buffer |
| 136 | |
| 137 | while True: |
| 138 | try: |
| 139 | request_message = _read_message_sync(stdin) |
| 140 | except Exception: |
| 141 | return |
| 142 | if request_message is None: |
| 143 | return # parent closed stdin (pool shutdown) -- exit cleanly |
| 144 | |
| 145 | response = _handle_worker_request(request_message, PythonObjectRuntime()) |
| 146 | try: |
| 147 | stdout.write(_encode_message(response)) |
| 148 | stdout.flush() |
| 149 | except Exception: |
| 150 | return |
| 151 | |
| 152 | |
| 153 | def _handle_worker_request(message: dict[str, Any], runtime: Any) -> dict[str, Any]: |
| 154 | object_id = message.get("object_id") or "" |
| 155 | method = message.get("method") or "GET" |
| 156 | payload = message.get("payload") or {} |
| 157 | correlation_id = message.get("correlation_id") |
| 158 | path_value = message.get("path") |
| 159 | |
| 160 | request = ObjectExecutionRequest( |
| 161 | object_id=object_id, |
| 162 | method=method, |
| 163 | payload=payload, |
| 164 | path=Path(path_value) if path_value else None, |
| 165 | correlation_id=correlation_id, |
| 166 | ) |
| 167 | |
| 168 | try: |
| 169 | result = object_execution.execute_object(runtime, request, roots=None) |
| 170 | except BaseException as exc: # keep the worker loop alive no matter what happened |
| 171 | result = _worker_error_result(request, exc) |
| 172 | |
| 173 | return result.to_dict() |
| 174 | |
| 175 | |
| 176 | def _worker_error_result(request: ObjectExecutionRequest, exc: BaseException) -> ObjectExecutionResult: |
| 177 | import traceback |
| 178 | |
| 179 | now = _utc_timestamp() |
| 180 | return ObjectExecutionResult( |
| 181 | object_id=request.object_id, |
| 182 | method=request.normalized_method(), |
| 183 | path=request.path, |
| 184 | ok=False, |
| 185 | result=None, |
| 186 | error=ObjectExecutionError( |
| 187 | type=object_execution.WORKER_ERROR_TYPE, |
| 188 | message=f"Object worker failed: {type(exc).__name__}: {exc}", |
| 189 | traceback="".join(traceback.format_exception(type(exc), exc, exc.__traceback__)), |
| 190 | ), |
| 191 | started_at=now, |
| 192 | finished_at=now, |
| 193 | duration_ms=0.0, |
| 194 | correlation_id=request.correlation_id, |
| 195 | ) |
| 196 | |
| 197 | |
| 198 | # --------------------------------------------------------------------------- |
| 199 | # Pool manager (parent side) |
| 200 | # --------------------------------------------------------------------------- |
| 201 | |
| 202 | _LAZY = object() # sentinel occupying a queue slot: "spawn a worker on demand" |
| 203 | |
| 204 | |
| 205 | class _Worker: |
| 206 | __slots__ = ("process", "pid") |
| 207 | |
| 208 | def __init__(self, process: "asyncio.subprocess.Process"): |
| 209 | self.process = process |
| 210 | self.pid = process.pid |
| 211 | |
| 212 | |
| 213 | def _error_result( |
| 214 | request: ObjectExecutionRequest, |
| 215 | method: str, |
| 216 | error_type: str, |
| 217 | message: str, |
| 218 | started_at: str, |
| 219 | started_perf: float, |
| 220 | ) -> ObjectExecutionResult: |
| 221 | return ObjectExecutionResult( |
| 222 | object_id=request.object_id, |
| 223 | method=method, |
| 224 | path=request.path, |
| 225 | ok=False, |
| 226 | result=None, |
| 227 | error=ObjectExecutionError(type=error_type, message=message), |
| 228 | started_at=started_at, |
| 229 | finished_at=_utc_timestamp(), |
| 230 | duration_ms=_duration_ms(started_perf), |
| 231 | correlation_id=request.correlation_id, |
| 232 | ) |
| 233 | |
| 234 | |
| 235 | def _result_from_dict(data: dict[str, Any]) -> ObjectExecutionResult: |
| 236 | error = None |
| 237 | error_data = data.get("error") |
| 238 | if error_data: |
| 239 | error = ObjectExecutionError( |
| 240 | type=error_data["type"], |
| 241 | message=error_data["message"], |
| 242 | traceback=error_data.get("traceback", ""), |
| 243 | ) |
| 244 | |
| 245 | path_value = data.get("path") |
| 246 | return ObjectExecutionResult( |
| 247 | object_id=data["object_id"], |
| 248 | method=data["method"], |
