Merge pull request #56 from vast-ai/obfuscate-mtoken

Obfuscate mtoken in logs

lib/backend.py

@@ -5,7 +5,7 @@ import base64
 import subprocess
 import dataclasses
 import logging
-from asyncio import wait, sleep, gather, Semaphore, FIRST_COMPLETED, create_task, get_running_loop
+from asyncio import wait, sleep, gather, Semaphore, FIRST_COMPLETED, create_task
 from typing import Tuple, Awaitable, NoReturn, List, Union, Callable, Optional
 from functools import cached_property
 from distutils.util import strtobool
@@ -30,7 +30,7 @@ from lib.data_types import (
     BenchmarkResult
 )
 
-VERSION = "0.1.0"
+VERSION = "0.2.0"
 
 MSG_HISTORY_LEN = 100
 log = logging.getLogger(__file__)
@@ -47,7 +47,7 @@ class Backend:
     This class is responsible for:
     1. Tailing logs and updating load time metrics
     2. Taking an EndpointHandler alongside incoming payload, preparing a json to be sent to the model, and
-       sending the request. It also updates metrics as it makes those requests.
+    sending the request. It also updates metrics as it makes those requests.
     3. Running a benchmark from an EndpointHandler
     """
 
@@ -66,19 +66,21 @@ class Backend:
     unsecured: bool = dataclasses.field(
         default_factory=lambda: bool(strtobool(os.environ.get("UNSECURED", "false"))),
     )
+    report_addr: str = dataclasses.field(
+        default_factory=lambda: os.environ.get("REPORT_ADDR", "https://run.vast.ai")
+    )
+    mtoken: str = dataclasses.field(
+        default_factory=lambda: os.environ.get("MASTER_TOKEN", "")
+    )
 
     def __post_init__(self):
         self.metrics = Metrics()
         self.metrics._set_version(self.version)
+        self.metrics._set_mtoken(self.mtoken)
         self._total_pubkey_fetch_errors = 0
         self._pubkey = self._fetch_pubkey()
         self.__start_healthcheck: bool = False
 
-        # NEW: FIFO queue + worker count
-        self.request_queue: "asyncio.Queue[tuple[EndpointHandler[ApiPayload_T], web.Request, asyncio.Future]]" = asyncio.Queue()
-        # If parallel allowed, let multiple workers drain the queue (order preserved by FIFO per worker; overall start order is FIFO).
-        self._num_workers: int = 1 if not self.allow_parallel_requests else int(os.environ.get("WORKERS", "4"))
-
     @property
     def pubkey(self) -> Optional[RSA.RsaKey]:
         if self._pubkey is None:
@@ -96,22 +98,6 @@ class Backend:
         timeout = ClientTimeout(total=None)
         return ClientSession(self.model_server_url, timeout=timeout, connector=connector)
 
-    async def _worker(self):
-        while True:
-            handler, request, fut = await self.request_queue.get()
-            try:
-                # Skip if already cancelled while waiting in the queue
-                if fut.cancelled():
-                    continue
-                res = await self.__process_enqueued_request(handler, request)
-                if not fut.cancelled():
-                    fut.set_result(res)
-            except Exception as e:
-                if not fut.cancelled():
-                    fut.set_exception(e)
-            finally:
-                self.request_queue.task_done()
-
     def create_handler(
         self,
         handler: EndpointHandler[ApiPayload_T],
@@ -125,59 +111,26 @@ class Backend:
 