| 249 | path=Path(path_value) if path_value else None, |
| 250 | ok=bool(data["ok"]), |
| 251 | result=data.get("result"), |
| 252 | error=error, |
| 253 | started_at=data["started_at"], |
| 254 | finished_at=data["finished_at"], |
| 255 | duration_ms=data["duration_ms"], |
| 256 | correlation_id=data.get("correlation_id"), |
| 257 | ) |
| 258 | |
| 259 | |
| 260 | def _request_with_resolved_path( |
| 261 | request: ObjectExecutionRequest, |
| 262 | roots: Iterable[Path] | None, |
| 263 | ) -> ObjectExecutionRequest: |
| 264 | if request.path is not None: |
| 265 | return request |
| 266 | return ObjectExecutionRequest( |
| 267 | object_id=request.object_id, |
| 268 | method=request.method, |
| 269 | payload=request.payload, |
| 270 | path=resolve_object_id(request.object_id, roots), |
| 271 | correlation_id=request.correlation_id, |
| 272 | ) |
| 273 | |
| 274 | |
| 275 | class WorkerPool: |
| 276 | """Async pool of persistent worker subprocesses. |
| 277 | |
| 278 | Workers are spawned lazily -- nothing is started until the first |
| 279 | ``execute()`` call needs one, and at most ``size`` workers ever run at |
| 280 | once. Each worker handles one request at a time; ``execute()`` acquires |
| 281 | a free worker (or spawns one if a lazy slot is available), ships it one |
| 282 | request over the wire protocol above, and returns it to the pool when |
| 283 | done. A worker that times out or crashes is killed/discarded and its |
| 284 | slot is put back as "lazy" so a replacement spawns on next use. |
| 285 | """ |
| 286 | |
| 287 | def __init__( |
| 288 | self, |
| 289 | size: int, |
| 290 | *, |
| 291 | env: dict[str, str] | None = None, |
| 292 | python_executable: str | None = None, |
| 293 | ) -> None: |
| 294 | if size <= 0: |
| 295 | raise ValueError("WorkerPool size must be positive") |
| 296 | self.size = size |
| 297 | self._env = dict(env) if env is not None else None |
| 298 | self._python_executable = python_executable or sys.executable |
| 299 | self._loop: asyncio.AbstractEventLoop | None = None |
| 300 | self._queue: asyncio.Queue | None = None |
| 301 | self._known_pids: set[int] = set() |
| 302 | self._closed = False |
| 303 | atexit.register(self._atexit_kill) |
| 304 | |
| 305 | # -- queue / lifecycle ------------------------------------------------- |
| 306 | |
| 307 | def _queue_for_current_loop(self) -> asyncio.Queue: |
| 308 | loop = asyncio.get_running_loop() |
| 309 | if self._loop is not loop: |
| 310 | # A pool must not straddle event loops: asyncio subprocess |
| 311 | # transports are bound to the loop that created them. If we're |
| 312 | # being used from a new loop (process restart of the loop, or a |
| 313 | # test harness that opens a fresh loop per call), hard-kill any |
| 314 | # previously known workers by pid (loop-independent) and start |
| 315 | # over with fresh lazy slots. |
| 316 | self._reap_stale_workers() |
| 317 | self._loop = loop |
| 318 | self._queue = asyncio.Queue() |
| 319 | for _ in range(self.size): |
| 320 | self._queue.put_nowait(_LAZY) |
| 321 | assert self._queue is not None |
| 322 | return self._queue |
| 323 | |
| 324 | def _reap_stale_workers(self) -> None: |
| 325 | for pid in self._known_pids: |
| 326 | _hard_kill(pid) |
| 327 | self._known_pids.clear() |
| 328 | |
| 329 | def _atexit_kill(self) -> None: |
| 330 | for pid in list(self._known_pids): |
| 331 | _hard_kill(pid) |
| 332 | |
| 333 | async def _spawn_worker(self) -> _Worker: |
| 334 | kwargs: dict[str, Any] = { |
| 335 | "stdin": asyncio.subprocess.PIPE, |
| 336 | "stdout": asyncio.subprocess.PIPE, |
| 337 | "stderr": None, # inherit -- worker stderr flows to server logs |
| 338 | } |
| 339 | if self._env is not None: |
| 340 | kwargs["env"] = self._env |
| 341 | process = await asyncio.create_subprocess_exec( |
| 342 | self._python_executable, |
| 343 | "-m", |
| 344 | "object_worker_pool", |
| 345 | "--worker", |
| 346 | **kwargs, |
| 347 | ) |
| 348 | worker = _Worker(process) |
| 349 | self._known_pids.add(worker.pid) |
| 350 | return worker |
| 351 | |
| 352 | async def _kill_worker(self, worker: _Worker) -> None: |
| 353 | process = worker.process |
| 354 | self._known_pids.discard(worker.pid) |
| 355 | if process.returncode is not None: |
| 356 | return |
| 357 | try: |
| 358 | process.kill() |
| 359 | except ProcessLookupError: |
| 360 | return |
| 361 | try: |
| 362 | await asyncio.wait_for(process.wait(), timeout=1.0) |
| 363 | except asyncio.TimeoutError: |
| 364 | pass |
| 365 | |
| 366 | async def _exit_code(self, worker: _Worker) -> int | None: |
| 367 | process = worker.process |
| 368 | if process.returncode is not None: |
| 369 | return process.returncode |
| 370 | try: |
| 371 | await asyncio.wait_for(process.wait(), timeout=0.5) |
| 372 | except asyncio.TimeoutError: |
| 373 | return None |
| 374 | return process.returncode |
| 375 | |
| 376 | # -- execution ----------------------------------------------------- |
| 377 | |
| 378 | async def execute( |
| 379 | self, |
| 380 | request: ObjectExecutionRequest, |
| 381 | *, |
| 382 | timeout_seconds: float, |
| 383 | roots: Iterable[Path] | None = None, |
| 384 | ) -> ObjectExecutionResult: |
| 385 | """Execute one object method on a warm worker, enforcing timeout_seconds.""" |
| 386 | if self._closed: |
| 387 | raise RuntimeError("WorkerPool is closed") |
| 388 | |
| 389 | started_perf = time.perf_counter() |
| 390 | started_at = _utc_timestamp() |
| 391 | method = request.normalized_method() |
| 392 | resolved_request = _request_with_resolved_path(request, roots) |
| 393 | |
| 394 | message = { |
| 395 | "object_id": resolved_request.object_id, |
| 396 | "method": method, |
| 397 | "payload": resolved_request.normalized_payload(), |
| 398 | "correlation_id": resolved_request.correlation_id, |
| 399 | "path": str(resolved_request.path) if resolved_request.path is not None else None, |
| 400 | } |
| 401 | # Encode before taking a worker: a payload JSON can't carry (bytes, |
| 402 | # say) used to raise out of here holding the worker, and once every |
| 403 | # worker had leaked that way each request waited forever. |
| 404 | try: |
| 405 | encoded = _encode_message(message) |
| 406 | except (TypeError, ValueError) as exc: |
| 407 | return _error_result( |
| 408 | resolved_request, |
| 409 | method, |
| 410 | object_execution.WORKER_ERROR_TYPE, |
| 411 | f"Request can't be sent to an object worker: {type(exc).__name__}: {exc}", |
| 412 | started_at, |
| 413 | started_perf, |
| 414 | ) |
| 415 | |
| 416 | queue = self._queue_for_current_loop() |
| 417 | try: |
| 418 | slot = await asyncio.wait_for(queue.get(), timeout=timeout_seconds) |
| 419 | except asyncio.TimeoutError: |
| 420 | return _error_result( |
| 421 | resolved_request, |
| 422 | method, |
| 423 | object_execution.TIMEOUT_ERROR_TYPE, |
| 424 | ( |
| 425 | f"No object worker came free for {resolved_request.object_id} " |
| 426 | f"within {_format_seconds(timeout_seconds)} seconds" |
| 427 | ), |
| 428 | started_at, |
| 429 | started_perf, |
| 430 | ) |
| 431 | worker = slot if isinstance(slot, _Worker) else None |
| 432 | |
| 433 | if worker is None: |
| 434 | try: |
| 435 | worker = await self._spawn_worker() |
| 436 | except BaseException as exc: |
| 437 | queue.put_nowait(_LAZY) |
| 438 | if not isinstance(exc, Exception): |
| 439 | raise |
| 440 | return _error_result( |
| 441 | resolved_request, |
| 442 | method, |
| 443 | object_execution.WORKER_ERROR_TYPE, |
| 444 | f"Failed to start object worker: {type(exc).__name__}: {exc}", |
| 445 | started_at, |
| 446 | started_perf, |
| 447 | ) |
| 448 | |
| 449 | worker_died = False |
| 450 | try: |
| 451 | response = await asyncio.wait_for( |
| 452 | _talk_to_worker(worker, encoded), |
| 453 | timeout=timeout_seconds, |
| 454 | ) |
| 455 | except asyncio.TimeoutError: |
| 456 | worker_died = True |
| 457 | await self._kill_worker(worker) |
| 458 | result = _error_result( |
| 459 | resolved_request, |
| 460 | method, |