     #######################################Private#######################################
     def _fetch_pubkey(self):
-        command = ["curl", "-X", "GET", "https://run.vast.ai/pubkey/"]
-        result = subprocess.check_output(command, universal_newlines=True)
-        log.debug("public key:")
-        log.debug(result)
-        key = None
-        for _ in range(5):
-            try:
-                key = RSA.import_key(result)
-                break
-            except ValueError as e:
-                log.debug(f"Error downloading key: {e}")
-                time.sleep(15)
-        if key is None:
-            self._total_pubkey_fetch_errors += 1
-            if self._total_pubkey_fetch_errors >= MAX_PUBKEY_FETCH_ATTEMPTS:
-                self.backend_errored("Failed to get autoscaler pubkey")
-        return key
+        report_addr = self.report_addr.rstrip("/")
+        command = ["curl", "-X", "GET", f"{report_addr}/pubkey/"]
+        try:
+            result = subprocess.check_output(command, universal_newlines=True)
+            log.debug("public key:")
+            log.debug(result)
+            key = RSA.import_key(result)
+            if key is not None:
+                return key
+        except (ValueError , subprocess.CalledProcessError) as e:
+            log.debug(f"Error downloading key: {e}")
+        self.backend_errored("Failed to get autoscaler pubkey")
+       
 
     async def __handle_request(
         self,
         handler: EndpointHandler[ApiPayload_T],
         request: web.Request,
     ) -> Union[web.Response, web.StreamResponse]:
-        """use this function to enqueue requests for FIFO processing"""
-        loop = get_running_loop()
-        fut: asyncio.Future = loop.create_future()
-
-        # If the client disconnects while waiting in the FIFO, cancel the future so the worker skips it
-        cancel_watch = create_task(request.wait_for_disconnection())
-        def _cancel_if_disconnected(_):
-            if not fut.done():
-                fut.cancel()
-        cancel_watch.add_done_callback(_cancel_if_disconnected)
-
-        try:
-            await self.request_queue.put((handler, request, fut))
-            return await fut
-        except asyncio.CancelledError:
-            # Propagate cancellation to ensure aiohttp doesn't expect a response body
-            raise
-        finally:
-            # Best-effort cleanup of the watcher
-            cancel_watch.cancel()
-
-    async def __process_enqueued_request(
-        self,
-        handler: EndpointHandler[ApiPayload_T],
-        request: web.Request,
-    ) -> Union[web.Response, web.StreamResponse]:
-        """
-        This contains the original __handle_request logic and is invoked by workers,
-        ensuring FIFO execution via asyncio.Queue.
-        """
+        """use this function to forward requests to the model endpoint"""
         try:
             data = await request.json()
             auth_data, payload = handler.get_data_from_request(data)
@@ -185,11 +138,8 @@ class Backend:
             return web.json_response(data=e.message, status=422)
         except json.JSONDecodeError:
             return web.json_response(dict(error="invalid JSON"), status=422)
-
         workload = payload.count_workload()
-        request_metrics: RequestMetrics = RequestMetrics(
-            request_idx=auth_data.request_idx, reqnum=auth_data.reqnum, workload=workload, status="Created"
-        )
+        request_metrics: RequestMetrics = RequestMetrics(request_idx=auth_data.request_idx, reqnum=auth_data.reqnum, workload=workload, status="Created")
 
         async def cancel_api_call_if_disconnected() -> web.Response:
             await request.wait_for_disconnection()
@@ -230,8 +180,6 @@ class Backend:
         acquired = False
         try:
             self.metrics._request_start(request_metrics)
-
-            # Preserve existing semaphore behavior for serializing requests when requested
             if self.allow_parallel_requests is False:
                 log.debug(f"Waiting to aquire Sem for reqnum:{request_metrics.reqnum}")
                 await self.sem.acquire()
@@ -241,7 +189,6 @@ class Backend:
                 )
             else:
                 log.debug(f"Starting request for reqnum:{request_metrics.reqnum}")
-
             done, pending = await wait(
                 [
                     create_task(make_request()),
@@ -309,14 +256,8 @@ class Backend:
                 self.backend_errored(str(e))
 
     async def _start_tracking(self) -> None:
-        # Start the FIFO workers alongside existing loops
-        worker_tasks = tuple(self._worker() for _ in range(self._num_workers))
         await gather(
-            self.__read_logs(),
-            self.metrics._send_metrics_loop(),
-            self.__healthcheck(),
-            self.metrics._send_delete_requests_loop(),
-            *worker_tasks,
+            self.__read_logs(), self.metrics._send_metrics_loop(), self.__healthcheck(), self.metrics._send_delete_requests_loop()
         )
 