| 461 | object_execution.TIMEOUT_ERROR_TYPE, |
| 462 | ( |
| 463 | f"{method} timed out for object {resolved_request.object_id} " |
| 464 | f"after {_format_seconds(timeout_seconds)} seconds" |
| 465 | ), |
| 466 | started_at, |
| 467 | started_perf, |
| 468 | ) |
| 469 | except (asyncio.IncompleteReadError, ConnectionResetError, BrokenPipeError, OSError): |
| 470 | worker_died = True |
| 471 | exit_code = await self._exit_code(worker) |
| 472 | self._known_pids.discard(worker.pid) |
| 473 | result = _error_result( |
| 474 | resolved_request, |
| 475 | method, |
| 476 | object_execution.WORKER_ERROR_TYPE, |
| 477 | f"Object worker exited without a result (exit code {exit_code})", |
| 478 | started_at, |
| 479 | started_perf, |
| 480 | ) |
| 481 | except Exception as exc: |
| 482 | # A reply we can't read (too large, not JSON) leaves the worker's |
| 483 | # stream in an unknown state: replace the worker. |
| 484 | worker_died = True |
| 485 | self._known_pids.discard(worker.pid) |
| 486 | _hard_kill(worker.pid) |
| 487 | result = _error_result( |
| 488 | resolved_request, |
| 489 | method, |
| 490 | object_execution.WORKER_ERROR_TYPE, |
| 491 | f"Object worker sent a reply that couldn't be read: {type(exc).__name__}: {exc}", |
| 492 | started_at, |
| 493 | started_perf, |
| 494 | ) |
| 495 | except BaseException: |
| 496 | # Cancelled mid-conversation: same, and the slot still goes back |
| 497 | # (finally) so the pool never shrinks. |
| 498 | worker_died = True |
| 499 | self._known_pids.discard(worker.pid) |
| 500 | _hard_kill(worker.pid) |
| 501 | raise |
| 502 | else: |
| 503 | result = _result_from_dict(response) |
| 504 | finally: |
| 505 | queue.put_nowait(_LAZY if worker_died else worker) |
| 506 | |
| 507 | return result |
| 508 | |
| 509 | async def shutdown(self) -> None: |
| 510 | """Best-effort termination of every known worker. Never blocks long.""" |
| 511 | self._closed = True |
| 512 | atexit.unregister(self._atexit_kill) |
| 513 | |
| 514 | try: |
| 515 | loop = asyncio.get_running_loop() |
| 516 | except RuntimeError: |
| 517 | loop = None |
| 518 | |
| 519 | if self._queue is None or loop is not self._loop: |
| 520 | # Either never used, or being shut down from a different event |
| 521 | # loop than the one that spawned the current workers (asyncio |
| 522 | # subprocess transports can't be awaited from another loop). |
| 523 | # Hard-kill by pid instead of the graceful kill()+wait() path. |
| 524 | self._reap_stale_workers() |
| 525 | return |
| 526 | |
| 527 | workers: list[_Worker] = [] |
| 528 | while not self._queue.empty(): |
| 529 | slot = self._queue.get_nowait() |
| 530 | if isinstance(slot, _Worker): |
| 531 | workers.append(slot) |
| 532 | for worker in workers: |
| 533 | await self._kill_worker(worker) |
| 534 | self._known_pids.clear() |
| 535 | |
| 536 | |
| 537 | async def _talk_to_worker(worker: _Worker, encoded: bytes) -> dict[str, Any]: |
| 538 | process = worker.process |
| 539 | assert process.stdin is not None and process.stdout is not None |
| 540 | process.stdin.write(encoded) |
| 541 | await process.stdin.drain() |
| 542 | return await _read_message_async(process.stdout) |
| 543 | |
| 544 | |
| 545 | def _hard_kill(pid: int) -> None: |
| 546 | try: |
| 547 | os.kill(pid, signal.SIGKILL) |
| 548 | except (ProcessLookupError, PermissionError, OSError): |
| 549 | pass |
| 550 | |
| 551 | |
| 552 | # --------------------------------------------------------------------------- |
| 553 | # CLI entry point |
| 554 | # --------------------------------------------------------------------------- |
| 555 | |
| 556 | |
| 557 | def main(argv: list[str] | None = None) -> int: |
| 558 | args = sys.argv[1:] if argv is None else argv |
| 559 | if "--worker" in args: |
| 560 | _worker_main() |
| 561 | return 0 |
| 562 | print("usage: python -m object_worker_pool --worker", file=sys.stderr) |
| 563 | return 2 |
| 564 | |
| 565 | |
| 566 | if __name__ == "__main__": |
| 567 | raise SystemExit(main()) |