     def backend_errored(self, msg: str) -> None:
@@ -348,7 +289,7 @@ class Backend:
         message = {
             key: value
             for (key, value) in (dataclasses.asdict(auth_data).items())
-            if key != "signature"
+            if key != "signature" and key != "__request_id"
         }
         if auth_data.reqnum < (self.reqnum - MSG_HISTORY_LEN):
             log.debug(
@@ -358,7 +299,7 @@ class Backend:
         elif message in self.msg_history:
             log.debug(f"message: {message} already in message history")
             return False
-        elif verify_signature(json.dumps(message, indent=4), auth_data.signature):
+        elif verify_signature(json.dumps(message, indent=4, sort_keys=True), auth_data.signature):
             self.reqnum = max(auth_data.reqnum, self.reqnum)
             self.msg_history.append(message)
             self.msg_history = self.msg_history[-MSG_HISTORY_LEN:]
@@ -377,10 +318,10 @@ class Backend:
                 with open(BENCHMARK_INDICATOR_FILE, "r") as f:
                     log.debug("already ran benchmark")
                     # trigger model load
-                    payload = self.benchmark_handler.make_benchmark_payload()
-                    _ = await self.__call_api(
-                        handler=self.benchmark_handler, payload=payload
-                    )
+                    # payload = self.benchmark_handler.make_benchmark_payload()
+                    # _ = await self.__call_api(
+                    #     handler=self.benchmark_handler, payload=payload
+                    # )
                     return float(f.readline())
             except FileNotFoundError:
                 pass
@@ -455,7 +396,7 @@ class Backend:
                         )
                         # some backends need a few seconds after logging successful startup before
                         # they can begin accepting requests
-                        await sleep(5)
+                        # await sleep(5)
                         try:
                             max_throughput = await run_benchmark()
                             self.__start_healthcheck = True

lib/data_types.py

@@ -65,12 +65,12 @@ class ApiPayload(ABC):
 class AuthData:
     """data used to authenticate requester"""
 
-    signature: str
     cost: str
     endpoint: str
     reqnum: int
-    url: str
     request_idx: int
+    signature: str
+    url: str
 
     @classmethod
     def from_json_msg(cls, json_msg: Dict[str, Any]):
@@ -190,11 +190,12 @@ class SystemMetrics:
         self.additional_disk_usage = disk_usage - self.last_disk_usage
         self.last_disk_usage = disk_usage
 
-    def reset(self):
+    def reset(self, expected: float | None) -> None:
         # autoscaler excepts model_loading_time to be populated only once, when the instance has
         # finished benchmarking and is ready to receive requests. This applies to restarted instances
         # as well: they should send model_loading_time once when they are done loading
-        self.model_loading_time = None
+        if self.model_loading_time == expected:
+            self.model_loading_time = None
 
 
 @dataclass
@@ -285,6 +286,7 @@ class AutoScalerData:
     """Data that is reported to autoscaler"""
 
     id: int
+    mtoken: str
     version: str
     loadtime: float
     cur_load: float

lib/metrics.py

@@ -28,6 +28,7 @@ def get_url() -> str:
 @dataclass
 class Metrics:
     version: str = "0"
+    mtoken: str = ""
     last_metric_update: float = 0.0
     last_request_served: float = 0.0
     update_pending: bool = False
@@ -142,12 +143,16 @@ class Metrics:
     def _set_version(self, version: str) -> None:
         self.version = version
 
+    def _set_mtoken(self, mtoken: str) -> None:
+        self.mtoken = mtoken
+
     #######################################Private#######################################
 
     async def __send_delete_requests_and_reset(self):
         async def post(report_addr: str, idxs: list[int], success_flag: bool) -> bool:
             data = {
                 "worker_id": self.id,
+                "mtoken": self.mtoken,
                 "request_idxs": idxs,
                 "success": success_flag,
             }
@@ -180,6 +185,10 @@ class Metrics:
             return  # nothing to do
 
         for report_addr in self.report_addr:
+            # TODO: Add a Redis subscriber queue for delete_requests
+            if report_addr == "https://cloud.vast.ai/api/v0":
+                # Patch: ignore the Redis API report_addr
+                continue
             sent_success = True
             sent_failed  = True
 
@@ -200,11 +209,14 @@ class Metrics:
 
     async def __send_metrics_and_reset(self):
 
+        loadtime_snapshot = self.system_metrics.model_loading_time
+
         def compute_autoscaler_data() -> AutoScalerData:
             return AutoScalerData(
                 id=self.id,
+                mtoken=self.mtoken,
                 version=self.version,
-                loadtime=(self.system_metrics.model_loading_time or 0.0),
+                loadtime=(loadtime_snapshot or 0.0), 
                 new_load=self.model_metrics.workload_processing,
                 cur_load=self.model_metrics.cur_load,
                 rej_load=self.model_metrics.workload_rejected,
@@ -222,17 +234,25 @@ class Metrics:
 
         async def send_data(report_addr: str) -> bool:
             data = compute_autoscaler_data()
-            full_path = report_addr.rstrip("/") + "/worker_status/"
+            log_data = asdict(data)
+            def obfuscate(secret: str) -> str:
+                if secret is None:
+                    return ""
+                return secret[:7] + "..." if len(secret) > 7 else ("*" * len(secret))
+            
+            log_data["mtoken"] = obfuscate(log_data.get("mtoken"))
             log.debug(
                 "\n".join(
                     [
                         "#" * 60,
                         f"sending data to autoscaler",
-                        f"{json.dumps((asdict(data)), indent=2)}",
+                        f"{json.dumps(log_data, indent=2)}",
                         "#" * 60,
                     ]
                 )
             )
+
+            full_path = report_addr.rstrip("/") + "/worker_status/"
             for attempt in range(1, 4):
                 try:
                     session = await self.http()
@@ -252,11 +272,15 @@ class Metrics:
 
         self.system_metrics.update_disk_usage()
 
+        sent = False
         for report_addr in self.report_addr:
-            success = await send_data(report_addr)
-            if success is True:
+            if await send_data(report_addr):
+                sent = True
                 break
-        self.update_pending = False
-        self.model_metrics.reset()
-        self.system_metrics.reset()
-        self.last_metric_update = time.time()
+
+        if sent:
+            # clear the one-shot loadtime only if we actually sent *this* value
+            self.system_metrics.reset(expected=loadtime_snapshot)
+            self.update_pending = False
+            self.model_metrics.reset()
+            self.last_metric_update = time.time()

start_server.sh

@@ -9,7 +9,7 @@ ENV_PATH="$WORKSPACE_DIR/worker-env"
 DEBUG_LOG="$WORKSPACE_DIR/debug.log"
 PYWORKER_LOG="$WORKSPACE_DIR/pyworker.log"
 
-REPORT_ADDR="${REPORT_ADDR:-https://cloud.vast.ai/api/v0,https://run.vast.ai}"
+REPORT_ADDR="${REPORT_ADDR:-https://run.vast.ai}"
 USE_SSL="${USE_SSL:-true}"
 WORKER_PORT="${WORKER_PORT:-3000}"
 mkdir -p "$WORKSPACE_DIR"

workers/comfyui-json/client.py

@@ -98,6 +98,7 @@ def call_text2image_workflow(
         endpoint=route_response["endpoint"],
         reqnum=route_response["reqnum"],
         url=route_response["url"],
+        request_idx=route_response["request_idx"],
     )
     
     # Build the payload for the worker request

workers/comfyui/client.py

@@ -82,6 +82,7 @@ def call_custom_workflow_for_sd3(
         endpoint=message["endpoint"],
         reqnum=message["reqnum"],
         url=message["url"],
+        request_idx=message["request_idx"],
     )
     workflow = {
         "3": {